Available freely, today.

Ninety-three per cent of AI Mode is zero-click. That is the architectural fact every marketing function is now operating against — whether the strategy deck has caught up or not. The unit of value has shifted from rank to citation, from click to extraction, from keyword to fragment. The shift is not coming. It is here.

I have spent the last year watching this conversation play out inside marketing organisations. The pattern is identical. A team asks how to recover lost clicks. The question is wrong before it finishes. The click was the tax retrieval used to charge — and retrieval has changed. Rebuilding traffic infrastructure for an outcome the architecture has eliminated is the most expensive mistake a senior marketer can make in 2026.

So I wrote a manual (with my team of agents, of course).

Why this is a manual, not another book

The Agentic CMO — the second edition coming in June — makes the strategic argument: marketing organisations that orchestrate human and artificial intelligence as a unified force will outcompete those that adopt AI as a tool stack. Cited is the discipline that argument depends on. It is the field kit. The book argues; the manual operates.

Cited is built for the senior manager doing the delivery — SEO/GEO, content, analytics, brand, marketing operations. Fifteen chapters, nine appendices, three governing questions in every chapter: what carries over from SEO, what is genuinely new, and what you do Monday morning. The chapter pattern repeats deliberately so it can sit open on a second monitor while you audit a robots.txt, fix an Entity Bible, or push back on a vendor pitch.

What’s inside:

• A continuity map of the eight SEO disciplines — what carries over to GEO, what re-weights, and where the agency proposal is double-charging you for work you already own.

• The dual-layer citation model — direct citation versus indirect ranking inheritance — and the schema misconception that has cost the industry two years of misallocated budget.

• A field guide to six engines — Google AI Overviews, ChatGPT Search, Perplexity, Microsoft Copilot, Anthropic Claude, and the Gemini/Meta AI/Grok long tail — each with its retrieval stack, citation logic, and paid-placement reality.

• A crawler portfolio framework — robots.txt, ai.txt, llms.txt — set against the post-July-2025 Cloudflare default-block reality and a register of the crawlers worth knowing by name.

• The Entity Bible — eight platforms in priority order, the Wikidata canonical fields, and the sameAs graph that resolves you as one entity across engines.

• A measurement discipline that doesn’t require a vendor — fixed prompt battery, citation-share dashboard, honest limits on what attribution still cannot do — followed by a 90-day playbook and an operating-model spec for who owns what.

The bias is against hype. Roughly eighty per cent of classical SEO carries over to GEO with re-weighting; Chapter 2 names which agencies are charging premium rates to rebuild work organisations already own. Chapter 5 corrects the schema misconception that has cost the industry two years of misallocated budget. Chapter 8 puts YouTube and multimodal where the evidence says they belong, not where most marketing programmes pretend they do. The manual ends with a chapter named What We Don’t Yet Know — because some of the most consequential questions in this field are not settled, and pretending otherwise would waste your reading time.

Every empirical claim carries an evidence weight inline: Strong, Moderate, Anecdotal, Theoretical. Vendor-funded data is flagged at first reference. The source canon, in Appendix G, lists every reference with funder, sample size, and replication status. Where the canon disagrees, the disagreement is named — not smoothed.

Why now

The manual is the answer to a question I’ve been asked too often to keep answering one CMO at a time: where do I start? Cited starts at the audit, runs through the engines, and lands at the operating model. It assumes you are the senior manager who has to make the work happen.

Two questions worth sitting with this week. Is your weekly dashboard headlined by traffic or by citation? And of the eight SEO disciplines on the continuity map, how many are passing today — and how many is your agency about to charge you to rebuild?

Cited is at the-agentic-cmo.com/cited.

Also available in Kindle format at a nominal price here.

The price is zero. The work it asks you to do is not.

The US-China AI competition is usually framed as a strategic choice for Western enterprises — which side of the divergence to back. From the practitioner seat, that framing has been overtaken by events. Cursor uses Alibaba’s Qwen as the open-source base for its internal model. Airbnb relies heavily on Qwen for user-facing features. Andreessen Horowitz partner Martin Casado estimates that roughly 80% of startups building on open-source AI stacks are running Chinese models. Western frontier labs still hold a clear lead on benchmark capability and on the licensing position that matters for regulated workloads. The question for marketing leaders is how to architect a procurement and governance model that uses both without accumulating risk.

Focus On: The structural divergence

The US and Chinese AI ecosystems are no longer running the same playbook. American hyperscalers will spend a combined $650 billion on AI capex this year, with Microsoft alone having spent $80 billion in 2025. Alibaba’s announced AI investment is $53 billion over three years. The compute gap is wide and widening, but it cuts both ways. Chinese labs, constrained on advanced silicon, organised around open weights and inference efficiency. The result is a different theory of value creation — closed frontier capability on the American side, open-weight diffusion across the economy on the Chinese side.

The token-share data tells the story. Chinese AI providers crossed 45% of OpenRouter traffic by April 2026, up from less than 2% a year earlier. A documented enterprise benchmark — 50,000 financial documents per day — costs $4,200 per month on a Western frontier model and $210 on DeepSeek V4, with accuracy within two percentage points. That gap is not a temporary subsidy. ByteDance owns Douyin and CapCut. Kuaishou operates China’s second-largest short-video platform. The training corpora available to Chinese labs are the corpora they themselves run. DeepSeek V4 trains and runs on Huawei Ascend silicon, no Nvidia. Domestic Chinese chips reached 41% of the country’s AI chip market in 2025.

The convergence is happening at both ends

Here is the part many miss. While Western boardrooms argue over decoupling, the open-versus-closed axis those arguments rest on is collapsing. Alibaba broke with its open-source approach in late April 2026, releasing its third proprietary AI model. Meta is reportedly considering keeping its forthcoming flagship — codenamed Avocado — closed, departing from the Llama tradition. The momentum at the open-weight end runs the other way: Chinese models accounted for 41% of Hugging Face downloads between February 2025 and February 2026, compared with 36.5% for US models, with Baidu, ByteDance, and Tencent all expanding their open-source release cadence by an order of magnitude. Both moves are commercial responses to the same observation: open weights drive adoption and proprietary weights capture revenue, and at frontier scale you cannot have both with the same model. Any marketing organisation standardising on a single stack — Western or Chinese, open or closed — is locking in a position the underlying market is actively unwinding.

The architecture: sovereignty-tiered routing

The CMO question is architectural. The model that holds up under stress, from conversations with peers running enterprise AI procurement, is a three-tier routing structure organised around data sensitivity rather than vendor origin.

The volume tier handles high-throughput workloads — classification, extraction, summarisation, translation, content variants, A/B test assets. These run on the cheapest production-quality model that meets the workload’s accuracy bar, frequently a self-hosted open-weight model regardless of origin, with the data never leaving the organisation’s cloud region. Documented savings versus a single-frontier-model default range from 60 to 90% depending on workload mix.

The frontier tier handles workloads where reasoning quality, agentic capability, or licensing position justifies the premium. These run on Western frontier models with commercial guardrails and enterprise contracts — Anthropic, OpenAI, Google. Smaller share of token volume, larger share of governance attention. Both are correct.

The sovereignty tier sits across the other two. It governs which model classes are permitted for which data classes — customer PII, financial data, regulated content, brand-sensitive creative — with the policy written down where finance, legal, and procurement can sign it. As I argued in earlier issues on agentic AI layers, routing tasks across heterogeneous models is itself an agentic problem. The marketing organisations that built that orchestration layer for text workloads have a structural head start when it extends to image, audio, and video.

What this means for the next twelve months

Three concrete consequences. Procurement language should describe model classes — frontier-licensed, open-weight self-hosted, hosted-API third-party — rather than vendor names. Vendor relationships will shift faster than procurement cycles can absorb if the contracts are written around individual companies. The policy that governs which tier handles which workload needs to live in writing, not in the engineering team’s heads, because engineering teams are already making model selection decisions and the policy gap is what creates compliance debt. The architecture investment compounds: the marketing organisations that build the routing layer this quarter will run more creative tests, in more markets, at lower cost than the ones that wait.

The sharpest version of this trade-off is already visible in advertising production, where the cost, quality, and IP triangle has broken in public over the past three months. More on that next week.

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

Las Vegas, April 2026. Two conferences in the same week — Adobe Summit and Google Cloud Next.

For anyone who read The Agentic CMO when it was published in June 2025 — and for anyone who didn’t — these were the days the book’s central claim became falsifiable.

The claim was that marketing automation was ending and marketing autonomy was beginning. That a new role would emerge — call it the supermarketer, the agentic CMO, the orchestrator — and it would not be a single human being doing more work. It would be a human orchestrating a stable of agents, each with bounded authority, each accountable to outcomes the human had not personally executed but had agreed to be measured on.

Eighteen months ago, that was a thesis. In April 2026 it was on stage with named customers behind it.

Adobe demonstrated CX Enterprise Coworker taking a single goal — improve cross-sell by three per cent — and assembling audiences, creative, decisioning, and execution behind it. The human approved the plan. The agents ran it. The platform measured it. Vodafone, on Google’s stage, talked about hundreds of agents in production saving millions of euros a year. Comcast rebuilt the Xfinity Assistant on Google’s Agent Development Kit and runtime. L’Oréal launched a Beauty Tech Agentic Platform on Google Cloud. Adobe Marketing Agent is now generally available inside Microsoft 365 Copilot and in beta inside Claude Enterprise, ChatGPT Enterprise, Gemini Enterprise, and IBM watsonx Orchestrate.

The supermarketer is now a working demo with logos behind it.

The Question Neither Vendor Answered

And this is the part the keynotes were careful not to say out loud.

We now know how to build the supermarketer. We do not know how to manage one.

The hardest unsolved problem in agentic marketing is not model quality. It is not runtime cost — though that one is closer than most CMOs realise. It is not even the data substrate, though Google’s Agentic Data Cloud and Adobe’s Brand Intelligence both acknowledge that agents fail without governed business context. The hardest unsolved problem is organisational.

When Adobe demonstrated CX Enterprise Coworker, the chosen scenario was a three per cent cross-sell improvement. Watch the demo carefully and you can list the open questions on one hand.

Who set the three per cent? At what level of seniority did that goal originate, and against what counterfactual? “Improve cross-sell by three per cent” is a conversational instruction. It’s not a budget. It’s not a forecast. It’s not a promise to a board. The agent can execute against it. It cannot decide whether the number should have been five.

Who carries the cost? Both Adobe and Google disclosed essentially no public pricing for their flagship agentic offerings. Read the silence. It is not oversight. It is an admission that cost-to-serve is not yet predictable, even for the vendors. Inference cost, runtime cost, memory cost, context-storage cost, observability cost, human-review cost — these accrue at different rates against different agents under different workloads. A SaaS seat is the wrong unit of analysis. The right unit is unit economics of work delivered. Most marketing organisations do not have a P&L granular enough to count it.

Who approves the action? Adobe’s example included a human approval step. So did Google’s. Neither vendor explained how the approval ladder works at scale. If you are running fifty agents, you cannot approve every action. If you are running five hundred, you cannot approve any of them and still call it management. Bounded autonomy — the framework I built the second edition of the book around — requires a governance authority. The authority does not yet exist on most org charts.

Who is accountable when the agent underperforms? The CMO? The Head of MarTech? The platform vendor? The agency that integrated it? In the conventional marketing organisation, accountability follows execution. In the agentic organisation, execution belongs to the agent. Accountability does not transfer cleanly upward. Someone has to volunteer for it.

Who decides when to expand the agent’s authority? An agent that is good enough to recommend an audience is, with one configuration change, good enough to assemble the audience. With another, good enough to launch the campaign. With another, good enough to commit budget without a human in the loop. Each step changes the risk profile of the organisation. There is no chief autonomy officer to make the call. There is no controller for the autonomy ledger.

A Small Pilot, A Large Problem

When I ran into this at S&P Global, the symptom was banal. We piloted an agent capable of drafting personalised outbound communications at scale. The technology worked. The pilot stalled — not on output quality, not on integration, not on data. It stalled on a single question: at what volume does this stop being marketing-led and start being compliance-led? Nobody had been hired to answer that. Nobody had been empowered to answer that. The org chart had no row for it.

That was a small pilot in a tightly regulated industry. Multiply it across a Fortune 500 marketing function with eight regions, twelve product lines, three hundred markets, and several hundred agents on the way, and the absence of the governance row stops being an inconvenience. It becomes the rate-limit on everything else.

What Adobe Got Right And Why It Isn’t Enough

To Adobe’s credit, the Summit announcements came closer to the operating-model question than anything I’ve seen from a major vendor. Naming Brand Intelligence and Engagement Intelligence as two separate, opposed reasoning engines inside CX Enterprise is the first time a vendor has admitted at the architectural level that marketing has two clocks running. One engine optimises for engagement economics on short horizons — conversion, lifetime value, return on spend. The other optimises for brand integrity on long horizons — consistency, voice, equity, trust. Without the second engine constraining the first, the agentic stack will eat the brand alive within a quarter.

The architecture is correct. It is not the answer to the question I am asking. Two engines make a platform. They don’t make a job description. The agent that launches the campaign is not the agent that approves the goal, controls the budget, decides when to expand autonomy, or carries the consequence when the customer complains. The org chart still has to exist somewhere — and it will exist in your company, not in Adobe’s.

A Word On The Pricing Silence

A short note on the pricing silence, because I think it’s been under-read. Both vendors are betting that the unit economics of agentic work will become predictable before any enterprise procurement team forces them to publish a rate card. They may be right. They may not. What is certain is that the chief marketing officer who treats the agentic platform as a SaaS line item will, within twelve months, find a multi-million-pound surprise on the inference invoice. The chief financial officer will not be in a forgiving mood.

If you are negotiating either contract — Adobe CX Enterprise, Gemini Enterprise Agent Platform, or both in tandem — separate the costs of model, runtime, memory, storage, context, observability, and human-review workflow. Refuse contracts that bundle them. The bundle is the vendor’s risk transfer to you.

What This Means For Your Next Quarter

The Agentic CMO thesis was that marketing automation was ending and marketing autonomy was beginning. April 2026 is the month I would mark on the calendar as the inflection. The demos work. The customers are named. The protocols — MCP, A2A, Agent Payment Protocol — are open enough that vendor lock-in has shifted shape rather than disappeared. The supermarketer is in sight.

The hardest job in marketing for the next twenty-four months is not building the supermarketer. It is writing the job description.

It is deciding who owns the autonomy ledger, who carries the inference budget, who escalates the goal, who approves the expansion of authority, who answers when it goes wrong. None of that is a procurement question. All of it is an organisational question. None of the keynotes addressed it. None of the vendors will.

The first wave of agentic marketing rollouts will not fail on technology. They will fail on accountability.

The CMOs who win the next cycle are not the ones with the best stack. They are the ones who hired the right person to be accountable for it — and then wrote the org chart that lets that person actually do the job.

Two questions worth sitting with this week.

First: if your CFO asked tomorrow who carries the P&L for agentic marketing across your organisation, could you give a name? Not a function. A name.

Second: if your CEO asked who has authority to expand an agent’s scope from recommendation to execution to autonomous spend — and who has authority to revoke it — could you draw the line on a single page?

If either answer is no, start there. The rest is procurement.

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

Anthropic, Cloudflare, AWS, Microsoft — each has shipped a cheaper, faster way to connect AI agents to your systems. The MCP standard itself survives. Its position in the architecture does not. What was sold as the universal interface for AI agents is being demoted to plumbing — still useful, but no longer the place where decisions get made. The shift happened in under twelve months, this week we explore why.

Focus On: The Token Bill That Forced the Industry’s Hand

The arithmetic is brutal. A single GitHub connection through MCP loads 93 separate command definitions into an AI agent’s working memory — 55,000 tokens of overhead before any work gets done. (Tokens are the units AI vendors meter and bill against — every cost projection in your AI budget eventually traces back to them.) Connect three common business tools — GitHub, Slack, Sentry — and 143,000 of an agent’s 200,000-token working memory is gone before it reads its first instruction. Cloudflare’s published benchmark shows the ceiling: their full developer platform exposed through MCP consumes more than 1.17 million tokens per request. The replacement pattern reduces that to roughly 1,000. A 99.9% reduction.

Anthropic’s engineering team published the canonical case study in November. A workflow that consumed 150,000 tokens under MCP dropped to 2,000 under the new approach. Accuracy improved at the same time — agent success rates climbed from 49% to 74% on identical evaluations. Independent benchmarks from Speakeasy and Scalekit confirm the order of magnitude. Thirty-two times cheaper. Sometimes a hundred.

The procurement question this raises is the one I rarely hear in vendor reviews. What does the connection layer cost in tokens, at our scale, on our actual workflows? Most enterprise AI budgets price the model and ignore the connection overhead entirely. That overhead is now the line item driving the bills nobody warned the CFO about.

The Category Error CIOs Are Still Making

Forrester’s framing is the clean one. MCP is wiring. It moves data between AI agents and the systems your business already runs. Wiring is not governance — the rules about who can do what, with which data, under whose authority. CIOs who treated MCP as a control layer are discovering the control layer has to be procured separately, urgently, and from a different vendor. Cloudflare, Strata Identity, AWS Bedrock AgentCore and a half-dozen specialised gateways are positioning into that gap.

Clutch Security research, surfaced by Strata Identity in January, found the average 10,000-person organisation already runs 3,056 unsanctioned MCP connections — 15% of employees with two each, on average. None governed by enterprise identity systems. Most using personal credentials granting far more permission than the user actually needs. Your CISO doesn’t know they exist. Your finance team is paying for the tokens. That’s not a future risk. That’s last quarter’s invoice.

What This Means at Your Next Vendor Review

The move is straightforward, even if most enterprise teams haven’t made it yet. Treat MCP as backend wiring. Buy the cheaper, sandboxed alternatives every major vendor now ships. Procure governance separately, from a vendor whose actual business is governance — not from the AI platform that sold you the wiring in the first place. Stop signing agent contracts whose pricing model conveniently leaves out the connection overhead.

The vendor signal to watch is who already ships the cheaper pattern by default. Cloudflare’s own MCP server explicitly warns you the old approach costs 244 times more. Anthropic’s Tool Search Tool defers loading commands until the agent actually needs them — cutting cost and improving accuracy in the same move. The vendors who haven’t shipped equivalents are the ones who’ll need to explain why on their next earnings call.

There is a single question worth taking into your next AI vendor evaluation. Did anyone in the room ask what the connection layer would cost, at our scale? If the answer is no, what you signed was a budget approval, not an architecture choice. The two are not the same thing.

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

Most CMOs have never heard of Magistrate Judge Ona Wang. Her 13 May 2025 preservation order has done more to reshape enterprise AI procurement than any vendor announcement of the past twelve months. On 5 January 2026, District Judge Sidney Stein affirmed her ruling: OpenAI was ordered to produce a 20 million de-identified sample of consumer ChatGPT logs to plaintiffs in a copyright case the company’s users are not party to.

Read that twice. A third party sued your AI vendor. Your conversations can become discoverable data in litigation involving that vendor.

Focus On: The Sovereignty Threshold

Three forces are converging, and most marketing leaders are watching only one of them.

The first is legal. The discovery dispute applies to consumer tiers — Free, Plus, Pro, Team — and to API users without a Zero Data Retention agreement. ChatGPT Enterprise, Edu, and ZDR API customers are excluded. That carve-out matters, but it cuts both ways. If your employees touch consumer-tier AI with company data — and they do — you are exposed in lawsuits you’ve never read. The principle that third-party litigation can compel production of user conversations has now been tested and held.

The second is behavioural. In April 2025, ChatGPT shipped a GPT-4o update that turned the model markedly sycophantic — flattering users into agreement. OpenAI rolled it back within days and published a post-mortem. In July, Grok started calling itself MechaHitler on X. Both were vendor-side changes. Both happened without warning. The model behaviour you validated last quarter may no longer be the behaviour running today, and your recourse is a support ticket.

The third makes the first two matter. The performance gap between leading open-weight and closed-weight models narrowed sharply through 2024 and 2025. Stanford’s AI Index puts the Chatbot Arena gap at 8.04% in January 2024 and 1.70% by February 2025 — a near-collapse in thirteen months. As Simon Willison put it in MIT Technology Review’s primer on local models: “A couple of years ago, I’d have said personal computers are not powerful enough to run the good models. And I kept on being proved wrong time and time again.”

Strategic Implications

The economics no longer favour the procurement default. I’m not going to give you a universal break-even threshold — anyone who does is selling you something. Token volumes, latency requirements, fine-tuning needs, and governance overhead make every workload its own calculation. What I will tell you is that the calculation has changed, and most enterprise AI deployments I’ve seen have never been re-run since the contracts were signed.

Closed-source frontier APIs remain the right answer for genuinely hard reasoning, frontier coding, and edge cases where the marginal capability difference matters. They are increasingly the wrong answer for high-volume retrieval, classification, summarisation, and routine drafting — workloads where a self-hosted 14B or 70B open-weight model now performs at adequate quality, with full data sovereignty, predictable behaviour, and no third-party legal exposure.

That’s a hybrid stack. It’s also a marketing operations decision most CMOs have outsourced to whichever vendor sold them their first proof of concept.

The Three Questions

Bring these to your next AI vendor review. Don’t accept the first answer.

If a third party sues our AI vendor next quarter over training data, what happens to our prompts, our customer conversations, our internal data? If our vendor ships a model update tomorrow — quieter, louder, more cautious, less so — what is our recourse, and how long does it take? Of the workloads we run on frontier APIs today, how many would actually break if we swapped in a self-hosted open-weight alternative — and how many are we paying frontier prices for out of procurement default?

The CMOs who answer these in 2026 will spend the rest of the decade running a deliberate hybrid stack — frontier where it matters, sovereign where it doesn’t. The CMOs who don’t will find out which question matters most when their general counsel forwards them a subpoena.

Wang’s order is a procurement story dressed in legal clothing, not a privacy one. Read it before your competitors do.

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

Three years into the gen AI wave, the failure mode has changed shape. It’s no longer “we don’t understand the technology.” It’s “we keep funding the easy 20%.”

PwC put a number on it in their 2026 AI Predictions: technology delivers about 20% of an AI initiative’s value. The other 80% comes from redesigning work — so agents handle routine tasks and people focus on judgment. BCG reaches the same shore from a different boat: 70% of AI’s transformative value depends on changes to people, organisation, and process. Algorithms and infrastructure account for 30%.

And yet boards keep approving the 30%. Let’s delve into this.

Focus On: The ratio that should anchor every 2026 AI budget

McKinsey’s most useful contribution to this conversation isn’t another adoption statistic. It’s a planning constraint. Buried in their analysis is what’s now being called the $1:$3 rule: for every dollar spent developing an AI model, organisations should budget three dollars for change management. Workflow redesign, role redefinition, governance, training, and the unglamorous work of teaching people to trust outputs they didn’t produce themselves.

Most organisations are running closer to $1:$0.30. They have it inverted.

The evidence that this matters is now abundant. McKinsey’s State of AI 2025 tested twenty-five organisational variables against EBIT impact from gen AI. The single strongest predictor wasn’t model sophistication, data estate size, or technology budget. It was whether the organisation had fundamentally redesigned its workflows. Only 21% had. The 6% who qualify as AI high performers — the ones attributing more than 5% of EBIT to AI — were nearly three times as likely to have rewired the work itself.

Here’s what makes this awkward. Deloitte’s 2026 State of AI in the Enterprise found that education — not workflow or role redesign — was the number-one way companies adjusted their talent strategies for AI. Only 34% are truly reimagining the business. The other two-thirds are running training programmes on tools their operating models can’t absorb.

The vendor incentive trap

I’ve been watching this pattern accumulate for eighteen months across industries, and the diagnosis is clear. The vendor incentive structure rewards line items: licences, GPUs, copilots, platforms. The board approval process too often rewards line items. The CFO benchmarks against line items. Few approve a forty-million dollars line item called “rewire the operating model” — but that’s what every credible source now says is the binding constraint.

The contrarian read is even sharper. Jing Ho at Ardonio looked at the same McKinsey data and argued that spending more on AI actually makes you twice as likely to fail, once you correct for survivorship bias. Heavy investors include both the high performers and an enormous tail of expensive failures. The variable that separates them isn’t budget. It’s whether the budget sat downstream of a workflow redesign or upstream of one.

What this means for next year’s plan

Forrester now reports that up to 25% of planned 2025 AI spend is being deferred to 2027 in organisations that couldn’t demonstrate Q1 or Q2 returns. The bill has come due. Boards are losing patience. The deferrals will accelerate.

If you’re sitting in front of next year’s AI budget, the first question isn’t which model, which vendor, or which platform. It’s whether the spend you’re about to approve sits downstream of a workflow you’ve already committed to redesigning — or whether you’re hoping the technology will somehow do the redesigning for you. It won’t. It never has.

The 6% aren’t winning because they bought better models. They’re winning because they were willing to break things their peers wouldn’t touch. By the end of 2026, AI spend that doesn’t sit downstream of an operating-model decision will look exactly like the digital transformation spend of 2018: real money, real dashboards, no EBIT.

The only question is whether your board notices before or after the deferral starts.

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

Jack Dorsey and Sequoia’s Roelof Botha just published a very interesting article on organisational design. “From Hierarchy to Intelligence” traces a line from Roman legions through Prussian staff officers to the modern org chart, and arrives at a striking thesis: middle management exists to route information, AI now routes information better, so middle management can be replaced.

The historical sweep is genuinely compelling. And on the coordination question, I think they’re right. But the essay treats coordination as the whole of what middle management does — and that’s where I part company.

Focus On: What the Middle Actually Does

The essay’s practical application is already visible. Block cut 40% of its workforce in February, restructured around three roles (individual contributors, DRIs, player-coaches), and eliminated permanent middle management entirely. The stock rose 24%. Wall Street loved it.

The “world model” concept — a continuously updated operational picture that replaces the context a manager used to carry — is a real contribution to how we think about coordination at scale. Where I think the argument overreaches is in treating coordination as the entirety of the middle-management function. It’s one of three jobs. It’s the most visible, the most tedious, and — yes — the most automatable. AI can track what’s blocked, who’s building what, where resources sit. Conceded.

The second job is judgment apprenticeship. Middle management is where future executives learn to make decisions with incomplete information, negotiate competing priorities across functions, and recover from being wrong. Northeastern University researcher Ravi Kalluri calls what happens when you automate these moments away the “Judgment Gap Crisis” — algorithmic management creates checkers, not leaders. When Amazon’s warehouse algorithms set quotas that managers can’t override, those managers stop practising the very skills that would qualify them for senior roles. The system improves. The human plateaus.

The third job is ethical safeguarding. An IMD analysis published the same week as Dorsey’s essay described what happens when the judgment layer becomes a rubber stamp: Cigna’s medical directors, employed specifically to review insurance claim denials, signed off on algorithmic decisions in batches — one physician denied over 60,000 claims in a single month, spending 1.2 seconds per case. That’s not management failure. That’s management elimination disguised as management.

The Double Squeeze

Even granting the full coordination thesis, there’s a question the essay doesn’t engage: who leads these companies in 2035?

The pipeline is collapsing from both ends simultaneously. On one side, companies are eliminating the middle-management roles that historically served as the training ground for executives. Gartner predicts that by the end of this year, organisations using AI to flatten their structures will have removed roughly half of their middle-management positions. On the other side, the generation entering the workforce doesn’t want the roles that remain. Only 6% of Gen Z workers say their primary career goal is reaching a leadership position. Seventy percent dismiss middle management as high-stress, low-reward. DDI’s Global Leadership Forecast reports that 77% of CHROs already lack confidence in their leadership bench strength — and that was before the current wave of AI-driven restructuring.

This isn’t a distant concern. It is a compounding structural deficit.

The Flatter Middle, Not the Empty Middle

I don’t think Dorsey is wrong that the middle will get flatter. It will. The status-meeting chairs, the slide-deck consolidators, the human routers of weekly updates — those roles are already gone in well-run organisations, with or without AI. The question is what you build in the space that remains.

My argument: the surviving middle layer must be deliberately reconstructed as a leadership forge. Not fewer managers doing the same job. Fewer managers doing a fundamentally different one — training judgment, connecting functions, providing the ethical override that no world model can supply. Dorsey’s own “player-coach” concept is one of the most promising ideas in the essay — but it needs a structure around it to fulfil its potential. A player-coach without a deliberate leadership development architecture risks becoming a senior IC who mentors on the side. Put that same player-coach inside a system designed to produce the next generation of executives as a primary output, and you have something genuinely different.

Companies that eliminate the middle wholesale will save on headcount for three years and face a succession crisis in seven. Those that redesign it — compressing layers while intensifying the developmental density of the roles that remain — will have a structural advantage that compounds as aggressively as any AI model.

The real test of any AI-flattened structure isn’t whether it ships faster this year. It’s whether, in 2035, anyone inside it will have practised leading long enough to run the company. The centurion wasn’t just routing information to the legate. He was turning raw recruits into commanders. Remove the centurion and you get a more efficient army — until you need a new general.

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

We all run operating system on our machines. macOS and Windows coordinate processes, manage memory, allocate resources — invisibly, continuously, without anyone clicking “approve” on each task. Our laptops had this for forty years.

The enterprise never had the equivalent.

Instead, it has humans. Thousands of them, spending 30% or more of their working hours acting as the connective tissue between applications — copying data from one system, pasting it into another, cross-referencing a third, drafting an email to summarise what the fourth one said. The most expensive middleware in history is us!

That gap is about to close. What’s emerging in 2026 isn’t another chatbot layer or another vendor’s AI assistant bolted onto an existing product. It is something structurally different: an agentic orchestration layer that sits atop the enterprise’s data estate, connects its siloed applications, and — here’s where it gets interesting — delivers a fully personalised, anticipatory interface to every employee. An operating system, in the truest sense, for the enterprise itself.

Focus On: The Rise of the Agent OS

The term “Agent Operating System” entered the lexicon through open-source projects, not enterprise vendors. OpenClaw demonstrated the concept at the individual and small-team level: a skills-based architecture where autonomous agents run on schedules, draw from multiple data sources via the Model Context Protocol, maintain persistent memory across conversations, and deliver outputs through whatever channel the user prefers. Telegram, Slack, email, WhatsApp. The agent wakes up, does the work, and reports back.

In February 2026, a new entrant called OpenFang pushed the concept further. Built from scratch in Rust — 137,000 lines of code compiled into a single binary — OpenFang explicitly positions itself not as a chatbot framework but as an operating system for autonomous agents. Its key innovation is what it calls “Hands”: pre-built autonomous capability packages that run independently, on schedules, without requiring a human to prompt them. A Researcher Hand that monitors your competitors at 6 AM and delivers a scored intelligence report to your dashboard before you’ve had coffee. A Lead Hand that generates and qualifies prospects overnight. A Browser Hand that automates multi-step web workflows with mandatory approval gates before any financial transaction.

The distinction between a library and an operating system matters enormously here. CrewAI, AutoGen, LangGraph — the incumbent Python-based multi-agent frameworks — are libraries. They require a developer to wire things up, manage state, handle failures. An OS does all of that invisibly. It manages agents, allocates resources, enforces permissions, schedules tasks, maintains audit trails. Just as macOS manages your applications, an Agent OS manages your agents. That is a category maturation, not an incremental improvement.

Now here’s the thing. What open-source projects demonstrate at the individual level today is what enterprise platforms will deliver at scale within 18 to 24 months. The hyperscalers and SaaS vendors know this. Microsoft has Agent 365. Salesforce has Agentforce. SAP has Joule. ServiceNow has its AI Agents platform. Nvidia launched its Agent Toolkit at GTC 2026 with 17 enterprise software companies as launch partners, including a reference architecture where Slack becomes the primary conversational interface for Agentforce agents — drawing from data stores in both cloud and on-premises environments.

But there’s a problem with every vendor-specific solution. They are scoped to their own ecosystems. A Salesforce agent orchestrates Salesforce workflows beautifully. It doesn’t know what’s happening in your SAP instance, your Workday environment, or the proprietary analytics platform your data science team built three years ago. Enterprise environments are heterogeneous by design and by history. The orchestration layer that actually works must be cross-stack.

This is where the Model Context Protocol becomes the connective tissue. MCP — open-sourced by Anthropic in late 2024, adopted by OpenAI and Google DeepMind in 2025, and donated to the Agentic AI Foundation under the Linux Foundation in December 2025 — is an open standard that lets AI agents connect to any tool, data source, or other agent through a universal interface. Think of it as what REST did for the web: a structured architectural pattern that replaces fragmented proprietary integrations with a shared protocol. Gartner predicts 75% of gateway vendors will have MCP features by the end of 2026. The protocol already sees 97 million monthly SDK downloads across Python and TypeScript. It won’t be the only standard — Google’s A2A protocol handles agent-to-agent communication — but MCP is rapidly becoming the default for how agents access enterprise data and tools.

The 2026 MCP roadmap, published in March, makes enterprise readiness one of its four top priorities. The gaps are real and acknowledged: no standardised audit trails, authentication still tied to static secrets, undefined gateway behaviour, configuration that doesn’t travel between clients. Multi-tenancy isn’t yet addressed. These are solvable problems, but they explain why enterprise deployment is lagging individual adoption. The protocol works. The governance around it is still being built.

The Data Layer Is the Actual Bottleneck

I want to be direct about something. The conversation about agentic AI is dominated by talk of models and agents — which LLM, which framework, which orchestration platform. That conversation, albeit important, often misses the key prerequisite.

The decisive factor is the data.

An agent is only as capable as the data it can access in real time. If your enterprise data sits in dozens of siloed systems — each with its own access controls, its own schema, its own latency characteristics — then your agents will be precisely as limited as the human who previously had to open twelve browser tabs to assemble a picture of what was happening. The agentic layer doesn’t solve bad data architecture. It amplifies it.

This is why the most consequential announcements in enterprise AI this quarter haven’t come from model providers. They’ve come from data infrastructure companies. Oracle, at its AI World Tour in London on 24 March, announced what it calls “Unified Memory Core” — a converged data engine where transactional data, vector embeddings, graph relationships, and spatial data coexist in a single architecture. Their argument: the database itself should be the operating system for enterprise intelligence. They also launched “Vectors on Ice” — enabling AI vector search directly on Apache Iceberg data lake tables — which unifies search across operational databases and data lakes in a single query.

Google Cloud and Ab Initio made a complementary move in February: joint data connectors, metadata connectors, and agents that enable agentic AI to operate across multi-cloud data estates. The key principle here deserves emphasis. Data can remain distributed and heterogeneous — that’s reality for every large enterprise. But metadata must be unified and standardised. Without consistent lineage, governance, and semantic context, agents cannot reason reliably. They hallucinate. They act on stale information. They make decisions that look plausible and are wrong.

Databricks offered perhaps the most vivid proof of concept with Lakewatch, its new agentic SIEM that runs security directly on the lakehouse. The insight that matters well beyond the security use case: organisations already possess the context they need to act. HR systems, collaboration platforms, application logs, transaction data — it all sits in the lake already. Traditional tools can’t access it without expensive duplication. Lakewatch lets agents correlate across any data source without moving files or switching tools. The architecture, not the model, is the breakthrough.

EY’s CIO Playbook on Agentic AI makes this explicit. It calls for increased focus on data lakes, data streaming and pipeline infrastructure, APIs and middleware, AI-friendly cloud architecture, and data governance as prerequisites for agent workloads. Not nice-to-haves. Prerequisites. And the data tells a sobering story: according to industry research, 70% of organisations discover their data infrastructure is inadequate after launching AI initiatives. The moment of truth typically arrives six months into a project, when a successful pilot needs to scale and the foundational data architecture simply can’t support the workload.

I wrote in a previous issue of Chronicles of Change about the migration of SaaS profit pools — how value is shifting from per-seat licensing toward orchestration and data access. The agentic layer accelerates this shift, but the framing matters. SaaS applications aren’t being replaced. They’re being repositioned as data infrastructure. The CRM, the analytics platform, the content management system — these persist. What changes is how humans interact with them. And that change is the real story.

The Death of the Generic Enterprise Interface

The endgame of the agentic layer isn’t better automation. It is the end of the generic enterprise interface.

When an orchestration layer sits between the employee and the data estate — perceiving context, maintaining memory, understanding permissions and role — every employee gets an interface shaped by their specific needs in their specific moment. Not a dashboard designed for a role. Not a report generated on a schedule. A living, anticipatory surface that synthesises information from multiple sources in real time and presents exactly what that person needs to make their next decision.

Gartner’s Strategic Predictions for 2026 put a number on this: GenAI and AI agent use will create the first true challenge to mainstream productivity tools in 35 years, triggering a $58 billion market reconfiguration. EY predicts that traditional user interfaces will fade as agents interact via APIs, making dashboards and heavy UIs largely obsolete. IDC expects AI copilots embedded in nearly 80% of enterprise workplace applications by 2026, and 85% of executives surveyed believe employees will rely on AI agent recommendations for real-time, data-driven decisions.

Zoom’s March 2026 announcement made this tangible. Its expanded agentic AI platform introduces “AI-first Chat” — custom agents available inside the messaging interface, topic-based summaries, and “For You” prioritisation that surfaces what matters most. The UX explicitly moves toward a single conversational surface that anticipates needs. Salesforce and Nvidia’s GTC reference architecture does the same from the CRM side: Slack becomes the command centre for corporate AI, with Agentforce agents pulling from both cloud and on-premises data stores.

For marketing leaders this is where the transformation gets personal. My ideal agent layer draws simultaneously from our CRM, our analytics platform, our competitive intelligence feeds, our content management system, and our social listening tools. It synthesises them into a single decision surface. Not a dashboard with five tabs. A conversation that already knows what I need because it has the context of my calendar, my current projects, my team’s capacity, and the market signals that changed overnight.

This matters profoundly because the value of enterprise data — particularly structured, high-quality, governed data — increases in the agentic paradigm. Data providers become more critical, not less. The risk isn’t disintermediation. The risk is invisibility. If your data isn’t accessible via MCP or equivalent agent-friendly protocols, it won’t surface in the agentic workflow. It won’t inform the agent’s recommendation. It won’t exist in the employee’s anticipatory interface. For any company whose value proposition rests on data — and that includes the company I work for — agent accessibility is now a first-order strategic priority.

What This Demands of You

The numbers paint a picture of urgency tempered by reality. Gartner says 40% of enterprise applications will embed AI agents by the end of this year — up from under 5% in 2025. The agentic AI market is projected to reach $52 billion by 2030. BCG reports 58% of companies have already integrated agents into operations. KPMG’s Q4 AI Pulse Survey shows enterprises projecting an average of $124 million in AI deployment over the coming year, with half planning $10–50 million specifically for agent security infrastructure.

And yet. Deloitte’s 2025 Emerging Technology Trends study found only 11% of organisations have agentic solutions in production. Thirty-five percent have no formal agentic strategy at all. The gap between ambition and execution is wide enough to drive a transformation programme through — which is exactly what the next 18 months demand.

Three things need to happen.

Your data estate must become agent-ready. That means unified metadata, governed access, real-time pipelines, and semantic context that agents can reason over. If you’re treating your data lake as a reporting backend, you’re building the agentic equivalent of a house on sand.

Your integration architecture must support MCP or equivalent open standards. Every data source, every application, every internal tool that agents will need to access must be reachable through a standardised protocol with proper authentication, audit trails, and permission scoping. Without this, you’ll get what Mirantis aptly calls “Shadow Agents” — unvetted AI running on developer laptops, accessing critical systems without governance. The shadow IT problem, reprised in a more dangerous form.

Your organisation must decide what the employee interface of the future looks like. Not which vendor’s copilot to deploy — that’s a procurement question. The strategic question is whether you’re building toward a world where every employee has a personalised, anticipatory decision surface powered by the full breadth of your enterprise data. Because that is what’s coming. The only variable is whether you design it or inherit whatever your vendors decide to ship.

The enterprises that will define the next era are those that recognise something counterintuitive: the AI model is the least important part of the stack. The data layer is the nervous system. The orchestration layer is the musculature. The personalised interface is the skin — the surface through which every employee touches the intelligence of the entire organisation.

The operating system your enterprise never had? It’s being built now. The question is whether you’ll architect it, or whether it will arrive piecemeal, ungoverned, and shaped by someone else’s priorities.

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

Jensen Huang spent two hours at GTC 2026 talking about watts.

Not FLOPS. Not parameters. Not training benchmarks. Watts — and the rate at which they convert into tokens. For anyone tracking the trajectory of enterprise AI, that inversion of Nvidia’s two-decade messaging hierarchy tells you more than the $1 trillion demand pipeline he projected through 2027. Let’s see why.

Focus On: Why Inference Speed Is the New Binding Constraint

The thesis Huang presented is very simple. Data centres are power-constrained systems. Within a fixed energy envelope, the only variable that matters is how many useful tokens you extract per watt consumed. He even offered a formula for the C-suite: Revenue = (Tokens per Watt) × (Available Gigawatts). It reads like a manufacturing equation — because that’s precisely what it is. Huang now describes data centres as “token factories,” and once you adopt that framing, every infrastructure decision becomes an optimisation problem with a single dependent variable: token velocity.

The numbers highlight the urgency. Nvidia’s forthcoming Vera Rubin platform generates 700 million tokens per second in the same power envelope where Blackwell manages 22 million. The Groq 3 LPX rack — the first product from Nvidia’s $20 billion acquisition of inference specialist Groq — delivers 35x more throughput per megawatt than its predecessor, targeting 1,500 tokens per second for agentic workloads. Nvidia’s own head of AI infrastructure said it clearly: CPUs are “becoming the bottleneck” in agentic workflows. Not GPUs. CPUs — the orchestration layer that coordinates agents, manages memory, and moves data between reasoning steps.

This matters to enterprise leaders beyond infrastructure teams because it redefines the economics of AI deployment. Training was a one-time capital expenditure. Inference is continuous, compounding with every agent deployed, every reasoning chain executed. Huang laid out a token pricing spectrum — from free tier through to $150 per million tokens at the premium end — and that spectrum maps directly to the quality of intelligence your organisation can afford to deploy. Slow inference becomes a revenue ceiling.

What This Means for Your Organisation

I’ve written previously about the transition from marketing automation to marketing autonomy — the shift from systems that execute predefined workflows to agents that reason, decide, and act within bounded parameters. Huang’s GTC presentation supplies the infrastructure economics that make that transition either feasible or prohibitively expensive. If your agentic AI deployments depend on commodity inference — and most enterprise pilots today do — you are building on an architecture that won’t support the workloads you’re planning for 2027.

The implication isn’t that every CMO needs to become a chip architect. It’s that the gap between organisations running agents on premium inference infrastructure and those running on the equivalent of economy class will produce qualitatively different capabilities. Real-time personalisation at scale, multi-agent orchestration across campaign and commercial workflows, autonomous research and competitive intelligence — these require token velocity that current commodity infrastructure cannot deliver within acceptable latency and cost parameters. As one technical analysis of GTC said, the information hierarchy has inverted: performance per watt got the stage, FLOPS got a passing mention.

Huang’s formula — Revenue = Tokens per Watt × Gigawatts — deserves a place in your next technology strategy review. Not as an abstraction, but as a forcing function. Your CFO will eventually ask what your AI agents cost per unit of output. When that conversation arrives, the answer had better be denominated in tokens.

If your AI agents’ response time were measured and priced like a utility — tokens per second, cost per million — would your current infrastructure strategy survive the audit?

Nvidia is building entirely separate chip architectures just for inference speed. What is your organisation doing to ensure its agentic investments don’t hit a latency wall before they hit an ROI?

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

Global data centres consumed 415 terawatt hours of electricity in 2024, with the IEA projecting consumption to surge to 945 TWh by 2030 — more than Japan’s entire annual electricity consumption. Meanwhile, the US Bureau of Labor Statistics projects 81,000 electrician job openings per year through 2034, many just to replace workers retiring out. The disconnect is stark: AI demands infrastructure that requires the precise human expertise we’ve neglected for two decades.

I recently witnessed an interesting discussion where energy industry leaders discussed infrastructure bottlenecks. A panelist described literally searching “almost to the North Pole” for a spare transformer — the only one available in North America. Otherwise, they faced years waiting for an Italian manufacturer. When pressed about root causes, the conversation kept returning to the same constraint: workforce.

This isn’t a supply chain problem masquerading as a skills gap. It’s the opposite. AI’s infrastructure demands expose how systematically we’ve underinvested in the trades that make digital transformation possible. Let’s delve in.

Focus On: The Dependency Chain Nobody Maps

The marketing connection is direct. Your AI-powered personalisation engine depends on data centre capacity. Data centre capacity depends on electrical infrastructure. Electrical infrastructure depends on transformers, transmission lines, and generation facilities. All of which depend on skilled trades working at industrial scale.

Yet most AI strategies ignore this dependency chain entirely. The IEA’s Energy and AI report projects that in the United States alone, power consumption by data centres will account for almost half the growth in electricity demand between now and 2030. The US economy is on track to consume more electricity processing data than manufacturing all energy-intensive goods combined — aluminium, steel, cement, and chemicals. The construction workforce required to support that doesn’t exist.

The Numbers Don’t Work

The Bureau of Labor Statistics shows electrician employment is projected to grow 9 percent from 2024 to 2034 — “much faster than average” — but that’s against baseline demand. The AI boom requires infrastructure expansion far beyond historical trends. The Associated General Contractors of America’s 2026 outlook survey found contractors report growing difficulty finding qualified workers, with data centres and power facilities driving the sharpest demand surge.

Construction costs for gas plants have tripled in a decade, from less than $1,000 per kilowatt to around $3,000 per kilowatt, according to industry participants at the roundtable. Labour scarcity drives much of this inflation. When you can’t find pipe fitters, commissioning schedules slip. When electrical workers are booked two years out, projects queue.

The transmission grid tells the same story. Seventy percent of US transmission lines are over 25 years old, and many key components are approaching end of life — infrastructure that was already stressed before AI demand materialised. Upgrading requires not just equipment but skilled workers to install it. The queue for major transmission projects now extends beyond 2030.

The Corporate Response Exposes the Gap

Microsoft’s 2024 agreement with Constellation Energy to restart the Crane Clean Energy Center — the shuttered nuclear plant at Three Mile Island — it’s admission that grid capacity can’t meet AI demand through conventional market mechanisms. Google signed a deal with Kairos Power for small modular reactors, while Amazon invested in X-energy and acquired a nuclear-powered data centre campus from Talen Energy. Elon Musk’s xAI built the Colossus supercomputer in Memphis using dedicated natural gas turbines when grid power proved insufficient.

This shift to behind-the-meter generation creates new infrastructure demands. On-site generation requires even more specialised expertise — nuclear operators, gas turbine technicians, electrical engineers qualified for industrial-scale installations. The same skilled trades shortages, concentrated at individual sites rather than distributed across the grid.

The transformer story I mentioned earlier illuminates how supply chains break under skilled labour constraints. It’s not that transformers don’t exist; it’s that manufacturers lack the electrical workers to build them fast enough, utilities lack the workforce to install them quickly, and grid operators lack experienced technicians to commission them safely.

The Marketing Dependency Most Miss

This creates a direct business risk for marketing organisations. Cloud service availability depends on data centre reliability. Data centre reliability depends on electrical infrastructure quality. Infrastructure quality depends on skilled installation and maintenance. Marketing technology stacks inherit these dependencies whether you map them or not.

The risk compounds through concentration. Northern Virginia’s “Data Center Alley” hosts the largest cluster of data centres on the planet, with over 300 facilities in Loudoun County alone. If skilled labour constraints slow grid upgrades in that region, it impacts cloud services globally. Your AI marketing tools inherit that geographic risk through infrastructure dependencies you may not have considered.

I’d argue this represents a different category of technology risk than most CMOs are tracking. Not cyber security or vendor lock-in, but physical infrastructure limitations that constrain the AI services marketing increasingly depends on.

What This Means for Your Organisation

Start mapping your AI dependency chains beyond software. Identify which cloud regions host your data, which utilities serve those regions, and what infrastructure investments are planned. Your AI roadmap has physical infrastructure assumptions baked in — stress-test them.

Consider geographic diversification of your cloud workloads, particularly if you’re concentrated in high-demand data centre markets like Northern Virginia or Georgia. The infrastructure bottleneck creates regionalised availability risks that traditional disaster recovery planning doesn’t address.

Evaluate your AI vendors’ infrastructure strategies. Companies building dedicated power generation or maintaining multiple grid connections may offer better availability than those depending entirely on standard utility service. Behind-the-meter generation requires more skilled labour per megawatt but offers independence from grid constraints.

Adjust your AI investment timelines for infrastructure reality. The electrical workers needed to support data centre expansion are ageing out, not scaling up. Your AI strategy may be constrained by factors that have nothing to do with model performance or training efficiency.

Have you mapped the physical infrastructure that your AI strategy depends on — and identified where skilled labour constraints might impact your timelines?

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of my employer.

Seventy-eight percent of American consumers say explicit labelling of AI-generated content is “very important” or “the most important factor” in maintaining trust. That’s from a Gartner survey of 335 U.S. consumers in October–November 2025. Not a loose preference. A stated condition of the relationship.

Meanwhile, 64% of customers told Gartner they’d prefer companies didn’t use AI in customer service at all. And in a separate survey of 1,539 consumers that same autumn, 68% said they frequently question whether the content they encounter online is even real.

Marketing organisations are spending billions to automate content production, personalisation engines, and creative workflows. Their customers are developing an immune response to the output.

Focus On: The Trust Deficit You’re Manufacturing

Last week, I wrote about the AI dividend trap — the risk that efficiency gains get harvested as budget cuts rather than reinvested in growth. That was the supply-side argument. This is the demand side, and it’s worse. Because while you’re debating internally what to do with the savings, your audience is already making decisions about what to do with you.

The Gartner data isn’t isolated. Fifty-three percent of consumers distrust AI-powered search results. Sixty-one percent want the option to turn AI summaries off entirely. People aren’t passively sceptical. They’ve moved to active verification.

The Authenticity Premium Is Measurable

The Association of National Advertisers chose two Words of the Year for 2025 — a first in the award’s twelve-year history. The words: “Agentic AI” and “Authenticity.” One acknowledges what’s happening; the other asks for grace.

Research by Colleen Kirk at the New York Institute of Technology, published in the Journal of Business Research, puts numbers behind the instinct. When consumers believe emotional marketing content was written by AI rather than a human, they judge it as less authentic — even when the content is identical. Engagement drops. Purchase intent weakens.

McDonald’s Netherlands learned this in December 2025. Their AI-generated Christmas ad was pulled within three days of its YouTube debut, after backlash so intense the company first disabled comments, then delisted the video entirely. The production company defended the craft — seven weeks, thousands of takes. The audience didn’t care about the craft. They cared about the provenance.

Coca-Cola’s trajectory is more instructive. After its 2024 AI holiday campaign drew criticism for uncanny human faces, the 2025 version switched to AI-generated animals and leaned into archival Santa artwork as the only human presence. CARMA’s media intelligence data: 10.2% positive sentiment, 32% negative. Consumer opposition to AI in ads dropped from 49% to 46% year-on-year — not enthusiasm, but declining resistance. The gap between what brands tell themselves about AI reception and what the data shows is itself a trust problem.

The Collision Course

Gartner’s January 2026 predictions: 60% of brands will use agentic AI for one-to-one customer interactions by 2028. At the same time, 78% of consumers demand clear AI labelling as a condition of trust. Brands are accelerating toward personalised AI interactions while consumers are demanding disclosure and human provenance. This tension doesn’t resolve itself through better prompting.

Emily Weiss at Gartner coined a term worth adopting: the “acoustic brand” — brands that deliberately shun AI and position around human-made provenance, the way vinyl became a differentiator in music. Every piece of undisclosed AI-generated consumer-facing content is an implicit statement about how much you value the relationship versus how much you value the margin.

I’d go further. It’s a brand positioning decision — and most CMOs are making it by default rather than by design.

What This Means for Your AI Operating Model

Establish a disclosure architecture. Not a legal disclaimer buried in terms and conditions — a genuine framework for when and how you communicate AI involvement. The brands that build this now will own the trust position when regulation mandates it.

Audit your content pipeline for trust exposure. Map every consumer touchpoint where AI generates or shapes the content. Score each for emotional weight. High-trust touchpoints — crisis communications, loyalty programmes, brand storytelling, service escalation — need human authorship or meaningful editorial oversight.

And separate the AI dividend conversation from the consumer experience conversation. Last week’s argument was about governance — protecting AI savings from becoming permanent cuts. This week’s argument is about the consumer’s willingness to tolerate AI in their experience at all. These are two different strategic conversations, and collapsing them into one produces incoherent decisions.

Three-quarters of your audience just told you they want to know when AI made the content. Have you built the architecture to tell them?

Follow me

That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.

Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.

AI can make marketing look cheaper faster than it can make marketing look more valuable. In a CFO’s spreadsheet, that’s a dangerous asymmetry. If the narrative becomes “same work, fewer people, less spend,” marketing doesn’t get rewarded for productivity — it gets downsized for it.

I’ve seen this before. Every wave of automation generates a brief window during which functions can either bank the savings or reinvest them. Once savings get absorbed as permanent cuts, the signal hardens: marketing is a cost centre. That window closes quickly, and it doesn’t reopen.

Focus On: A Governance Choice, Not a Technology Choice

The most consequential decision confronting CMOs right now isn’t which model to deploy or which vendor to select. It’s a governance choice — one that determines whether AI accelerates growth or merely accelerates headcount reduction.

The efficiency narrative harvests savings, shows margin improvement this quarter, and accepts that marketing becomes a managed cost line. The gro…

Read more

Thirty-seven percent of employees who lost their middle manager to restructuring reported feeling directionless, according to Korn Ferry’s 2025 Workforce survey of 15,000 professionals. That finding should give pause to every marketing leader rethinking their org chart. The coordination layer in marketing — the briefing, routing, reporting, quality control apparatus that middle managers have operated for decades — isn’t overhead. It’s critical infrastructure. And infrastructure redesigned carelessly creates more problems than it solves. This week, we look at how this important layer is going to be augmented by AI and how to facilitate this change.

Focus On: The Mismatch Between AI-Ready Tools and Human-Only Org Charts

Marketing departments now sit at the epicentre of AI deployment. The Duke/Fuqua CMO Survey shows 88% of marketers use AI daily, with generative AI adoption surging 116% year-over-year. Yet only 7% of organizations report “great progress” on reinventing the manager role, ac…

Read more

Infrastructure is strategic again. For fifteen years, the cloud migration narrative ran in one direction: shed your servers, embrace elastic compute, convert capital expenditure to operational expenditure, and never look back. In this essay, we explore why I am convinced that AI is now fracturing this narrative.

Focus On: Why AI Reverses the Cloud Calculus

The enterprise cloud thesis was built on a set of assumptions that held remarkably well from 2008 to 2023. Shared compute was cheaper than owned compute. Scale advantages accrued to hyperscalers. Data gravity pulled workloads toward whichever provider held the most of your estate. The economics were clear, and the migration was rational.

AI disrupts every one of those assumptions. Not because the technology is new — enterprises have run machine learning workloads in the cloud for decades — but because generative AI introduces two forces that the original cloud calculus never anticipated: uncontrolled data leakage at the prompt layer, a…

Read more

Most organizations now claim to be “using AI.” The claim is almost meaningless. McKinsey’s 2025 State of AI survey reports that 88 per cent of companies employ AI in at least one business function—yet only a third have managed to scale it across the enterprise. BCG’s parallel research is even more sobering: 60 per cent of organisations generate no material value from their AI investments. The integration model is clearly failing them.

The distinction separating the productive minority from the stalled majority is deceptively simple: are you embedding AI into how your organization thinks and operates, or bolting it onto existing processes like a chrome accessory on an aging car?

Focus On: Five Signs of Genuine Embedding

Having observed this pattern across multiple industries in previous business transformation initiatives, I’ve identified five reliable indicators that an organisation has moved beyond superficial adoption. Let me adapt them for the age of AI:

Workflows have been redesigned,…

Read more

The Confidence Trap: Why Your AI Sounds Right When It’s Wrong

Ask three leading AI models for the title of a researcher’s doctoral dissertation and you will receive three different answers — each stated with total certainty, each entirely fabricated. That finding comes not from an AI sceptic’s blog but from OpenAI’s own research team, who published a paper in September 2025 demonstrating that hallucination is not a defect awaiting a patch but a structural consequence of how every major language model is trained, evaluated, and rewarded.

Focus On: The Incentive Architecture of Fabrication

The paper, authored by Kalai, Nachum, Vempala, and Zhang (arXiv:2509.04664), makes a disarmingly simple argument: language models hallucinate because every incentive in the training pipeline rewards confident guessing over honest uncertainty, at every stage, without exception.

Start with how these models are examined. Nine out of ten industry benchmarks employ binary grading, where a correct answer scores…

Read more

Two million visitors in a few days. Mac Minis sold out across multiple retailers. A social network populated entirely by AI agents, posting manifestos and arguing with each other while their human creators slept. The catalyst wasn’t a product launch from Google, Microsoft, or Anthropic. It was a side project by an Austrian developer named Peter Steinberger, released under the name OpenClaw—and it may be the clearest signal yet that 2026 is the year agentic AI enters personal life.

Focus on: When agents leave the enterprise

For two years, the agentic AI conversation has been an enterprise affair. Salesforce’s Agentforce, Adobe’s Agent Orchestrator, Microsoft’s multi-agent orchestration in 365—these are systems designed for procurement teams and marketing departments, governed by IT policies and wrapped in compliance frameworks. I’ve written extensively in this newsletter about the organisational implications: governance, orchestration, the shift from human-directed to human-governed AI. …

Read more

London, 2014. A black cab to Heathrow ran £70+. Uber charged £25. We all knew, somewhere in the back of our minds, that the maths didn’t add up—that venture capital was paying the difference. We took the ride anyway.

A decade later, Uber charges airport surcharges and takes a cut up to 50%, at times becoming even more expensive than taxis. The economics caught up. They always do.

I’ve been thinking about this pattern as I experiment with generative AI. The tools feel almost free. OpenAI distributes credits generously. Anthropic, perhaps the most “expensive” of all, bundles usage into accessible tiers. Google subsidises Gemini through search. For individual hobbyists and marketing teams testing content automation or personalisation at scale, the marginal cost of experimentation appears close to zero.

It isn’t. And—here’s the part that matters—it won’t stay that way.

Focus On: The subsidy and the meter

The price is wrong (for now)

Let me be clear: the cost decline is real. Independent benchma…

Read more

Forty-two percent of executives now admit that introducing AI is tearing their organisations apart. In an era of unprecedented technological exuberance, this statistic demands scrutiny. We possess models of extraordinary sophistication—systems capable of generating compelling narratives, analysing vast datasets in moments, and personalising customer experiences at scale never before conceivable. Yet the chasm between capability and deployment grows wider by the quarter. Satya Nadella recently articulated this phenomenon as a “model overhang”: a surplus of advanced AI capacity sitting largely unused while marketing departments struggle to integrate these tools into meaningful workflows. The obstacle to marketing transformation is no longer technological. It is organisational, cultural, and fundamentally human.

The Essay: The Adoption Cliff

The Spectacle-Substance Divide

The marketing industry’s relationship with generative AI resembles a prolonged infatuation—intense fascination without m…

Read more

CES 2026, my favourite January event, offered plenty of spectacle—rollable screens, AI-powered mirrors, a parade of concept vehicles. I think this year the real signal was in the side halls and partner demos, where humanoid robots appeared in numbers that would have seemed implausible two years ago. Unitree, Figure, Agility, Apptronik, and a dozen Chinese entrants all showed working prototypes. The implicit message: general-purpose robotics is no longer a research curiosity. It’s a product category in formation.

Yet walk into any factory, warehouse, or retail environment today and you’ll find precisely zero humanoids doing useful work at scale. The gap between demo and deployment remains vast—and understanding why matters for any executive weighing robotics investments.

Focus on: the four bridges to commercial reality

A humanoid must cross four thresholds simultaneously before it becomes economically viable outside controlled pilots.

Safety without fences

Industrial robots have operated be…

Read more