Over the past eighteen months, the question "should an executive use AI?" has stopped being a meaningful one — it has turned into "how exactly, and on which tasks?" The difference between these two questions is orders of magnitude: the first resolves itself in the next hype cycle, the second requires managerial reflection that remains rare in open sources.
This is a working note — not a manual, not a guide. Its foundation is my own practice of applying AI in marketing and my participation in the research programme MIT AI Adoption. But it is not a rollout report — it is also my own thinking along the way: where I was wrong, what I revised, what I am still uncertain about. The purpose is to show how the decision-making contour of a CEO and CMO changes when part of the cognitive load is delegated to a model.
Three tasks where AI genuinely shortens the decision cycle
Out of the whole range of AI applications, three classes of tasks are genuinely worth an executive's attention — they yield a verifiable gain in both time and quality. The rest more often creates the illusion of productivity rather than productivity itself.
First — synthesising large volumes of weakly structured information. A transcript of an hour-long call with the team, a thirty-page agency report, six months of project correspondence, a brief spread across ten documents — all of this used to require a dedicated person with the single task of reading and summarising. A model does that work in minutes and does not tire. It does not replace analysis — it removes the preparatory layer. A CMO who walks into a meeting having already read a summary of 200 competition cases from an awards jury operates at a different depth than one reading those same cases on top of a full load — in the evening, at an angle, an hour before the meeting.
Second — generating alternatives and stress-testing your own position. A good managerial habit: before taking a decision, formulate three alternatives and ask them to be picked apart for strengths and weaknesses. Previously this was done by the team — at the cost of time, politics, and often blurred framings, because not everyone will tell the boss directly that an idea is weak. Now a model gives a fast, honest, and disinterested counter-argument. It does not replace the team, but it radically raises the quality of the conversation with them: you arrive at the meeting not with a raw idea but with a first round of discussion already completed.
Third — a draft first pass on tasks with a high "blank page" cost. A strategic memo for the board, a response to a regulator's objection, a letter to a key partner after a difficult conversation — everything where 60% of the time went not into thinking but into resistance to starting itself. The model removes that resistance and returns you directly to substantive work — editing rather than composing from scratch. The first draft is almost always poor on the merits, but it exists — and what exists can be worked with.
Where AI creates the illusion of a decision, not a decision
The symmetric side — tasks where applying AI looks appropriate but in fact worsens the quality of the decision. Being precise here matters, because this is not talked about enough — for some reason only successes tend to get discussed.
Final judicial choice. AI generates well and summarises well, but ranks poorly under conditions where the criteria of choice are incommensurable. "What matters more — protecting brand position or defending the quarter's P&L?" — is not a task for a model. It is a managerial choice, and delegating it to AI produces a formally argued but substantively empty answer. The model will readily write elegant logic for whichever position you have already taken — this is its main weakness and main danger simultaneously.
Work with unique internal data without a retrieval infrastructure. An AI that has no access to your cohorts, your economic models, and your history of decisions operates on the public slice of the market — meaning it gives an answer that any market-side analytics could give. The value of such a consultant is close to zero. It emerges when AI works in conjunction with a corporate retrieval system over proprietary data — but that is no longer "asked ChatGPT", that is a full-scale infrastructure project.
Communication with a person where subjectivity matters. An automatically generated letter to an employee you are letting go, to a colleague you are congratulating, to a partner after a break — does not help; it creates long-term damage to trust, even when the recipient does not know exactly that the letter was written by AI. The difference is legible at the level of tone — and it is destructive to relationships that took years to build.
What changes in the requirements for a CMO
The role of the CMO is visibly shifting — in three directions at once, and this is already reflected in the profiles that owners and boards look for.
From executor to data architect. Where the CMO used to rely on an analytics team and receive prepared cuts, now it is critical to understand yourself: where the data comes from, where attribution bottlenecks are, how your company's retrieval architecture looks. AI in itself gives no advantage — the advantage comes from the quality of the data it is connected to. This resonates with my position on the CMO–CDO partnership that I discussed in detail in my column on ROI marketing.
From delegation to co-thinking. AI does not "do the work for you" — it changes the rhythm of your own work. A good CMO in 2026 knows how to think "in pairs" with the model: formulate the question precisely, receive a draft, see the weak points, come back with a refinement. This is a separate cognitive skill — and it cannot be bought as a subscription to a tool, no matter what you pay for it.
From knowing the stack to knowing the boundaries. The core competency is understanding where AI applies and where it does not. The classification "AI helps / AI hinders" becomes part of the CMO job description alongside budget and team. Without that frame, any tool sooner or later becomes another line item without a clear return.
How my system is built: AI as a second memory
Everything described above is the frame — where to apply, where not. But the frame has a concrete implementation, and I went through it myself. The turning point was not when I started using AI, but when I connected it to accumulated context — Claude with Obsidian.
Until then, AI was a useful but limited instrument. The main problem — the absence of memory: every new chat began from scratch, and every time I had to re-explain the project, the agreements, the history of decisions. That gets tiresome quickly, honestly. The issue turned out to be not the quality of the model but the absence of long-term context — and it is precisely at this point that AI stops being a chatbot and becomes a working instrument.
Now my work and life are structured through Obsidian: separate large areas — main work, philanthropic project, PR, personal projects, family, health — and within each are projects, and within projects the accumulated context: documents, meetings, minutes, decisions, tasks, memory, research, correspondence. Through MCP connectors AI receives access to email, Telegram, files, and other sources — and helps to sort, index, and structure all of it. The result is not a folder of files but a full-fledged digital memory of the project: AI begins to remember the project better than I do.
This is not an abstract metaphor of "an AI assistant" — it is a direct extension of the thesis from the section above: AI without access to your data delivers market analytics, AI with access is a managerial instrument. The difference between these two modes of operation lies not in the model but in the context infrastructure around it.
An important limitation I took away over the year: contemporary models are too inclined to agree with the human — they are simply not good enough opponents. AI today is a poor challenger but an excellent thought partner: it helps you think together well, but it poorly replaces a critically thinking human. The same principle as in the section on final judicial choice: AI enriches thinking, but does not replace verification. That is why AI most strongly amplifies people with high expertise — and is especially risky for junior specialists who cannot yet distinguish a good model answer from a bad one. AI creates the illusion of competence, the error goes unchecked, ends up in the work — and quality quietly degrades.
Another shift I observe in practice: offline meetings are starting to lose ground to online ones — not because they are more convenient, but because online is fully digitised. A Zoom call is recorded, transcribed, indexed, and automatically enters the project memory; an offline meeting is lost context that later has to be reconstructed separately. The more digital the work becomes, the more important the ability to automatically preserve context — and this is no longer about tools, but about what team management will look like in the coming years.
The estimate of time saved that such a system delivers is consistent with the broader market picture.
The quality of this system depends directly on the quality of the information structure — the classical garbage in, garbage out rule holds. Chaotic files, poor naming, disconnected notes — AI inherits that chaos, and an error accidentally recorded in the system's memory then spreads across all projects. That is why the most important skill now is being able to structure knowledge: naming conventions, links between files, cleanliness and freshness of data. Not "using AI" — precisely this.
From instrument to environment: AI-native management
Everything described above is the level of a personal operating system. But the scale of the phenomenon is wider: the same thing is happening at the company level, and it concerns not only the CMO but any top executive accountable for P&L and a team. Fully AI-native organisations do not exist yet — it is too early to talk about them — but the direction is already visible: a world where all work is digitised, every project has accumulated memory, all context is available to the model, and the model itself does not take decisions in the manager's place but becomes an intelligent layer above management — highlighting risks and gaps, helping to hold the cadence of projects, but not taking on the human's responsibility.
Next — a thesis stated more sharply than usual in business media: the gap between those who have learned to embed AI in daily work and those who continue to work the old way will become one of the main sources of managerial advantage in the coming years. And this is not about "the ability to write prompts" — that is already a cliché, if not a banality. It is about a deeper layer of skills: building digital memory, structuring knowledge, managing context, working alongside AI rather than instead of yourself.
Alongside — an honest observation about the quality of public discussion on this topic. Materials about how AI actually changes the work of an executive, not a developer, are almost non-existent. The field is polarised: on the one hand, advice at the level of "ask ChatGPT to write a letter"; on the other, materials targeted exclusively at engineering teams. The applied middle layer — what it looks like for a practising top executive — barely exists. That is why the best way to understand the topic is not to consume content about AI, but to try to automate your own work: only through your own processes can you see where the model delivers real leverage, where it breaks, and where you still need a human.
What I took away from MIT AI Adoption
Since the autumn of 2025, I have been participating in the MIT research programme on AI application in business. The format is not a course but a laboratory: dissection of real adoption cases, discussion with practitioners whose implementation has already passed through the "worked / did not work" cycle. Three observations surprised me most.
Successful cases are not about the model — they are about the process. Companies that extracted measurable business value spent their main effort not on choosing an LLM but on rethinking the process into which AI is embedded. The same customer service — not "we put a bot at the door" but rerouted queries, reassembled the knowledge base, redefined the support team's KPIs. The model is the last 15% of the effort. The first 85% is organisational work that an outside consultant cannot do for you.
Failures are most often — because of the absence of feedback. Companies deployed AI agents but did not build a system in which employees could correct the model on real errors. As a result, the agent produced confident but systematically incorrect answers, and after six months the team quietly returned to manual work. The absence of a feedback loop is the main invisible cause of failures.
Small and mid-sized businesses win faster than large ones. A paradox: SMBs extract operational gains from AI within 3–6 months; large corporations take 12–18 for the same result. The reason is not budget or team quality but the speed of decision-making and simpler approval chains. A good signal for marketing: pilots should be run in the fastest-deciding contours, not "where the budget is".
Practical takeaway for the executive
If I summarise everything I see in my own practice and in the MIT research: AI does not remove any of your managerial responsibilities. It gives you the opportunity to approach them from a higher preparation ceiling — provided that you understand the boundaries of applicability.
The practical frame I use myself and recommend to colleagues:
- Tasks of synthesis, generating alternatives, and drafting — delegate to the model without hesitation.
- Final choice, work with proprietary data, and communication with people where subjectivity matters — keep for yourself.
- Build your own feedback loop — record where the model gave a useful insight and where it erred. Within 2–3 months you will have your own map of where AI works in your context.
- Invest not in subscriptions but in retrieval infrastructure. AI without access to your data delivers market analytics. AI with access is a managerial instrument.
And finally — the least obvious point. In 2026, it is more important for an executive not to "know how to use AI" but to have a position on AI: one that can be defended before the board, explained to the team, and held through the next turn of the hype cycle. That kind of position is worth every hour spent on developing it.
Sources: ¹ McKinsey & Company, The State of AI: How organizations are rewiring to capture value (2025) · ² MIT NANDA, The GenAI Divide: State of AI in Business 2025 · ³ MCP Enterprise Adoption, industry review of the Model Context Protocol ecosystem (2026) · ⁴ McKinsey State of AI (2025); Pew Research Center, US worker survey (October 2025) · ⁵ Microsoft, Work Trend Index (2025) · ⁶ Gartner, agentic AI forecast (press release, 2025) · ⁷ World Economic Forum, Future of Jobs Report (2025)