What Should a CEO Actually Do About AI?

The Decision Landed on Your Desk, and You Cannot Delegate It

In BCG’s January 2026 AI Radar survey of roughly 2,400 executives worldwide, including 640 CEOs, nearly three quarters of CEOs said they are their organization’s main decision maker on AI, twice the share of the year before. Sixty-five percent put accelerating AI in their top three priorities, and corporations expect to double AI spending in 2026, from about 0.8 percent of revenue to about 1.7 percent (BCG, “As AI Investments Surge, CEOs Take the Lead”). Those figures skew to large global companies, but the percentage is a useful sanity check at any size: on $20MM of revenue, doubling means moving from roughly $160,000 to roughly $340,000, counting tools, data work, training, and outside help.

Notice what climbed to the corner office. Most technology decisions do not, and the ones that do are usually about money. This one is on your desk because it is not really a technology decision. It touches who decides what, what “good” means, which work is worth doing at all, and what you owe the people already on the payroll. Those are the CEO’s questions in any decade. AI made avoiding them expensive.

Why the Pilots Stall

Pick a workflow that one of your people does well. Now write down how they do it, well enough that a competent stranger could produce the same result on Tuesday without calling anyone. Most companies cannot do that for their most important work. The standard lives in a head, the exceptions live in hallway conversations, and the definition of done is whatever the founder nods at.

That is the Invisible Operating System: the real way the company runs, which is nowhere on paper. People tolerate it because they infer. They read the room, notice the boss’s mood, and correct for missing instructions without being asked. Software does not infer. It does what the instructions say, at volume, with no hunch about whether this particular customer is the exception everyone in the office already knows about.

I watched the analog version of this at NFFS, a construction company, long before any of this software existed. Field crews made calls that cost real money. Managers froze on decisions I would have made in thirty seconds. Two people argued over whose job something was while it sat undone. None of that was a talent problem. It was a company where the standard and the decision rights had never been written down, and every one of those failures is one an AI tool would reproduce faster.

So the same experiment produces two outcomes. Where the standard is written, AI compounds: it absorbs the explicit work and people move up into judgment. Where the standard is implicit, it produces confident output nobody can evaluate, someone senior rewrites it, and the pilot dies without a funeral. There is a second half to this, which is the data the tool is standing on: automation pointed at numbers your team already argues about will produce faster arguments. That case is made in The Data Layer Moat. Written standards and trustworthy data are the two things a tool needs underneath it, and most companies are short both.

Your AI Operating Model: Decide, Execute, Judge

Before you buy anything, take one workflow and split it into three columns. Every piece of work in your company already divides this way, whether or not anyone has said so out loud.

Column	What it is	Who holds it
Decide	Choosing what to do, what to stop, what a miss is worth. Trade-offs with consequences.	A named human. Always.
Execute	Producing the work once the decision and the standard exist. Repeatable, describable, checkable.	People, software, or increasingly both.
Judge	Deciding whether the output is good, against a written standard, before it reaches a customer.	A named human, with the standard in hand.

Two rules make this work. First, one specific owner per box. Shared ownership is the oldest way to guarantee nobody owns it, which is why the Accountability Framework demands what exactly, who specifically, by when, and shared understanding, on every commitment. Second, the tool is never the judge. If nobody can say what good looks like for a piece of work, that work is not ready to be automated. It is ready to be defined.

Run this on three or four core workflows and the real answer usually surfaces before you spend a dollar. Most companies find that the execute column is thick with work nobody has documented, and the judge column is one tired person: the founder. That is the founder bottleneck in new clothes, and pointing software at it makes the queue longer, not shorter.

Preparing for AI Agents: What Has to Be True First

An assistant answers when spoken to. An agent takes a goal and acts: it sends the email, updates the record, moves the order. The difference matters to a CEO because an agent is not a tool your team uses. It is closer to a worker your company has hired, and everything that goes wrong with an under-briefed new hire goes wrong with it, only faster and at three in the morning.

Four things have to be true before you point an agent at a workflow.

• The standard is written. Not a demo, not a prompt someone keeps in a notes file. A page describing what done looks like, what the edge cases are, and what to do when the situation is not covered.
• The blast radius is bounded. Name in advance what the agent may do without a human, and what it may never do without one. Money out, contracts, pricing, and anything a customer sees for the first time belong in the second list until you have months of evidence.
• The work is checkable. If you cannot tell within a day whether the output was right, you cannot supervise it, and you will find out from a customer instead.
• A human owns the outcome by name. Not the vendor and not the committee. When the agent is wrong, one person’s number moved, and that person is the one who decides whether it keeps running.

Where to start, in a $3MM to $50MM company, is usually the boring high-volume work with a clear right answer: quote and proposal assembly, accounts-receivable follow-up, RFP and compliance responses, service scheduling, onboarding paperwork, first-pass drafting of recurring reports. What to keep human, for now, is anything that sets a price, opens or ends a relationship, or requires knowing something about a customer that is not written anywhere.

The Five Moves, In Order

1. Tie it to the plan, not to the news cycle. Name the two or three outcomes this year depends on. AI work that does not serve one of them is a hobby with a budget. If those outcomes are not written and owned, the problem is upstream, in strategic planning, not in the tool budget.
2. Pick two workflows, not fourteen. A company running twelve pilots is not moving twelve times faster; it is producing twelve orphans. Choose one with real volume and one that is bleeding the leadership team. The reasoning is the same as how many priorities a company can actually carry.
3. Write the standard down before you automate anything. One workflow, one page, written by the person who does the work best and edited by the person who judges it. Budget a few hours plus a week of arguing about the edge cases. This is the unglamorous half of the job and the half that creates the compounding, and it pays whether or not you ever buy the software.
4. Assign the three columns by name and publish them. Who decides, who executes, who judges, for this workflow, starting this month. Ambiguity here is what turns an experiment into a turf argument.
5. Put it on the scoreboard and review it on a cadence. One number that says whether this is working, reviewed in a meeting you already run. If the people doing the work cannot tell at the end of a day whether they won it, the new way reverts to the old way and nobody reports it. That is the Invisible Scoreboard, and it applies to an AI workflow the same as everything else. Reviewing it means meetings that force decisions, not status theater.

Four of those five moves are operating discipline you should be doing anyway. That is not a dodge; it is the reason the gap between companies is widening faster than the gap between their tools. BCG’s survey points the same way: the 15 percent of CEOs seeing the strongest returns are not distinguished by better software. They made AI a top priority, funded it at scale, and upskilled nearly three quarters of their people.

Five Ways CEOs Get This Wrong

• Buying licenses and calling it a strategy. A seat for everyone is a procurement event. Adoption without a standard produces a thousand private workflows and no institutional capability.
• Handing it to IT, or to the youngest person in the room. Both know more than you about the tools, and neither can set decision rights or standards. Those are yours.
• Starting with the headcount math. Leading with who you can cut buys a short-term number and teaches everyone still there that honesty about their own workflow is dangerous. There is a harder version of this argument at the industry level in The Race We Don’t Want to Win.
• Waiting for certainty. A two-year wait for the landscape to settle is a decision, made passively, that your standards stay in people’s heads for two more years. Almost everything in this guide has value whether or not you ever automate the workflow.
• Blaming the team for a system problem. Before concluding your people cannot adapt, check what the company gave them. Tools, time, and training are the company’s obligation; talent and tenacity are the person’s. The order matters, and it is the 5 T’s of Execution.

Data, Access, and the Policy You Owe Your Team

The first question inside most companies is not strategic. It is “am I allowed to paste this into that?” More than half the executives in BCG’s survey still name data privacy and cybersecurity as concerns, and in a mid-size company the exposure usually is not a sophisticated attack. It is a good employee trying to work faster with a customer list, a contract, or a payroll file.

One page, written by you and your leadership team, closes most of it: which tools are approved, what categories of information may never go into an unapproved one (customer records, employee data, contracts, anything under an NDA, credentials), what has to be reviewed by a human before a customer sees it, and who to ask when the answer is not obvious. Publish it, then make it easy to comply with by approving a tool that is good enough for real work. A policy with no sanctioned option does not stop the behavior. It moves it somewhere you cannot see.

What About the People?

Say the part everyone is waiting to hear. If work in your company gets faster, what happens to the people doing it now? Your team will decide how much to help you based on their read of that answer, and they will read it from what you do rather than from the all-hands slide.

For most mid-size companies the constraint is not too many people. It is too few people operating at the top of their capability, because the standard was never written and the reps were never handed over. Work absorbed by software should move people up into the judge column and into the customer-facing judgment that no tool covers. None of that means the broader contraction is not real; I have argued elsewhere that it is coming. It means the companies that come through it will be the ones whose people were already operating at the top of their capability, not the ones that got to a lower headcount fastest. If moving people up is your intent, say so plainly and prove it with the first workflow. If it is not, do not pretend; people can tell, and a company that lies about this loses exactly the people it needs.

Are You Ready to Start?

Three questions, answered straight, tell you where you are.

• Can you name the two or three outcomes this year depends on, and does each have one owner? If not, start there. Everything else is downstream of it.
• Take your most important repeated workflow. Is the standard written anywhere? If the answer is “everybody just knows,” you have found your first month of work.
• Does anything reach a customer without passing a named human against a written standard? If so, fix that before you add volume to it.

These are the same questions that decide whether a company can grow past its founder at all, which is why this guide keeps landing on operating fundamentals rather than on tools. The longer sequence is in how to scale a company, and the short version of the diagnosis is the Scale Readiness Diagnostic.