Somewhere between the last board meeting and the next one, a director will ask the CEO for the AI plan. Not a deck about AI. A plan. Most CEOs, including good ones, will answer with a list of initiatives, a vendor or two, and a slide that says "exploring." The board nods. Nothing changes. Ninety days later the same question comes back with a sharper edge.
I have sat on both sides of that table. As a CEO I took a company through a Nasdaq IPO and answered to a public board. As a managing partner for a decade I sat on boards and asked the questions. And more recently, as CEO of Product Genius, I built and launched an AI product and rolled it out nationally in under six months, which is the part most AI advisers have not done. Here is what a board actually wants to hear about AI, and how to have real wins in production before you have to say it.
Start with the numbers the board has already read
Directors read the same surveys you do. Two of them frame every AI conversation in 2026. Alvarez and Marsal surveyed a hundred private equity executives and operators: 73 percent expect AI to raise the value of their companies within a year, and 8 percent say it is materially moving EBITDA today, as PitchBook reported. FTI Consulting surveyed 200 fund and operating leaders: 95 percent say their AI initiatives met or beat the business case, and only about a third say their companies use AI in day-to-day operations. MIT's NANDA initiative put a number on the gap inside companies generally: 95 percent of enterprise generative AI pilots deliver no measurable P&L impact.
So the board's real question is not "do we have an AI strategy." It is "are we one of the 8 percent, and how would we know." Answer that, and the rest of the meeting takes care of itself.
The five questions, with the answer that fails and the answer that holds
One. What business problem are we solving with AI?
Before the workloads, answer the question underneath this one. Is AI a must for this company or a should? I ask every leadership team to name where they are going, why, and whether the thing in front of them is a must or a should. If it is a should, the where or the why is wrong, or the team is not committed, and no amount of activity fixes that. If all three line up, the how gets clear and focused. A CEO who tells the board AI is a "should" has answered the board's real question already, and not well.
The answer that fails: "We are exploring use cases across the company." The answer that holds: "AI is a must for us, and here is where and why. We ran a workload sweep. Of the recurring processes in the business, these four can be modeled, carry this volume at this cost per unit, and have a named owner. We are starting with the two that are boring and large." A board can diligence that. It cannot diligence "exploring."
Two. How much of the work can AI actually take, and how do we measure it?
The answer that fails: "Productivity gains across the organization." The answer that holds: "Our operating rule is that any process we can model can be mostly automated, with the exceptions routed to people. In customer service, support and claims that typically takes 30 to 70 percent of the work off the human queue. We measure the share of each workload's volume that no longer touches a person, and we report it as a line, not a story." Klarna's assistant took two thirds of its customer service chats in month one. Independent tests of off-the-shelf agents land near 38 percent. The spread is execution, and the board deserves to know which end you are on and why.
Three. What could go wrong, and who owns it?
The answer that fails: "We have a governance framework." The answer that holds: "Every automated workload has an exception path with an owner, and the failures we expect are these. In engineering, the constraint moves: Faros AI's 2026 telemetry across 22,000 developers found epics per developer up 66 percent, but review time up 91 percent and 31 percent of code reaching production unreviewed. So we are adding product and review capacity before we buy more multiplier." Boards trust the CEO who names the failure mode before it happens.
Four. Who owns AI in this company?
The answer that fails: "The CTO, with a cross-functional committee." The answer that holds: "One person owns the AI agenda at the level where trade-offs get made, and here is why that person is qualified." The test I apply: has this person shipped an AI product, not advised on one; have they run a P&L of real size; can they get an operating team to adopt something it did not ask for. If nobody inside passes all three, say so, and say how you are covering it. That is a more credible answer than a committee.
Five. What will be true by the next meeting?
The answer that fails: "We will have a roadmap." The answer that holds: "Two workloads in production, each with an internal owner and a measured share of volume off the queue, reported here as two numbers. Then the next two." A roadmap is a promise. Two numbers are evidence.
The ninety days between meetings
Here is how to have those two numbers.
Weeks one and two: the workload sweep. List every recurring process in the business. Score each on four questions: can it be modeled, what is its volume and unit cost, what happens to the exceptions, who will own it after the outside help leaves. Rank by cost times confidence. Pick two. Not ten.
Weeks three to ten: production, not pilot. Build the two workloads with the team that will own them, put them into production at a defined share of volume, and route exceptions to a named person. The reason MIT found 95 percent of pilots delivering nothing is that pilots have no owner and no exception path. Skip the pilot stage. A workload in production at 30 percent of volume with an owner beats a pilot at 100 percent of a demo.
Weeks eleven and twelve: the two numbers. Share of volume off the human queue for each workload, and the cost line it moved. Put them on one slide. Then the sweep list, ranked, so the board sees what is next and why.
That is the whole plan. It is not a strategy deck, and it should not be. The board did not ask for a document about AI. It asked whether the company can execute, and two workloads in production answer that in a way no deck can.
What to say if the honest answer is "not yet"
Sometimes the honest answer to question five is that nothing is in production and the organization is not ready. Say that. Then say what you are doing in the ninety days to change it, and who is accountable. Boards forgive "not yet" with a plan. They do not forgive "exploring" for the third meeting in a row.
Where I fit
I work with CEOs on this in three shapes. An AI roadmap, which is the workload sweep with numbers behind every line. An adoption sprint, which takes one or two workloads to production with your team before the next board meeting. And fractional AI leadership, where I hold the mandate across the company without the full-time hire. I am based in Alpharetta, Georgia, work with CEOs everywhere on video, and come on site when the work calls for it. If your board has asked the question and you would rather answer it with two numbers than a deck, the first conversation is forty-five minutes, and if I am not the right person for it I will say so.