AI for CEOs: How to Plan, Pilot and Measure AI in Your Business

Look around your business and a lot of the work still runs through a pair of human eyes. Someone is filling in an Excel model by hand. Someone else is building next week's board deck. Month-end reconciliation is a person cross-checking two systems, and the CRM only stays clean because somebody keeps tidying it. It works, and it's stable - and it's exactly where AI changes what the business can afford to do.

So it's time to start looking at this through the eyes of a plan - one you could act on next week. You form it from the work your senior team already knows is a waste, you pilot and measure the first process, and you put the support in place before anything scales.

Look at AI through the eyes of a plan

The instinct, when people hear that, is the frightening version: AI comes in, the headcount goes out. That's not what I'm describing, and in my experience it's not what works. Integrating AI means studying how your business already operates - the routines it has made stable, the steps a person repeats every week - and adding a little automation to those, one at a time. Used this way, AI is less a replacement than a junior - willing, quick, capable, and happy to take on almost any routine job in the company, as long as someone sets it up well.

The companies getting the most out of it are still a small group. McKinsey's state of AI survey, published in August 2026, puts its "high performers" (those attributing at least 5% of EBIT to AI and calling the impact significant) at about 6% of respondents. Nearly three-quarters of that group say they have fundamentally redesigned workflows because of their AI use. They're also twice as likely to say their senior leaders show commitment to it, and that they have defined processes to measure the impact.

PwC's 2026 predictions describe the alternative, where leadership crowdsources initiatives and tries to shape them into a strategy afterwards. Their verdict is that crowdsourcing "can create impressive adoption numbers, but it seldom produces meaningful business outcomes." Deborah Golden, Deloitte's chief innovation officer in the US, has a name for where a lot of AI initiatives end up: "pilot purgatory". The way I see it, without a plan from the top, enthusiasm from the bottom of the business runs into data it can't reach, a budget it doesn't hold and decisions it isn't allowed to make, which is the pattern in the diagram below.

How to form the plan

The plan starts with the senior team. Each of them needs a strong feel for the work in their area that they already know is a waste - the lengthy report updates, the spreadsheets, the accounts reconciliation, the product copy, the SEO monitoring.

Ask each of them to write down the jobs that are stable, repeated every week or month, and expensive in people's time. That's the waste list. Most of the five jobs above already have a known route, which makes forming the plan mostly a matter of choosing between them.

The table below puts each of them, and market research too, next to the Houtini page that covers it, with what was measured there.

Then cut the list down. PwC's advice is that top leadership picks a few areas, "often where business priorities, evidence of AI's value, and availability of talent and data align." In BCG's 2024 survey of 1,000 executives, the companies it calls leaders pursued about half as many opportunities as their less advanced peers, and expected more than twice the ROI from the ones they chose.

How to pilot and measure

The shape of the job is simple enough to hold in your head: pick one stable, expensive, repetitive process; back one person to automate it, with the design and the reliability taken care of for them; and when it works, bring the next person in. Do that a few times over a pilot and you're not running an AI project - you've got a growing group of people spending less time copying and pasting and more time thinking about the business.

Before the pilot starts, take a baseline, which is the set of numbers for the process as it runs today. Skip it and you'll end up arguing about whether the pilot worked. Once the pilot is running, take the same numbers again. There are four of them:

• how long it takes
• how often it runs
• what it costs in people's time
• how often it goes wrong

McKinsey looked at twelve adoption and scaling practices in its March 2025 survey. The one "with the most impact on the bottom line" was tracking well-defined KPIs for the AI work. A KPI (key performance indicator) is one of the handful of numbers you agree in advance to judge the work by, and for your pilot, those are the four baseline numbers above. Fewer than one in five respondents said their organisation was tracking KPIs for its AI work.

Anthropic's customer story on Novo Nordisk has the kind of before-and-after numbers you want: a development team of 11 people took the time spent producing clinical study documentation from more than ten weeks to ten minutes, and cut the resources needed for device verification protocols by 95%.

When the second set of numbers is in, decide whether to keep the new process, change it or stop it. Stopping counts as a result. After that, take the next item on the list, and the next person. AI turns out to be the most willing junior you'll ever take on. The only thing it needs from you is the decision to give it the room.

What it will need to support you

A pilot like that needs four things around it: sponsorship from the top, a small team with a mandate, an MCP gateway for the teams' AI tools and, if you want it, outside help to get the first pilot moving.

Sponsorship from the top

If the CEO backs their team to analyse the need for some procedural automation, congrats - you've solved the first barrier to entering the world of AI.

An internal hire who walks in cold without sponsorship from above isn't going to make these steps forward very quickly, in my experience, as wading through the politics and objections can be exhausting and dispiriting. Pure consulting that produces a strategy deck doesn't survive contact with the business either (although it can be very expensive, depending on the size of the consultancy you hire).

In McKinsey's March 2025 survey, of the 25 attributes it tested, a CEO's oversight of AI governance was the one most correlated with higher self-reported bottom-line impact. That's why I'd tell any MD that the decision about which problems get worked on, and who gets the budget and the people to work on them, sits at the top.

A small team with a mandate

The team is five to ten people, working as a new business inside the business that disrupts the old business. Their remit is narrow. They pick a problem that's simple but expensive, and ship the fix. Then they teach the people whose work has been changed how to operate the new tools with judgement, and build their experience in making and testing small procedural changes to get the answers they need before delivering the work. Then they move on to the next problem.

Most of the effort in that loop goes on people rather than software. BCG's 2024 research has a rule of thumb for it: the leaders put 10% of their resources into algorithms, 20% into technology and data, and 70% into people and processes.

Agentic AI, where the system takes a chain of actions on its own, is the second move, not the first. If you skip the boring fundamentals first, you're cooked.

An MCP gateway for the teams

If you lead the business, you only need to know this exists and why. The detail is for whoever runs IT, and I'll pick up with you again in the next section.

MCP (Model Context Protocol) is how an AI model like Claude connects to a business's tools and data. An MCP gateway is one controlled front door that the teams' AI tools go through to reach those systems. In Cloudflare's version, the controls set which users or groups may reach each MCP server, and which tools inside it they're allowed to use. Its portal logs the requests and tool executions, so an administrator can see what was done. Microsoft's gateway handles sign-in through Entra ID, and Docker's runs each server in its own container with restricted privileges, network access and resource usage. Spend sits in a separate layer: budgets on model calls belong to the AI gateway (one checkpoint that every model call goes through, with a limit set per team), which the guide to AI for CFOs covers.

I moved the MCP servers on my workstation behind Docker's gateway in August, and ten servers have run through it since 17 August, with me as the only user. Between 7 and 9 August, the clean-up job on that machine was killing 64, 93 and 53 orphaned processes a day; since 17 August it's killed one or none. For a large electronics manufacturer I work with, the work has been to develop alternative routes via a managed MCP gateway.

Outside help, if you want it

My advice is, if you're bringing in outside consulting, execute on the AI Audit. That starts with a feasibility audit, then a three to six month engagement with the handover built in.

Where to start

This week, put the one-page plan below in front of the senior team. Ask each of them to bring their waste list to the next meeting. Pick one process from those lists, name the owner, and take the baseline before anybody touches it.

AI plan: one page for the senior team

1. The waste list: each of us names the work in our area that is stable, repeated every week or month, and expensive in people's time.
2. Pick one: a stable, expensive, repetitive item that is easy to measure.
3. Owner: one named person, backed by the leadership team, with time set aside.
4. Baseline: how long it takes today, how often it runs, what it costs, how often it goes wrong.
5. Pilot: build it with the owner, check it against the baseline, and fix what breaks.
6. Measure: the same four numbers after the pilot, reported to the leadership team.
7. Decide: keep it, change it or stop it, then pick the next item from the list.
8. Support: a sponsor at the top, a small team with a mandate, an MCP gateway for the teams' tools and data, budgets on AI spend, and outside help if we want it.