The AI business case your board will approve

A board approves an AI investment when the proposal answers five questions in terms of money and risk: what does it cost in total, when does it pay for itself, what is the worst that can happen, how do we stop if it disappoints, and who owns it once the vendor has left. A technical plan describes what will be built. An investment proposal describes what the board is saying yes to and how much it can lose at most. The second document gets a signature. The first usually gets a polite request to "work it out a bit further".

This piece is about that second document. It is written for the owner or director of a company with 50 to 350 people who is convinced, or half convinced, and now has to sell it internally. To a supervisory board, to co-owners, to the bank, or to family shareholders who want the company to still be there in twenty years.

We have already written separately about two related subjects, and we will not repeat them here. How to calculate the return on your own figures is covered in Calculating AI ROI on your own figures. Which chapters belong in a roadmap for the board is covered in What belongs in an AI roadmap for the boardroom. This piece sits in between: how to turn a calculation into a proposal that the people around you are willing to approve.

Who actually has to say yes

In a mid-sized company, "the board" is rarely a board in the sense of a listed company. It is often a small group of people, each with their own reason to be careful. If you know that reason, you know which question needs answering first.

The pattern is always the same. Nobody in this list asks which language model you are using. Everyone asks, in different words, what it costs, what it delivers and what happens if it does not work. A proposal that opens with the technology answers a question nobody asked and leaves the real questions for the meeting, where you have to answer them unprepared.

There is something else at play. Mid-sized companies have adopted AI at a fast pace in recent years. Statistics Netherlands (CBS) reports that the share of companies with 50 to 250 employees using AI rose from 20 percent in 2023 to 45 percent in 2025, against 66 percent for companies with 250 or more people. So there is pressure to join in. That same pressure makes a board more alert, because by now everyone has heard a story about a tool that was bought and never used.

A handshake over a signed document. Each person at the table has their own reason to be careful.

The difference between a technical plan and an investment proposal

A technical plan is written for the people who will build it. It describes systems, integrations, data and planning. That document is necessary and it has to be good. It is simply not the document that gets you a decision.

An investment proposal is written for the people who carry the risk. It describes a decision: this amount, this period, this owner, and this condition under which we stop. Anything that does not contribute to that belongs in the appendix or in the technical plan.

The most common mistake is that someone shortens the technical plan and presents that. The board then gets a summary of a document about technology, with a cost line at the bottom. It feels like a proposal, but it lacks the parts that only appear in an investment proposal: the maximum loss, the stop moment, reversibility, and the question of who takes over the work when the vendor is done. You cannot summarise those parts out of a technical plan, because they were never in it.

The five questions a board asks

Below are the five questions, with what a good answer contains for each. The order is deliberate: money and risk first, technology only when someone asks about it.

1. What is the risk, and how large is it in euros?

"There are risks, but we have them under control" is not an answer. A board wants to know how much money is gone at most if things disappoint. Calculate that number and put it up front. It is the cost of the first phase, including the hours of your own people, up to the first stop moment.

Then there are risks that cannot be captured in euros: errors in the output, sensitive data ending up with a vendor, obligations under the AI Act. Name them briefly and in plain language, with what you are doing about each one. A person who checks every rejection before it goes out the door is a measure a supervisory director understands. An accuracy percentage without its consequence is not.

2. Is it reversible?

This is the question that carries the most weight with a careful board, and the one that is answered least often. Reversible means three things. The old process keeps running as long as the new one is not proven. You can switch vendors without starting over, because the data, the instructions and the agreements are yours. And you build nothing you cannot switch off again within a few weeks.

A proposal that makes this explicit changes the conversation. The question becomes: do we dare to take a small step that we can undo? That is a much smaller decision, and it is also the more honest one. Which agreements to make so you are not locked into one party is described in Build, buy or do it together.

3. When will we have earned it back?

A careful board rightly crosses out a payback period based on a benchmark or a vendor white paper. The only number that holds up is one based on your own process, with your own hours and your own rate. How to calculate it is in the ROI piece. For the proposal, what matters most is how you present it: as a range, with the assumptions alongside, and including the costs that are not on the quote. The hidden costs are exactly the line a lender checks first.

State as well when you will measure the payback period again. A board that knows a real number will be on the table after three months has to rely less on your estimate today.

4. Who owns it when the vendor leaves?

This is the question family businesses ask most sharply, and rightly so. An external party can build something good and still leave a problem behind: a system nobody inside understands, which stops working at the first ERP update. The proposal needs to contain a name. A person in the company who is responsible for the outcome, who works on the project along the way and who can make small changes on their own afterwards.

If you cannot fill in that name yet, that is not a detail to sort out later. It is a reason not to submit the proposal yet. What happens when that person is missing is described in What happens when the external party leaves.

There is a figure that often comes up here. The 2025 report by MIT Project NANDA stated that roughly 95 percent of the generative AI pilots it examined had no measurable effect on profit and loss, and that projects with a specialised external party succeeded in about 67 percent of cases, against roughly one third for fully internal builds. That report is not peer-reviewed and rests partly on interviews and self-reporting, so do not base a decision on it. The direction is still useful for your proposal: outside help raises the chance that something works, and that is exactly why you need to arrange up front what remains once that help has gone.

5. What happens if it fails?

A board only approves the entrance once it can see the exit. So write down in advance when you will stop. A stop moment is a date with a measurable condition: "If after eight weeks the system processes fewer than half of the orders without correction, we stop, and the loss is limited to the cost of phase 1."

That feels contradictory. You are asking for money and at the same time writing down when you will give up. In practice it makes approval easier, because the board no longer has to decide on the whole project. It decides on a small, well-defined step with a known maximum downside. And when the stop moment comes, the decision has already been made. Nobody has to have an awkward conversation in month five about a project that has too much in it to stop.

A hand-drawn line chart on paper. A board trusts a number you will measure again more than a forecast.

Build the case from back to front

The usual order is: idea, solution, cost, return. For a careful board the reverse order works better. Start with the damage you can oversee, then work towards the return.

Here is a hypothetical example, with invented but realistic numbers, to show what that looks like.

Suppose: a wholesaler with 120 employees. Orders come in by email, as PDFs and sometimes as a photo of an order list. Four people in the sales office retype them into the ERP. Together they spend about 30 hours a week on it, which at 46 working weeks comes to some 1,380 hours a year. CBS put the average labour cost per hour worked in 2024 at 45 euros. That makes the retyping cost about 62,000 euros a year.

Step 1: the maximum loss. Phase 1 is an eight-week trial in which a system reads the orders and prepares them, while the sales office keeps checking everything. The old process simply continues. Cost: 12,000 euros external and about 180 hours of your own people, together roughly 20,000 euros. That is the most the board can lose.

Step 2: the stop moment. After eight weeks we measure how many orders went through the system without correction. Below 50 percent, we stop. Above that, a second proposal follows for phase 2.

Step 3: the return, calculated conservatively. Suppose the system eventually takes over 60 percent of the retyping work. The saving is then about 37,000 euros a year. After deducting a licence of 400 euros a month, roughly 32,500 euros a year remains net.

Step 4: the total investment and the payback period. In this example phase 1 and phase 2 together cost about 35,000 euros. At 32,500 euros net a year, that is earned back in a little over a year after go-live.

Step 5: the owner. The sales office team lead is the owner, works on phase 1 for six hours a week and learns to adjust the instructions themselves.

Notice what this example does not do. It does not promise a payback period of three months. It does not count quality gains that cannot be measured yet. And it does not ask the board for 35,000 euros, but for 20,000 euros with a stop moment. That is a proposal a supervisory director can defend to their own conscience.

To be clear: this is a worked example, not a figure from experience. Your process, your hours and your rate will give a different outcome. Sometimes better, sometimes worse. A disappointing outcome is useful too: you have learned for little money that you should start somewhere else.

The proposal on one page

A board reads page one. It only reads the rest if page one raises questions. This is the structure of that first page, filled in with the hypothetical example above.

Two lines in this table do most of the work: maximum loss and stop moment. They are deliberately placed above the return. A board that first sees how much can be lost at most reads the return with a calmer mind.

Look at the last line as well. You are not asking for approval of an entire programme. You are asking for a decision on one step, with an agreement about the next decision. What goes into the full roadmap belongs in an appendix and is worked out in the roadmap piece.

What outside figures can and cannot do

A proposal sometimes needs a few outside figures to show that this is not the director's hobby. Use them sparingly and never put them in place of your own figures.

They are useful for showing where the sector stands. The CBS figure of 45 percent AI use among companies with 50 to 250 people belongs here. In trade and logistics, evofenedex reports, based on research by Brthrs Agency and GreenPT, that about 60 percent of companies are interested in AI applications and that only 7 percent have an established AI strategy. That tells a board something useful: interest is everywhere, a plan almost nowhere. A company that now takes a well-founded, small decision is neither falling behind nor gambling.

They are not useful as the basis for your payback period. An average across a thousand companies says nothing about the sales office of a wholesaler in Brabant. A careful supervisory director knows that, and once they spot it, they distrust the rest of the proposal as well.

How to bring the proposal in

The meeting is not where the decision is made. The decision is made in the conversations beforehand. A few practical points.

Discuss the proposal in advance with the most critical person. The aim is to collect the question you would otherwise get unprepared in the meeting. That question then belongs in the proposal.

Ask for the smallest sensible step. A board that says no to a programme of a hundred thousand euros often says yes to an eight-week trial. And a successful trial is a stronger argument for the next step than any proposal.

Let the owner speak for themselves. A team lead who explains which work disappears and what their people get in return is more convincing than a director doing it on their behalf. It also shows that the owner really exists.

Say what you do not know. "We do not yet know whether the system can handle the handwritten order lists, which is why we measure that in phase 1" is a sentence that builds trust. A proposal without uncertainties is not believed.

Agree how you will report back. A short update halfway through phase 1 and a decision point at the end. Then the board knows it will not spend eight weeks in the dark.

What the whole project can cost, including the items nobody mentions in advance, is in What does AI implementation cost for SMEs. Use it as a checklist before you submit the proposal.

A meeting room before the meeting starts. Whoever discusses the proposal in advance with the most critical person walks in better prepared.

Where the AI Readiness Scan fits

The hardest part of this proposal is the groundwork. Which process do you choose, how many hours are really in it, and which opportunities do you deliberately leave alone? That is the work of the AI Readiness Scan. The scan delivers a calculation of the return on your top 3 opportunities, based on your own figures rather than benchmarks, and a roadmap you can take straight into the boardroom. The scan starts from 3,900 euros, depending on scope and modules.

That page also shows an example calculation: 1,200 hours of manual work a year, 68,000 euros saved a year and a payback period of under three months. It is an illustrative example, not a promise. The worked example in this piece comes out at a little over a year, and that difference is exactly the point: the outcome depends on the process, which is why you have to calculate it on your own figures.

If you first want to see where you stand on your own, you can start with the free AI-Foto. It takes ten minutes and gives you a first picture of your starting position, with no obligation.

About this page

This piece was written on 5 October 2026 by the Radical AI team. The wholesaler example is hypothetical and intended as a calculation method, not as a figure from experience or a promise. The average labour cost of 45 euros per hour worked and the figures on AI use by company size come from CBS. The figures on trade and logistics come from the evofenedex article on research by Brthrs Agency and GreenPT. The MIT report is not peer-reviewed and rests partly on self-reporting. The example calculation of 1,200 hours and 68,000 euros comes from our own AI Readiness Scan page and is illustrative. All sources are listed below with a direct link. No client data was used.