Contents · 6 min read
01 / 05 · The decision
What has the opportunity earned so far?
In short
Your board is deciding whether an AI opportunity has earned more funds before the final result can be known. The task is to judge the next commitment, not predict the whole bet.
In your next board meeting, your team may show a working test and ask for more funds while customer use is still unknown. The test proves the idea is more than a claim, but your decision turns on what the next funds allow and how hard that step will be to reverse.
But it does not prove the final result. Customer use may still be unknown, and a small test says little about safe use at scale.
Ask the team to show its weakest claim and the proof still missing. It should also state what result would end the work.
That gives your board a clear view of what the bet has earned so far. You can then set the next proof by asking how hard the next commitment is to reverse.
02 / 05 · Reversal
How hard is the next commitment to reverse?
In short
An easy-to-reverse step needs less proof because your company can stop it within set limits. A step that changes live work or affects customers needs stronger proof first.
Start with the next decision in front of your board. Look at the funds at risk and the access being granted. Then look at any effect on customers.
A test with fixed limits can be easy to stop. Broad access to live systems is harder to undo because it can change real work.
This is why AI agent access should be granted in stages. Read access can show what the data supports, while a live write may need a named approver.
The two common errors sit at each end. Demanding proof of the final result can block all learning, but broad live access commits too much before use and ownership are known.
Once the board knows the next step is easy to reverse, it can fund a much smaller test.
03 / 05 · Small test
What should a small test prove?
In short
Fund only enough to test the question that could stop the bet. The test should prove that question, not claim that the full system is ready.
A team can learn a great deal without building for live use. It can work with approved files or read access, so the source data stays unchanged.
Name the stop question
Have the owner state the one unknown that could end the bet. If the test cannot answer it, do not fund that test.
Use the least access needed
Let the team work with a safe copy or read access when that can answer the question.
Set the result before work starts
Have the owner state what would support more funds and what would stop the work.
A working test is useful, but it does not prove that the full system is ready. AI pilots fail to scale when the task works but management has not changed the work around it.
If the test supports the bet, the next step may carry more risk. That is when your board should ask for stronger proof.
04 / 05 · Stronger proof
When must the evidence get stronger?
In short
Require stronger evidence before a commitment can affect customers or make a material change to live work. That evidence should come from real use or measured results that match the next decision.
Technical proof matters because the tool must do the job. But a sound tool does not show that customers will use it or that the change will help them.
Customer proof matters most when the next funds change what customers buy or use. In that case, delivery work should not become the main source of truth.
The AI-augmented product organization connects customer evidence before build work defines the answer. It also records the proof and the reason to stop each bet.
Measured results matter when the next step changes live work. The team should show what happened within clear limits, while a named person still approves any material change.
The board now has enough to link each rise in risk with a rise in proof. The last step is to release funds in the same stages.
05 / 05 · Funding stages
How should the board release funding?
In short
Release funds one stage at a time, and approve the next stage only after the last one answers the doubt that matters to it. Keep the stop rule visible so an old case does not keep drawing funds.
Use this review with your board:
Have the board name the next decision
State what the funds will allow the team to do and how hard that step is to reverse.
Have the owner show the new proof
Link each result to the doubt set at the last review.
Have the owner test the old case
Show whether the facts behind the bet have changed since funds were first approved.
Have the board compare the other uses of the funds
A capital request should show what will wait, as the capital story briefing sets out.
Have the board release only the next stage
Set the next proof and the result that would stop further work.
Your roadmap should record each bet and the proof behind it. It should also show the cost of the next step and its stop rule.
Do not wait for all doubt to vanish. Proof can arrive after the time to act has passed, so your board needs a set way to fund learning before the final numbers exist.
That set review turns an open bet into a series of clear board choices. The next decision belongs in the memo that asks for the funds.
Fund the next stage
The next step is to put the next funding stage before the board with the weakest claim and the result that would stop further work, backed by the proof earned so far. GRAIL's two-page AI board memo gives that decision a clear place in the next board meeting, so the request can state what the funds allow and what evidence the following stage must earn. It is laid out on the page and ready to copy.
A two-page memo for your next board meeting, laid out on the page, ready to copy. No sign-up.
Questions CEOs ask about this
How much evidence should we require before funding an uncertain AI opportunity?
Require enough evidence to support the next funding decision, not proof of the final outcome. When your team brings a working test to the board but cannot yet show customer use or safe use at scale, the strength of the case depends on what the next funds will commit. Your decision is how much proof that next step requires.
What can we learn without building a production system?
Your team can test a key question with approved files or read access. That can show what the data supports and whether the result merits a larger test, but it does not prove safe live use.
How should evidence differ for easy and hard-to-reverse bets?
An easy-to-reverse bet needs enough proof to justify a small test with fixed limits. A hard-to-reverse bet needs proof from actual use or measured results, plus a named person who approves material changes.
When is customer evidence more important than technical validation?
Customer evidence matters most when the next choice changes what customers buy or use. Technical proof can show that the tool works, but it cannot show that customers want the change.
How should uncertainty affect the amount we fund?
Keep the first amount small while the question that could stop the bet remains unanswered. Increase the amount only after the prior stage answers that doubt and makes the next commitment sound.
What milestones justify the next investment without making us wait too long?
Set each milestone around the doubt that matters to the next decision. Approve the next stage when that doubt has been cut enough for its level of risk, rather than waiting for proof of the final result.
Founder, GRAIL
Johan has spent fifteen years in management consulting and executive leadership, including as Chief Product Officer of a SaaS company. He founded GRAIL to help European leadership teams work out what AI changes at the core of their business. He is the author of The Execution Revolution and the host of the ThinkRoom podcast.
Related reading
Sources
- AI Agent Access to Your Systems: why access should rise only when the prior stage has earned it. (accessed October 29, 2026)
- Why AI Pilots Don't Scale: why a working task does not prove that the wider change is ready. (accessed October 29, 2026)
- What does an AI-augmented product organisation look like?: how bets should carry their evidence and stopping assumptions. (accessed October 29, 2026)
- What's Changing in Your Capital Story: why a capital request should be compared with other uses of the same funds. (accessed October 29, 2026)
- Your Best Quarter Is the Warning: why waiting for final proof can mean acting too late. (accessed October 29, 2026)
Next scheduled review: February 6, 2027