You're reading The AI Directive — practical AI intelligence for business leaders. Subscribe here if someone forwarded this.
THE SIGNAL
Discipline Is Not The Same As Delay
AI has created an awkward problem for leaders who like proper investment discipline.
The old questions still matter.
What is the business case? What is the payback period? What cost comes out? What value goes in? Who signs it off? How do we know whether it worked?
Those are not bad questions. Organisations that avoid them will spend a surprising amount of money creating very little evidence.
But there is another trap.
If every AI experiment has to survive the full machinery of a traditional business case before anyone is allowed to learn anything, the process will be too slow for the technology and too heavy for the certainty available.
That does not make AI special in a mystical sense.
It also does not make the problem new.
This is where some of the AI conversation gets carried away with itself.
Businesses have had to fund uncertainty before.
Product innovation works like this. So do innovation labs, test markets, prototypes and pop-up shops.
If you want to test a retail concept, you do not necessarily build the full shop, connect every corporate system, implement the complete EPOS estate, hire the permanent team, fit out the whole operating model and then discover whether anyone wants to buy the thing.
You might run a pop-up.
A real enough environment to learn something useful. A limited enough exposure that failure does not become a capital project with lighting.
Software learned a similar lesson through Agile. Do not spend six months pretending you know everything. Maintain a backlog. Run the sprint. Release the smallest useful thing. Learn from the evidence. Adjust.
None of that means "no discipline".
It means the discipline moves closer to the work.
The question is not whether AI should be exempt from investment logic.
It should not.
The question is which investment logic fits the level of uncertainty.
Last issue, I argued that the Board does not need a list of AI tools. It needs a portfolio view: use cases, value, risk, owners, evidence and next decisions.
This is where that portfolio needs funding logic.
The mistake is treating AI investment as either a free-for-all or a fully formed capital project.
There is a better middle ground:
staged funding, explicit evidence, and fast decisions.
Start small. Define the work. Set the value hypothesis. Name the owner. Set the control rules. Decide what evidence will count. Run the test. Then make the decision: kill, narrow, rescue, scale or hold.
This matters because AI is unusually visible.
Boards are asking about it. CEOs are reading headlines about huge investment, huge waste and huge competitive advantage. CFOs are trying to stop cost from leaking everywhere. Technology teams are being asked to move quickly without creating a mess. Functional leaders are trying to show they are doing something meaningful without pretending every use case has a neat five-year return model.
The old metrics can distort the decision.
If headcount is the forbidden number, organisations will often spend more cash on temporary resource, consultants or capital projects just to keep permanent roles off the books. The P&L looks cleaner. The real economics may not.
AI can fall into the same trap.
A small, sensible investment gets blocked because it looks like fixed overhead. A much larger platform spend looks more acceptable because it sits in a different bucket. The accounting shape starts driving the operating decision.
That is not investment discipline.
It is a governance problem dressed up as financial control.
In practice, weak AI ideas drift because nobody wants to sound anti-AI. Strong ideas get underfunded because they cannot prove mature ROI before they have been allowed to generate evidence. Expensive platforms get bought because they feel more strategic than small tests. Cheap experiments multiply until nobody knows what is happening.
That is not innovation.
That is portfolio leakage.
It is also not "write a 30-page business case before testing whether the assistant can produce a useful weekly report".
The answer is staged confidence.
Spend a little to learn whether the use case is real.
Spend more only when the evidence improves.
Stop when the evidence does not.
Scale when the value is repeatable and the controls are good enough.
FIELD NOTES
The Three-Gate AI Investment Model
If I were setting up AI investment discipline, I would use three gates.
Gate 1: Is this worth testing?
Do not ask for full ROI. Ask whether there is a real business problem, a credible AI role, a named owner and a low-risk way to learn.
Required evidence: use case, expected value, owner, affected process, data boundaries, test effort, review date and kill criteria.
Gate 2: Is this worth improving?
After the first test, ask what changed. Did cycle time improve? Did quality improve? Did people keep using it? Did the control model hold?
The answer is often not scale or kill. It may be narrow or rescue. That is fine. A pilot that becomes narrower is often getting healthier.
Gate 3: Is this worth scaling?
Scaling needs a higher standard: repeatable value evidence, named operational owner, support model, training plan, permission controls, cost model, benefit tracking and clear decision rights.
The trick is not to put Gate 3 paperwork in front of Gate 1 learning.
That is how you make an agile-looking process behave like wet concrete.
THE SHORTLIST
1. AI still needs investment discipline. It just needs discipline that can learn quickly.
2. Do not demand full ROI certainty before a small test. Do demand a value hypothesis, owner, evidence plan and kill criteria.
3. Scaling AI is a different decision from testing AI. Treat them as separate gates or you will either over-control learning or under-govern risk.
ASK ME ANYTHING
"How do I justify AI if I cannot promise FTE savings?"
— The CFO question, correctly asked
Do not force every AI case into an FTE-saving story.
Look for cycle time reduction, less rework, faster decision preparation, better consistency, lower risk, fewer handoff errors, higher throughput or reduced dependency on scarce expertise.
Then ask the harder question:
What will we do with the capacity or quality improvement if it appears?
AI value is not created when the tool produces output. It is created when the operating model changes enough for the organisation to benefit.
ONE THING
The right question is not "what is the ROI before we start?" It is "what evidence will justify the next pound?"
FROM THE EDITOR
If you only do one thing this week, take one AI idea and write its Gate 1 case.
One page.
Problem, owner, value hypothesis, test design, controls, evidence and kill criteria.
Next week: when AI starts taking action, governance has to move from policy into operations.
See you Tuesday.
— Toby