AI Reality Check
An honest read on which of your AI ideas will actually ship.
You bring the roadmap. I score every idea against the same six questions and give you a build order you can defend.
Who this is for
You have five to twenty AI ideas and no way to compare them.
They came from a board meeting, a competitor's website, a vendor pitch and an offsite. Everyone has an opinion about which ones matter and none of the opinions agree.
You don't need a strategy deck. You need someone with no stake in the answer to tell you which of these are real.
What you get
Six deliverables, all yours whatever happens next.
A scored table
Every idea marked real, theatre or flagged, with the reasoning written beside it so anyone can audit the call.
Effort and impact
Where each surviving idea sits, with the number it should move and where that number stands today — so the sequencing argument is visible rather than asserted. Where the number doesn't exist yet, that's the first finding.
A build order
The three to five worth doing, in the order I'd do them, and what to do first.
A kill list
What to stop, and what to do instead — often a rules engine, an off-the-shelf tool, or nothing.
Risk and governance
Where each use case lands under the EU AI Act, and what that means before you build.
A live readout
I present all of it to the people who have to act on it, and answer for every score.
Sample output — illustrative
Flagged means the idea is real but blocked until something changes — usually data that doesn't exist yet, or an owner nobody has named.
Sample output — effort against impact
The sequencing argument becomes visible rather than asserted. Filled dots are real, dashed are theatre, outlined are flagged.
How the scoring works
The same six questions, asked in the same order, every time.
Two gates first. Anything legally prohibited stops there. Anything with no realistic path to the data gets parked rather than scored. Then every surviving idea gets the same treatment, so the answer doesn't depend on who pitched it or how senior they are.
First — two gates
Fail either and it never reaches a score.
Question 1 — scored 0–3
Does this need AI at all?
Would a rule, a query or a lookup table do this better? If so, that's a recommendation worth paying for. Boring technology that works beats a model that impresses.
Question 2 — scored 0–3
Does the data exist?WEIGHTED ×1.5
Not “could we collect it” — does it exist today, at volume, clean enough, with the right to use it. For generative use cases the question changes: is there a test set and governed context, rather than training labels.
Question 3 — scored 0–3
What's the actual upside?WEIGHTED ×1.5
In euros or hours, against a documented baseline, as a range rather than a single confident number. If nobody can size it, that tells you something.
Question 4 — scored 0–3
What breaks if it's wrong?
Wrong answers are guaranteed. If a wrong answer moves money, breaks a regulation or harms a customer, the bar moves and the governance work comes first.
Question 5 — scored 0–3
Build or buy?
Plenty of genuine AI is a commodity now. Building it yourself is a choice that needs defending — and it changes your legal position, since building can make you a provider under the EU AI Act where buying leaves you a deployer.
Question 6 — scored 0–3
Who owns it after launch?
Models drift and quietly get worse. If nobody owns the monitoring and the retraining, you're not shipping a feature, you're shipping a liability.
Finally — one verdict
Scores resolve to one of three.
The rubric is the easy part.
Six questions you could ask yourselves. Most teams don't, and when they do the answers come back wrong — not because the questions are hard, but because of who's answering them.
The person who proposed the use case is scoring their own use case. The executive who announced the AI programme is sitting in the room where you decide it's theatre. The data team says the data exists, because saying otherwise means admitting a migration didn't land. None of these people are lying. They just can't give you a clean answer about their own work.
I have no stake in which way it goes. That's most of what you're paying for.
The rubric is calibrated against my own engagements and against the NIST AI risk framework and the EU AI Act. I'll tell you which parts are established practice and which are my judgment — you should know the difference.
What it costs
Fixed price, so the scope can't drift on you.
Up to ten use cases
- Intake questionnaire
- One working call, about ninety minutes
- Written report and build order
- Live readout
Up to twenty use cases
- Everything in the startup tier
- Stakeholder interviews
- Deeper governance and EU AI Act section
- Readout built for leadership
“I've worked with a lot of product people in AI and R&D. Fanni is the one who asks whether the thing should be built at all. She'd rather kill a bad idea in week one than let it eat a quarter.”
“Working with Fanni was straightforward in a way large-organisation projects usually aren't. She asked the awkward questions early, which meant we didn't discover them late.”
Where it comes from
Built where saying no had consequences.
I built this rubric inside a regulated European enterprise and ran forty AI proposals through it. Every score had to survive security review, legal and a works council.
Saying no there had consequences. That is where the six questions come from.
Most AI assessment frameworks are written by people selling AI. This one was built by someone who had to say no inside an organisation that had already announced it was saying yes.
What happens after
Three options, and I don't mind which you pick.
Take it and go
You have the roadmap, the reasoning and the order. Your team builds it. Plenty of clients do this and it's a good outcome.
No further cost Option 02I build the first one
Scoped, built, handed over with monitoring in place so it doesn't quietly rot after launch.
€850–1,100 / day · €6–12K fixed Option 03I own the roadmap
A fractional arrangement, about a day a week, embedded in your team and in the rooms where AI stalls.
€6,000–8,000 / monthThe fee isn't contingent on you picking two or three. A diagnostic that only pays if it recommends more work isn't a diagnostic.
Questions
Before you book.
Because it takes real work and the answer has value on its own. The kill list alone usually saves more than the fee. If you want a free directional read first, the scorecard takes four minutes.
Then I've saved you a quarter. That's a legitimate result, and it happens.
If you sell into the EU it applies to you. If you don't, I'll cut that section and go deeper elsewhere.
An intake questionnaire, one call of about ninety minutes, and an hour for the readout.
Yes, as standard.
Then we scope it as two rounds, or narrow to the ones with budget attached. I'll tell you which on the intro call.
Your turn
Bring me the list nobody's willing to question.
Twenty minutes, no pitch. If I'm not the right person, I'll say so and point you somewhere better.