The AI initiative
your board asked for.
The board wants “an AI strategy”. There are four pilots running that nobody can name, a vendor promising magic, an impressive demo built by an enthusiast — and no one who owns getting any of it into production. This is how I’d take it from there to working delivery. Method, tools, and the political reality nobody writes down. The difference from most advisors: I’ve run this transition on myself first — 4,751 commits of it.
Pilot purgatory.
The demo that impressed everyone
Someone built something clever in an afternoon. It was shown to management. Now it’s referenced in every meeting as proof that “we’re doing AI” — and it has never once run in production.
Pilots multiplying, nothing landing
Marketing has one. Finance has two. IT wasn’t told about either. Each pilot proves AI “works”; none has an owner, a budget line, a risk assessment or a path to operations.
The vendor with the magic deck
A supplier promises their platform “handles the AI part”. The price is remarkable in both directions. Nobody in the room can evaluate the claims — which is precisely what the deck is designed for.
The quiet compliance dread
Somewhere between GDPR, the EU AI Act and the sector regulator, someone senior suspects the pilots aren’t entirely legal. Nobody wants to be the one who asks — asking might mean stopping.
It’s not a technology gap. It’s a delivery gap.
The models are better than your organisation can currently absorb — that’s the honest starting point. What’s missing is everything around them: ownership, verification, and a definition of done. A demo needs none of those. A production system is nothing but those.
Most stuck AI initiatives share the same underlying condition: accelerated improvisation. Prompts, wrappers, a frontend that looks like a product, a story about velocity on top. No audit trail, no verifiable done-state, no hard limits on what the AI may do, and no mechanism that can stop a bad output once the system is moving. It looks fast. It is, in fact, parked.
That reflex — and the machinery that enforces it — is what I install. I’ve published exactly how it works: the four mechanical controls and the ten failure patterns every AI initiative eventually meets.
Everyone wants the credit. Nobody wants the ownership.
AI initiatives are political sugar: innovation credit for the board, conference material for the executives, budget justification for every department that claims one. Which is why pilots multiply — a pilot harvests the credit without accepting the operational risk. Production is where accountability starts, so production is precisely where nothing goes.
Below the surface runs something quieter: middle managers who suspect the technology is aimed at their headcount, and specialists who see their expertise commoditised. They rarely object openly. They express it as thoroughness — one more risk workshop, one more legal review, one more architecture board. Treat that as what it is: not obstruction, but fear that hasn’t been given an honest answer. Part of the navigator’s job is giving it one — including where the honest answer is uncomfortable.
And the vendor dynamic: when nobody internally can evaluate AI claims, procurement becomes theatre. The organisation buys confidence, not capability. The single cheapest de-risking move available is having one person in the room who has actually built with this technology and can ask the question the deck was built to avoid.
Four phases, from noise to production.
Chart the water
Before anything is built: an honest map. Not a strategy deck — an inventory with owners’ names on it.
- Every running pilot surfaced, including the unofficial ones — then most of them killed, kindly and in writing. Zombie pilots consume the credibility real delivery will need.
- Use-cases ranked by one criterion: value if it works, damage if it fails. The first production candidate is high-value, low-blast-radius — never the flashiest.
- Data reality checked against ambition: what may legally be used for what (GDPR, AI Act risk class, sector rules) — answered early, so compliance becomes a design input instead of a late veto.
- One named owner for the initiative. Not a committee. Committees harvest credit; owners ship.
One honest pilot
One use-case, built the way production will be built — not the way demos are built. Small, bounded, and gated from day one.
- The hallucination defence installed before the first output: evidence requirements, human-in-the-loop, and the rule that uncertainty is a state — not something the system papers over with fluent text.
- “Done” defined in writing before building: which user outcome, verified how, by whom. A green test is not done. A convincing demo is definitely not done.
- Real users, real data (within the legal map from phase 1), real failure logging. The pilot’s job is to find the failure modes cheaply — not to hide them from the steering committee.
- Weekly honest reporting: what worked, what hallucinated, what it cost. The first time leadership sees an AI initiative report its own failures candidly is usually the moment trust actually starts.
Harden what works
The pilot survived contact with reality. Now it earns production — through gates, not enthusiasm.
- The verification regime made mechanical: checks that block, not guidelines that recommend. Unknown = blocked.
- Adversarial review before go-live: someone whose explicit job is to break it — wrong inputs, malicious inputs, the confident-but-wrong output, the compliance edge case.
- Operating model settled: who watches it, who answers for it, who can switch it off, and what the rollback is. If it can’t be switched off cleanly, it isn’t ready.
- The audit trail a regulator would ask for — built now, while it’s cheap, not reconstructed later under pressure.
Scale what survived
Only now does “strategy” get written — from evidence, not hope. One production success converts more sceptics than any roadmap.
- The pattern from the first delivery becomes the template: same gates, same evidence bar, same ownership model, next use-case.
- Competence transferred deliberately: your people run it, I make myself progressively unnecessary. The exit is designed, not drifted into.
- An honest “not with AI” list published internally — the use-cases examined and rejected, with reasons. It builds more credibility than the successes.
- Governance handed over as a living system — the failure catalogue, the gates, the review cadence — so quality survives my departure.
Published, not proprietary.
The Hallucination Defence
Four mechanical controls and ten real failure patterns — the quality machinery your initiative will need on day one.
Read it →The Workshop
The one-man factory where all of this was learned under production pressure — with the receipts on display.
Inside the factory →A technical case
What disciplined AI-era delivery looks like in practice: stack, gates, security-by-tests, measured numbers.
Read the case →Control · Stabilise · Correct
The delivery method from 18 years of regulated programmes — the same discipline the factory runs on.
The method →Not a strategy. A running system — and a capable organisation.
Somewhere between the demo and production?
That’s exactly the water I navigate. Interim or freelance, via broker or direct. Written, verifiable references on the site; personal referees at offer stage.
This scenario is an illustrative composite built from public regulation and market practice. It contains no confidential client information — a deliberate choice, and the same discretion I bring to an engagement.