Somewhere between "we should use AI" and the feature going live, a question got skipped: how will we know when it's wrong? Every AI feature has a failure rate. Some companies choose theirs, measure against it, and can defend it to a board. The rest find theirs out from customers.
We build and operate a banking platform with AI inside it. This page is the discipline we apply to our own product, offered as an engagement.
Every AI feature was approved on some version of that sentence. Almost nobody who said it had measured the human first.
There's rarely a baseline for the process the AI replaced: no error rate for the people who used to do it, no count of how often the old way got it wrong. Without that number, "better" is just a mood. So the first thing we measure on any engagement is the humans. Sometimes the AI genuinely beats them, and then you can finally prove it. That proof is the difference between "we think it's fine" and "we can show it's fine", and only one of those survives a board meeting.
The sentence also moves responsibility, quietly. A person's mistake is one mistake, caught by the colleague who checks their work and covered by your professional indemnity policy. The model's mistake repeats on every matching request until someone notices, and it's a defect in your product rather than an employee error. A tribunal has already held an airline to what its chatbot promised a customer. Worth asking your insurer whether AI output is covered; the pause before the answer is informative.
If the company carries your name, read that right-hand column again. The work your team ships with AI is signed by you, whether or not you've seen it.
Your team uses AI to build the product, and the product now has AI inside it. Those need different foundations, and most companies have poured neither.
Foundations before houses. The adoption programme on the main page is the first job; this page is the second. Companies keep doing the building without either, and it holds until it doesn't.
Sized to one AI surface at a time: a chatbot, an extraction pipeline, a scoring model. A second surface is a second scope, quoted the same way.
We measure the human baseline the AI replaced, score the surface across the nine dimensions we score our own platform on, put a number on what it costs at real volume, and sit with the owner to choose the failure rate you can live with. You leave with a risk register and a roadmap priced line by line. Start the harness within 90 days and the full audit fee comes off its price.
Book the free call first →An eval suite built from your real cases, tracing and cost telemetry, drift monitors, a regression process for model upgrades, and an incident playbook for the day something gets through. Every deliverable is deterministic: the evals exist and run, the dashboards are live, your team can drive both without us. Quoted once, from the audit's roadmap.
Talk about dates →Providers retire and revise models on their own schedule, not yours. When yours changes, we run the regression before your customers run it for you. Monthly eval and drift runs, a quarterly re-baseline, and a one-page assurance report written for your board.
Available after the harnessOne thing we won't sell you is a guaranteed accuracy number. The model is probabilistic; anyone guaranteeing its behaviour is guessing with confidence. What we guarantee is deterministic: the measuring exists, it runs, and you hear about a problem before your customers do.
And it stacks. Both adoption tracks from the main page plus this harness, quoted as one package from £70,000, with one Friday note and one final report covering the team, the code and the AI.
The standard we hold ourselves to is public: the AI usage policy template from our governance pack is free to copy, no email wall.
Future Proof Consulting is the advisory arm of Future Proof Solutions, the engineering practice that builds and operates the Manx Credit Union platform, live for a licensed financial operator at mcu.im.
AI assists on that platform and people decide: an assistant can walk a member to a decision, and the number of decisions a model may confirm is zero. The harness this page describes is the one that lets us sleep.
Useful first question from us: what does your AI answer, and how many times a day.
