Who is this for?
Teams already spending about £1.5k+/month on AI models, APIs, or agents — ideally £3k+. Below that, payback is usually too slow for a paid engagement.
What do you actually do?
We help you cut AI spend without silent quality loss. That means a written note on likely cause and save range, then (if you proceed) evals on your real work, an agreed pass mark before any switch, and a 30-day check.
How do you decide whether the cheaper model is good enough?
We run your current model and the cheaper one on the same examples from your own work, and check every answer on its own against the source and a written list of what it must get right.
An AI grader flags possible serious errors, as defined in writing before the test: a wrong person, number, date or amount; a decision, promise or condition nobody made; a condition dropped, reversed or stretched; or a final position reported as its opposite. Wording, style and a dropped “maybe” do not count.
Graders flag too much, so every flag is checked by a person before it counts. The checker works blind: they see the source and the answer, never which model wrote it or what the grader thought, and every answer is set out the same way whether it was flagged or not. They must first pass a practice test whose right answers are already known, and answers with known mistakes are hidden among the real ones: if the checker misses too many of those, their decisions are not used. They also check a sample of answers the grader passed, to catch errors it missed. Until a flag is decided, it counts against the cheaper model. Only confirmed errors go into the result.
For written work such as summaries and replies, we first run your current model against itself, so we know how much answers vary when nothing has changed at all.
Who sets the pass mark?
You do, in writing, before the test runs, and it never changes after a result is seen. The default is 5%: the cheaper model passes only if we can show, with 95% confidence, that it makes a serious error in fewer than 5 in 100 of your cases.
That needs at least 59 examples with no confirmed serious error, and more if one is found. A stricter limit needs more examples, and we show you how many before you choose. Every result states the limit it was tested against.
Which AI services see our examples?
Only redacted examples, never keys or system access. They go through Vercel AI Gateway to the models being compared (your current one and the candidate) and to the AI model that grades the answers. Flagged answers are checked by a person: us, or an independent checker bound by a written confidentiality and data-processing agreement. Every service, and any checker, is named in your engagement letter before anything is sent, and we use only services whose business terms state that API data is not used to train their models. More in our Privacy Notice.
Is there a free audit?
No. The first paid step is a £250 written diagnostic (credited if you continue). The intake form gets you a short offline note so you can decide whether that diagnostic is worth it.
What does it cost?
- £250 written diagnostic — credited toward full work if you continue
- £750–£1,500 full engagement — evals, switch support, 30-day check
One-off project pricing. Fine if you leave after the savings land.
What if we're not happy with the work?
The diagnostic fee is not refundable. It buys a fixed piece of work: we read your usage, work out where the money actually goes, and tell you what each change is worth. You keep those findings whether or not you go further, so there is nothing to hand back. It is deliberately small so that the risk of trying us is small — and the written note you get before it costs nothing at all.
The balance on a full engagement is different. If, within 14 days of delivery, you do not think the work was worth it, say so and we will refund it.
That is a judgement on the work, not on the savings. We do not guarantee savings. Anyone who does is guessing.
Why should we trust you?
You shouldn’t, on a promise. Trust the process you can see.
We never ask for API keys or live system access — redacted exports only. Before any cheaper setup is measured, we agree a quality pass mark in writing. If it fails that bar, you don’t switch. A full engagement includes a re-check 30 days later, on the setup you tell us is running, using your original examples plus fresh redacted ones from the month.
We do not guarantee savings. We do not invent case studies. The written note before paid work costs nothing; the diagnostic is a fixed piece of work you keep either way.
Do I need a live call on day one?
No. Start with the form. We reply in writing. A short screen-share is optional later if redacted exports aren’t enough.
What do you need from us?
Spend band, tools, and the jobs AI does (from the form). For paid work: redacted last-30-days usage/cost export and real examples of the work (redacted). We do not ask for live API keys or customer PII in the form.
Do you guarantee savings?
No. Savings are estimates until an eval passes your agreed quality bar. If the cheaper setup fails the bar, you don’t switch.
How is this different from a bill dashboard?
Dashboards show spend. We test whether a cheaper setup still passes your quality bar on examples of your real work, and report the hard cases separately, so a good average can’t hide them.
What usually causes the bill?
Common pattern: frontier (or near-frontier) models on every step of a multi-call loop, little or no caching, and no split between easy and hard work. Your note will name the cause for your setup — not a generic list.
How long does it take?
Written note: usually within two working days of a clear intake. Diagnostic and full engagement timing depends on how fast redacted exports and examples arrive.
Where are you based / who do I email?
UK. Contact hello@proofcut.co.uk or use the intake form.
Not financial advice. No savings guaranteed until an eval passes.