Home / Production Standard
The Production Standard
AI with an on-call rotation.
Anyone can demo an AI workflow. The difference between a demo and a system is what happens at 3 AM, on a holiday, when the model changes under you. Every system we ship runs under these six commitments — in writing, before we start.
The six commitments
What "production" actually means here.
Evals before launch
We write the test before the feature. Every AI system ships with an eval suite — a fixed set of real inputs with known-good outputs — run on every prompt change. If a tweak drops the pass rate, it doesn't ship. You get the eval report, not just our word.
Versioned prompts, logged runs
Every prompt is versioned like code. Every run is logged and diffable — input, model version, output, cost, latency. When something goes wrong, we don't guess: we replay the exact run and find the exact change.
A human approves every word that goes out
Nothing the AI makes reaches the outside world — a publish button, a customer inbox, a deploy — without a human approving it. Approval queues are part of the system design, not an afterthought.
A named engineer on call
Every production system gets a runbook and a named engineer who answers for it. Alerts go to a human who knows the system — not a ticket queue, not a chatbot.
A monthly reliability report
Uptime, error rates, eval pass rates, cost per run — in your inbox every month, in plain language. If the numbers slip, the report says why and what we're doing about it.
The handover test
Your team operates everything without us for a full week before we leave. If they can't, we haven't finished — more docs, more training, until they can. Our bar: your team can run it all without us. That's the bar.
The 90-day metric
Measured at day 90, not at launch.
Launch-day numbers are easy. We measure every system against the success criteria from its pilot, 90 days after handover — real usage, real edge cases, real Mondays.
If the system is short of its criteria, the fix is on us: we diagnose, repair, and re-measure until it meets the bar we signed up to. That's what "built to run" has to mean, or it's just a slogan.
How day 90 works
- Day 1–90monthly reliability reports track the success criteria in production
- Day 90full measurement against the pilot's success criteria, shared with you
- If shortwe fix it on our time until the criteria are met
AI data handling
Your data, answered straight.
The questions every serious buyer asks about AI systems — answered before you have to ask.
We build on enterprise AI providers with contractual no-training-use terms (the same tier that offers zero data retention options). The exact provider is your choice — we recommend, you decide — and it's named in the project docs before we start.
Never. We use providers and API tiers whose terms exclude training on customer data, and we put it in writing in the project agreement.
In your cloud accounts, under your controls, wherever possible. We work inside your repos and your infrastructure — data doesn't take a detour through ours. Anything that must pass through a third party is listed in the project docs with its retention policy.
Yes — NDAs before the first working session, DPAs where personal data is involved. No fuss.
Not yet — we'll say so plainly rather than imply it. What we do instead: work inside your cloud accounts under your security controls, keep audit logs on everything we touch, and follow the practices on this page whether or not a certificate says to.
The standard applies from day one
Start with the free audit.
Seven days, a ranked roadmap, and the same honesty you'll find on this page. Yours to keep — whether you hire us or not.