The demo always works.

Yes, every day, and it's a real part of why forty-one days is the number at the top of this page. The rules are the interesting bit. A named engineer reviews and signs every pull request, and generated code clears the same test, review and security bar as typed code — there is no lower one. Nothing generated reaches production unattended. We don't publish a percentage of AI-written lines because it measures nothing; what we'll show you is the review history. If your contract or your regulator forbids assisted development on your repositories, we'll work that way and tell you what it costs, which is usually about a fifth more.

So has most of the market, and the cause is usually the same. To be fast, the pilot skipped the integration — it ran against an export instead of the live system, on clean documents instead of the ones people actually send, without permissions, and without anyone writing down what a right answer looks like. That isn't incompetence; it's the deal you make to get a demo in three weeks. Our Survey starts by separating what the pilot genuinely proved from what it stepped around. Sometimes the finding is that it was right and needs eight weeks of unglamorous engineering. Sometimes it's that there was never a number underneath it. Both are much cheaper to learn in two weeks than in another year.

Almost never, and we'd be suspicious of anyone who said otherwise. The usual shape is a good team with no slack in it and nobody who has taken an AI feature all the way through evaluation, load and audit before. That path is what we bring, and handing it over is the point rather than a courtesy — the embedded arrangement, one of ours inside your team, exists specifically so your engineers own this in six months without us. If we think a role is genuinely redundant we'll say it plainly, but in five years that has come up twice.

You do, outright, on payment, and the MSA names the awkward parts explicitly: prompts, eval sets, fine-tuned weights, retrieval configuration and index-building code. Those are the assets a vendor can quietly hold hostage precisely because they don't look like software. The single exception is our own toolbox — the eval harness and delivery templates we bring to every engagement — which stays ours and comes to you under a perpetual, transferable licence, so it keeps working after we're gone. We just don't hand a competitor the workshop.

No. Client data goes only to providers under zero-retention, no-training terms, and which providers those are is named in your contract rather than in a policy page we could quietly edit. For regulated work we'll run inside your own cloud tenancy, or against models you host, and accept the performance cost of doing so. Separately, and this is the part people forget to ask about: retrieval is filtered by your existing permissions before anything reaches a model, so an assistant cannot answer out of a document the person asking wasn't allowed to open.

An engineer in Pune, where it is the middle of the working afternoon. That's the whole reason for the second office — nobody does their best incident response at 3 a.m. in their own time zone. Both offices are Commerce Wright employees on Commerce Wright payroll. We don't subcontract the night shift, and you can ask for the on-call rota.

You take everything. Repositories, prompts, eval suites and their full run history, Terraform, traces, credentials, documentation — all of it lives in your accounts and your repositories from day one, not ours. Offboarding support is thirty days at no charge, and we'll sit on a call with whoever replaces us. We've done it. It's not fun, but a client who can't leave isn't a client, they're a hostage.

Eleven years — a manufacturer we started with in 2015, three CTOs ago. Median tenure across the book is 4 years 7 months. We've lost six clients since 2014: three were acquired and absorbed into a parent's engineering group, two outgrew us and built the team in-house, which is a good ending and the one an embedded engagement is designed to produce, and one we handled badly in 2019. Ask us about that one on the call.