How we scope, build and hand over: the engagement models, the delivery process, and the things we say no to. Written down so you can hold us to it.
Four stages. The first one exists to find out whether the other three are worth doing.
We sit with the people doing the work, not only the people who own it, and write down how the process actually runs including the parts nobody documented.
Content is audited before anything is built. We connect to the systems that hold the answers, and build the evaluation set from your own history before tuning anything.
The interesting work is the ten percent the happy path ignores: the refund that is out of policy, the document that is missing a page, the customer who is already angry.
Documentation, a walkthrough with the people who will run it, and an exit path. If you want to take it in house, nothing about how we built it should stop you.
Pick the shape that fits where you are. Most clients start with an audit and decide afterwards.
We sit with the people doing the work, map where time actually goes, and come back with a ranked list of what is worth automating, what is not, and what it would cost. The output is yours whether or not you build it with us. Roughly half of every audit ends with us telling a team that one of their ideas is not worth doing yet, which is the point. Best for: teams who know AI should help somewhere but cannot yet name the workflow.
Two weeks, fixed fee, delivered as a written planOne workflow, scoped in writing before anything is agreed, priced as a whole rather than by the hour. You know the number, the date and the acceptance criteria before we start. Change requests are quoted separately rather than absorbed quietly, so the scope does not drift and neither does the invoice. Best for: teams with a known workflow who want it shipped without an open ended contract.
Four to twelve weeks, priced up front against a written scopeAn engineer joins your standups, your repository and your ticket queue, and works as part of your team rather than across a contract boundary. You direct the work. This suits teams who already know what they want built and need someone who has shipped retrieval, evaluation and agent systems before rather than someone learning on your codebase. Best for: product teams who have the roadmap and the context, and need the AI depth.
Monthly, minimum two months, part time or full timeA system that answers customers is not finished when it ships. Content drifts, edge cases arrive, model providers change their pricing and their behaviour. We monitor quality against the evaluation set, retune retrieval as your content changes, watch spend, and send a monthly report that says plainly what improved and what did not. Best for: teams with a live system and nobody on staff who owns it day to day.
Monthly retainer, cancellable with thirty days noticeYou own the client relationship, the contract and the margin. We design, build and support the system behind your brand. Proposals, documentation and dashboards all carry your logo, and non solicitation terms protect the relationship. Some partners present us as their internal AI division, others introduce us openly as their delivery partner. Best for: agencies, MSPs and consultancies whose clients are asking for AI.
Per project, at partner rates, under NDAA short list, because it is the most useful thing on this page. Every item here has cost us work.
If your documentation contradicts itself or is badly out of date, a retrieval system will faithfully surface the contradictions faster. We will tell you that the first sprint is content work, and that it is not AI work, and some teams do not want to hear it.
Tuning is the most over-prescribed answer in this field and it is an expensive way to not fix a retrieval problem. We check the cheap options are genuinely exhausted before quoting the expensive one.
An agent that can issue refunds, cancel orders or send email on your behalf needs a written boundary: what it may do unattended, what needs a human, what it may never do. If a client wants that boundary left vague, we decline.
We will not sign up to a number we do not believe. We would rather set a target from your own ticket mix and beat it than accept an industry figure and miss it.
If there is no way to tell whether the system is working, there is no way to tell when it stops. A project without an evaluation set is a project that degrades quietly, and we will not take one on.
An audit is two weeks and ends in a written plan. A first workflow is typically four to twelve weeks depending on how many systems it has to touch and what state the content is in. We would rather quote twelve and deliver in nine than the other way round.
Two weeks, fixed fee, ends in a written plan you own whether or not you build it with us.