Skip to content

AI solutions

AI integration

A model built into what already works: search across your knowledge base, suggestions inside your CRM, incoming requests handled. No rewrite required.

Get an estimateHow the work runs
Reply to your brief
within 24 hours
First call
30 minutes, no commitment
First working version
in 3–4 weeks

Process

What happens, and when

The work is split into stages, and each one has a named result. You see it yourself rather than reading about it in a report.

  1. Days 1–4

    Where it goes in

    We look at your product rather than at the model: where people lose time in it, and which screen they are looking at while they do. That yields a list of places worth wiring a model into — usually two or three, and we start with the one a single flag can switch off.

    You get — A shortlist of call sites, first one chosen

  2. Week 2

    Entry point and layer

    We work out how to get into the existing code: a queue, a webhook, one more endpoint, a field in the interface. The model call is wrapped in its own layer, so it can be switched off, cached or replaced without touching the rest. Latency gets counted here too: model answers arrive in seconds, and the screen was built for fractions of one.

    You get — A code boundary and a separate call layer

  3. Weeks 3–4

    Integration and failures

    We write the integration and test it by failure: the model is down, slower than the timeout, returns the wrong shape, hits a rate limit. In every case the product must behave as it did before — one suggestion short, but with no error on screen. The tests are written around that, not around a good answer.

    You get — An integration with a fallback for every failure

  4. Before it goes on

    Behind a flag

    It goes live behind a flag for a share of users, and we watch three numbers: latency, spend, and how often the suggestion is accepted. A suggestion everyone dismisses costs more than it looks, and removing it now is cheaper than arguing about it in six months.

    You get — A flagged suggestion with latency and spend measured

  5. After it is on

    Keeping the layer

    The model layer needs keeping: providers ship new versions and retire old ones, and answers shift without your involvement. Once a month, the same three numbers; when a cheaper or faster model lands, moving to it costs one run of the checks.

    You get — A swappable model layer with monthly figures

What the work covers

  • Where it goes in

    a list of the places a model would take manual work away, each weighed by the effort it takes and by what happens if the suggestion is wrong

  • The way into the code

    a background job, a webhook, one more endpoint, a new field on an existing screen — we take a boundary the project has already drawn and leave the rest alone

  • The model call layer

    the calls live in one place, the prompts sit as text beside the code, and the list of fields that go into a request is set here too — what leaves your perimeter stays visible

  • Behaviour on failure

    timeouts and retries, a fallback path for every call site, a queue for background work, and tests that trigger those failures on purpose

  • Latency and spend

    a cache over repeated requests, a monthly ceiling, a token counter per call site, and an alert when spend runs past the ceiling

  • Switched on behind a flag

    a share of users first, with a count of suggestions accepted and dismissed beside it, and an irreversible step is offered to a person rather than taken by the suggestion itself

What we need from you

  • Access to the repository and to a staging environment where answers can be checked before production
  • An account with the model provider in your name — the keys, the limits and the bill for usage stay yours
  • A list of the data that must not leave your systems — it decides what goes into a request and what gets added afterwards

Projects in this area

What people usually ask

Other service lines

All service lines

Contact

Send a description of the task

A reply with the scope, the timeline and a budget estimate comes within 24 hours.

Get an estimate

The first call is 30 minutes, with no commitment on your side.

  1. 01

    You describe the task

    Five questions in the form, or a plain email — whichever suits you.

  2. 02

    We answer within a day

    With the scope, the timeline and a budget estimate, based on what you told us.

  3. 03

    We talk for 30 minutes

    To clear up whatever is unclear. It commits you to nothing.