AI Implementation

Something works in a notebook and there is no agreed route to customers touching it

Your AI bill moves and nobody can say which feature caused it

A pilot has been running for months with no date for production

Promoting the prototype as it stands

Code written to prove an idea rarely survives contact with real load and real data. We review honestly, and if the right answer is to rebuild the pipeline rather than harden it, that is quoted openly rather than discovered later.

Limits nobody has tested

A spending cap that has never fired is an assumption. We trigger ceilings deliberately in production and watch them hold, because the alternative is finding out during a busy week.

Nobody owns it after launch

Pilots die when the excitement fades and no name is against the system. Handover names an owner, documents what changed, and leaves your developers able to maintain it without us.

A prototype proves that a model can do the task on the examples it was shown. Production requires five things the prototype never had to have: integration with the systems people already use, so the output lands where the work happens; behaviour on the messy inputs, the scanned PDF, the empty field, the customer who writes in three languages; cost that scales predictably with usage, with ceilings that have been watched firing; monitoring so a failure is noticed before a customer notices it; and an owner, named, with a runbook, so the system survives the departure of whoever was excited about it.

Most stalled pilots are missing three of the five. Our guide on why AI pilots never reach production sets out the pattern; this service is the fix.

Week one is an honest review of what exists. Sometimes the prototype can be hardened; sometimes the right answer is to rebuild the pipeline and keep the prompt, and we say which before any money is spent on the wrong one. The output is a written scope, an acceptance threshold and a fixed price. Week two is integration and instrumentation: the connections to your CRM, inbox, calendar or database, cost attribution per request, ceilings triggered on purpose, and a test run on a sample of your real inputs. Week three is monitoring, security review, handover and go-live, with your own developers walked through what changed so they can maintain it without us.

Where a project is larger than three weeks, it is split into phases, each with its own fixed price, so the budget cannot drift.

Published UK ranges for a first AI implementation run from £5,000 to £20,000 at fixed-price suppliers and considerably higher on day rates. Ours is £4,000 to £15,000 fixed, with the running cost stated before go-live. The larger difference is the running cost after launch: a system instrumented per request and routed to the cheapest capable model costs a fraction of one that sends everything to a frontier model, and that difference compounds every month.

Market figures from the published UK ranges surveyed in our automation cost guide, September 2026.

Much of this work is integration into systems another team owns. We work to their conventions, in their repositories where they prefer, document every change, and leave the system in a state their developers can maintain. Where a client has no developers, the handover pack and the runbook are written for the next developer they hire, and a monthly support line is available at a price agreed at go-live, which you can stop paying without losing the system.

Our prototype works on a laptop. What does production actually require?

Integration with the systems people already use, cost that scales predictably with usage, monitoring so failures are noticed before customers report them, security and data protection review, and somebody accountable for it once the excitement fades. The prototype is usually the easy fifth of the work.

Can you stop our AI costs running away?

Yes. We attribute cost per customer, per feature and per model, route routine work to economical models, and enforce spending ceilings in real time. A limit nobody has watched stop something is an assumption, so we trigger it deliberately in production before relying on it.

Will you work alongside our existing developers?

Yes. Much of this work is integration into systems another team owns, so we work to their conventions, document what we change, and hand back something their developers can maintain without us.

What about GDPR and data protection?

A written data processing agreement is signed before any personal data moves, and we are registered with the Information Commissioner under reference ZC152144. Client systems sit on the client own provider accounts wherever possible so ownership is never ambiguous.