Home / Case Studies / Enterprise AI
Enterprise AI

Enterprise AI copilots that survived contact with real users

A pilot that impresses in a demo is easy. We built the AI layer that people still use in month six, with the evaluation and cost controls that make it defensible.

4 hrs
given back per person, per week
86%
of answers accepted unedited
61%
weekly active use at month six
100%
of answers traceable
The challenge

Where it started

Several AI experiments had stalled: impressive demos, no measurable adoption, and no way to tell whether answers were right.

Leadership needed a cost per answer they could forecast and evidence they could show an auditor.

Client: InvexAI

What we built

The solution, in four parts

One workflow at a time

Each copilot scoped to a single job with a named owner and a measured baseline before build.

Grounded and cited

Retrieval over the company’s own content with permissions honoured, and a source next to every answer.

Evaluated continuously

A fixed test set scored on every release, so quality is a number rather than an opinion.

Cost and governance

Per-answer cost tracked from week one, model cards and usage policies written as the work went along.

Technology

What it runs on

Mainstream, well-supported technology — chosen so the client’s own team can own it afterwards.

ClaudeLangGraphPythonTypeScriptVector searchKubernetes

The difference was the evaluation harness. Once we could see quality as a number, the business stopped arguing and started using it.

Chief Technology Officer, InvexAI

More work

Other outcomes we’re proud of

Want an outcome like this one?

Tell us the number you need to move. We’ll tell you honestly what it takes and how quickly you’d see the first release.