Internal AI agents for enterprise functions, built on the client's own data

The situation

A materials science organisation deciding that generative AI needed an internal strategy, not a vendor contract. The question was how to put GPTs and agents in the hands of their own teams on their own data, without exposing that data to the public internet.

What we did

Engineers embedded into the client's AI and platform teams, building internal agents function by function.

Each agent grounded in the data that function actually works from: regulatory submissions for compliance, formulation and process data for R&D, supplier and spec history for procurement, operations and sales data for commercial.

Retrieval tuned per corpus. The chunking strategy and embedding model that work on a legal document do not work on a scientific paper or a batch record. We tuned them against the actual data each time, not defaulted.

Evaluation harnesses for every agent before it reached a user. If a prompt change made the agent wrong on five of fifty labelled cases, we knew before deploy, not after.

Guardrails against the obvious failure modes: hallucinated references, data from one tenant surfacing in another's response, sensitive data leaving the model's context into logs. These are table stakes in an enterprise AI programme and still routinely missed.

Logs that make an agent's decision auditable, because an enterprise using AI internally will eventually have to explain to legal or audit what the agent did and why.

Change management alongside: training the function teams to use the agents, collecting feedback, iterating the prompts and the agent's tool definitions. The engineering is a fraction of the work. Adoption is the rest.

The shape of it

A small to medium standing team embedded in the client's AI and platform groups.

The outcome

GPTs and agents in regular use across functions, grounded in the organisation's own data, with the evaluation and audit scaffolding to run them responsibly over time.