Grounded AI assistants: what actually works in production

Every team we meet has tried an AI assistant. Far fewer have one that people still open six months later. The difference is rarely the model. It is the boring engineering around it.

Retrieval that respects boundaries

An assistant grounded in your own documents is only useful if it retrieves the right documents for the right person. We scope retrieval per user and per tenant, index with metadata your access-control system already understands, and log which sources produced every answer.

Actions with a human in the loop

The moment an assistant can do things, not just say things, trust becomes the product. Every action gets a preview, an approval step for anything irreversible and an audit trail that a compliance team can read without help.

• Preview the change before it is applied
• Require approval for irreversible or external actions
• Record who approved what, with the full context the model saw

Evaluation is the roadmap

We build an evaluation harness in week one: a few hundred real questions with expected answers, run on every change to prompts, retrieval or model. It turns arguments about quality into charts, and it tells you when a cheaper model is good enough.

“The assistant that survives is the one whose failures are visible.”