AI, where it adds value

AI where the work genuinely needs language or judgement. Nowhere else.

Only when it pays back

The problem.

Your documentation is spread across a file share, a wiki, a ticketing system and years of email. People know the answer exists; finding it costs more than rewriting it. A model can help, but only if it respects who is allowed to read what, and only if someone can say whether it works.

How it works.

Three patterns we build, step by step

Permission-aware retrieval

Answers only from documents the person asking is allowed to open.

Asked in plain words Question
Top passages from the index Retrieve
Drop what they cannot open Check access
Or a refusal, never a guess Cited answer

Dropped: no access

Use whenPeople need answers from documents with different access rights.

Agent with human approval

The agent prepares the action. A person decides whether it runs.

Prepares the action Agent drafts
Reads it, says yes or no Person approves
Send, pay, update Action runs

If no: it stops

Use whenThe action touches money, customers, or records.

Evaluation gate before release

Agreed test cases decide when a system is ready, not a demo.

Real cases, agreed upfront Test set
The system answers each case Run
Accuracy, leakage, refusals Score
Only a pass goes through Release gate
Into production It ships

Fail: fix and run again

Use whenBefore anything reaches production.

Three AI engineering patterns. Permission-aware retrieval: a question retrieves passages, the ones the person cannot open are dropped, and the answer is cited or refused. Agent with human approval: the agent prepares an action and a person approves it, or it stops. Evaluation gate: an agreed test set is run and scored, and the system ships only when it passes, otherwise it is fixed and run again.

What we do.

  • Knowledge assistants that answer from your documents, filter every passage against the permissions of the person asking, and cite their source or refuse.
  • Agents scoped to one job, with a person approving any action that touches money, customers or records.
  • A test set built from your real cases, agreed before the build, that decides when the system is ready.
  • Logging, cost tracking and monitoring, because systems that handle language degrade in ways ordinary software does not.
When this is the wrong answer

The task follows fixed rules with no judgement in it: that is automation, cheaper and deterministic. Or your documents are out of date and contradict each other: retrieval will surface the wrong answer faster. Fix the source of truth first.