service · knowledge engineering
Teaching your operation to an agent
Your experienced people know things no document holds. This is how an agent learns them.
Capturing what your experienced people know — procedures, judgement calls, exceptions — into grounded, versioned knowledge an agent can actually work from.
- Claude / LLMs
- RAG / pgvector
- Evaluation harnesses
- Git-versioned knowledge
overview
What it is
An agent is only as good as what it has been given to work from, and the most valuable knowledge in most operations is not written down anywhere. It lives with the people who have handled the edge cases — the exception that always needs a second check, the customer type that gets treated differently, the step everybody quietly skips.
This is the unglamorous work that decides whether an agent is useful or merely fluent. We extract that knowledge, write it down in a form that can be retrieved and tested, build evaluation sets that prove the agent gets the real questions right, and hand it over with an owner and a review cadence so it does not rot.
what's included
Capabilities
Capture from people, not just documents
Structured sessions with the people who actually handle the work, aimed squarely at the exceptions and judgement calls the process document never mentions.
Tacit judgement made explicit
'You just know when to escalate' becomes a written, testable rule — or an honest note that this decision must stay with a person.
Structured for retrieval
Knowledge written and chunked so retrieval actually finds the right passage, with sources attached so every answer can be traced back.
Evaluation sets
A set of real questions with known-correct answers, so 'the agent knows our work' is something demonstrated rather than asserted — and regressions are caught when the knowledge changes.
Versioning & ownership
The knowledge base is versioned like code, with a named owner and a review cadence, so a policy change reaches the agent instead of quietly leaving it wrong.
Handover
Your team is taught to write, test, and ship knowledge changes themselves — the point is that you own it afterwards.
why it matters
Benefits
- The knowledge that currently lives in a few heads becomes a written asset your organisation keeps.
- Agents answer from your reality — your exceptions, your policies — instead of plausible generic guesses.
- Evaluation sets turn 'it seems to work' into something you can re-run after every change.
- Writing it down surfaces the contradictions and stale policies that were already costing you, agent or no agent.
in practice
Use cases
how we work
The engagement
Scope the domain
We pick one bounded area of work — a product line, a process, a queue — rather than trying to capture everything.
Capture sessions
Structured interviews and shadowing with the people who handle it, focused on exceptions and judgement.
Write, structure & test
Knowledge written for retrieval, with an evaluation set of real questions and known-correct answers.
Own it
Versioning, a named owner, a review cadence, and training so your team keeps it current without us.
get in touch
Let's talk
Have an operation you'd like an agent to run, or want to see how Teaching your operation to an agent could fit your work? Tell us about it and we'll come back with a tailored plan.