One number, agreed in the room
A scope needs an indicator attached before the build starts. It tells us when to stop tuning, and lets the team say the agent is not working without it becoming an opinion.
How we work
Five steps to production.
A useful agent is not a prompt. It is a clear boundary, clean data, a test set, and supervision that does not stop at launch.
- 01 Scoping Week 1 to 2
We pick a measurable process, not a topic.
One workshop to fix the boundary, the expected value, and the single number that decides success. If we cannot find a clear trigger, one decision and an outcome somebody already tracks, we say the process is not ready.
- Process map
- Three ranked use cases
- Success indicator
- Feasibility note
- 02 Data Week 2 to 3
We prepare and secure the knowledge sources.
Collection, cleanup, access rights and European hosting settled before anything goes live. Retention windows are chosen rather than inherited from a default.
- Source inventory
- Access model
- Retention policy
- Hosting decision
- 03 Prototype Week 3 to 4
A testable agent in two weeks, on your real cases.
A narrow scope wired to your tools and scored against a business test set your team wrote with us. Roughly a fifth of that set is cases the agent is expected to decline.
- Working agent
- Business test set
- Baseline score
- Escalation rules
- 04 Production Week 4 to 6
Production inside your tools, with guardrails.
Single sign-on, full logging, human escalation and explicit limits on every write. Actions that change a customer-visible field require approval until the test set says otherwise.
- SSO integration
- Audit log
- Approval workflow
- Runbook
- 05 Supervision Ongoing, 90 days included
We measure, correct and extend it monthly.
An answer dashboard, a monthly review and iterations on the cases nobody anticipated. The test set runs on every change, so a regression is caught before your users find it.
- Answer scoring
- Monthly review
- Regression runs
- Backlog of gaps
02
What we hold to.
The test set outlives the model
Written properly, a business test set is portable. When a new model appears we run it and get an answer the same afternoon. Model choice becomes a measurement, not a debate.
Every write is reversible
Reads happen freely. Drafts and notes happen automatically and can be undone. Anything customer-visible waits for a human. Every write carries the run that produced it.
We decline the vague middle
Undocumented processes, or ones where two teams disagree on the correct answer, are process problems wearing an AI costume. Fixing the process first is faster to deliver.
Start with a diagnostic.
Two weeks to find out whether an agent is the right answer for the process that costs your team the most time.
Book a diagnostic