Illustrative engagement. This write-up describes how we approach this kind of work. It is an anonymised archetype rather than a named client project — the figures are representative, not audited.
A support copilot that knows when to shut up.
An LLM copilot that drafts support replies from a company's own docs and ticket history — and escalates to a human instead of guessing.
Every one
Answers grounded in cited sources
Hundreds of tickets
Eval suite before launch
By design
Escalation instead of a guess
The challenge
Support was drowning, and an off-the-shelf AI widget had already been tried and switched off — it answered confidently from outdated help-centre articles and invented policy that the company then had to honour. Leadership wanted the efficiency but had lost trust in the category.
How we approached it
Built the eval harness first, from a few hundred real resolved tickets with known-good answers, so every prompt and retrieval change moved a number instead of a vibe.
Grounded every answer in retrieved, cited sources — if retrieval returned nothing above the confidence threshold, the copilot escalated rather than improvised.
Kept a human in the loop by design: the copilot drafts, an agent sends. Adoption came from making agents faster, not from replacing them.
Routed inference through a gateway with model fallbacks and per-request cost ceilings, and logged every call with customer identifiers stripped.
What shipped
Agents work from a draft instead of a blank box, and the answers carry links to the sources they came from. Because the system escalates when it is unsure, the failure mode is a slower reply rather than a wrong one — which is what earned it a second chance internally.
Could this be your story?
Tell us where you’re stuck. We reply within one business day with next steps.
Start a project- Senior engineers, never juniors
- Weekly demos against a roadmap
- Code, infra, and docs you keep