The AI assistant that passed risk review

T H E C H A L L E N G E:

Leadership wanted generative AI. Legal and risk would not
approve it. Every vendor demo failed the same three
points: answers could not be traced to a source, retrieval
had no concept of who was asking, and nobody could
prove the output was correct — today or after the next

model update. Meanwhile frontline teams kept hand-
searching plan documents, contracts and SOPs.

O U R A P P R O A C H :

We started with the data, not the model. Retrieval sits on
the governed layer of the existing platform, so the assistant
inherits the exact row- and column-level security the
warehouse already enforces — if a user cannot query it,
the assistant cannot retrieve it. Every answer carries
citations, governed operations are exposed as explicit MCP
tools, and an evaluation harness grades every release
before deploy.

W H A T W E D I D

– Built hybrid retrieval — vector plus structured — over
policy documents, contracts, SOPs and governed
warehouse tables
– Made retrieval entitlement-aware: the assistant inherits
the user’s own access rights and cannot surface what
they could not already query
– Exposed governed operations — eligibility checks,
record lookups, policy search — as explicit MCP tools,
not open database access
– Built the evaluation framework — golden question set,
groundedness and citation scoring, CI gate
– Implemented guardrails and a queryable audit trail — PII
redaction, injection defence, refusal policies, human
review for regulated determinations

T H E O U T C O M E

Approved by legal and risk — the governance model,
not the demo, cleared review
Minutes of document search reduced to one question
for frontline teams
Every answer carries citations to a named,
permissioned source
No entitlement violations — retrieval is bounded by
warehouse controls
Release confidence — nothing ships without passing
the golden set, and the pattern now extends to further
business units