Grounded policy answers for a distributed workforce — a permission-aware knowledge system on Google Cloud
A distributed organisation wanted staff anywhere to get a reliable answer to "am I entitled to…?" or "how do I apply for…?" in seconds — from the actual policy, with the actual citation — instead of waiting days for HR.
The challenge
HR and policy documents lived in several systems, in several versions, with different access rules for different staff groups. A general-purpose chatbot would happily answer from the wrong version — or from the open web. The organisation needed answers grounded only in current, approved documents, visible only to people entitled to them, and auditable when challenged.
The approach
- Frameentitlements questions first
- Groundsources + permissions
- Buildcited answers, declines
- EvaluateHR-written test set
- Operateclient team runs it
- Start with the questions HR actually gets. Leave, allowances, expenses and onboarding — the highest-volume, lowest-ambiguity topics — with an evaluation set written by the HR team.
- Ground at request time. Each question retrieves from the current approved documents with the user's own permissions; the model answers only from what was retrieved and cites every source.
- Decline by design. When retrieval returns nothing relevant, or policies conflict, the assistant says so and routes to a person — that behaviour is tested, not hoped for.
- Hand over operations. Dashboards for unanswered questions, a content-gap report for HR, and documentation so the client's team owns the system after launch.
What we built
- Permission-aware retrieval over HR and policy documents with version control
- Grounded answers with citations, provenance and explicit declines
- Evaluation harness and regression tests covering accuracy, leakage and tone
- Operations dashboards and a content-gap report for the policy owners
Results
- Staff get cited answers in seconds instead of days in an HR queue
- Every answer traceable to a document, section and version — zero uncited answers in evaluation
- HR time shifts from repetitive questions to the exceptions that need judgement
- Content gaps surfaced automatically, improving the policies themselves
What made it work
Grounding beats fine-tuning here
Policies change; retrieval at request time keeps answers current without retraining.
Test the declines
An assistant that says 'I don't know' at the right moment is what makes HR comfortable switching it on.
Measure the gaps
The unanswered-questions report became a tool for improving policy, not just the bot.
Government and enterprise programmes are described under confidentiality agreements; client identities, figures and architectures are generalised where required. Marketing results are reported as the clients described them.