Know what it may use
Approved collections, currency, ownership and access are designed before the interface.
AI assistants / Knowledge
Answer from approved knowledge and show where the answer came from.
A controlled-knowledge assistant retrieves relevant material from an approved library and uses it to prepare a concise answer with source references. GA Applications designs content ingestion, permissions, retrieval, answer boundaries and feedback so users can verify the guidance and escalate when sources are missing, outdated or contradictory.
Bounded requestWhich procedure applies when two approved sources disagree?
RequestRetrieveDraftReview
Draft with evidence
The assistant has prepared a response, but it is not an approved decision.
1Procedure register / current versionapproved source
2Change notice / effective 18 Julycheck currency
Pending review. No external action is available in this demonstration.
Approved sourcesVisible evidenceRole permissionsHuman reviewEvaluationDisable path
Useful because the job is bounded. Trustworthy because the boundaries are visible.
AI assistance belongs inside a designed operating process. It does not replace accountable people, authoritative records or competent professional judgement.
Let people inspect support
Citations, source context and uncertainty travel with the draft.
Keep decisions attributable
Permissions, review, escalation, logging and shutdown routes match consequence.
01
Build an owned, reviewable knowledge collection.
Policies, procedures, product information and guides need authority, effective date, audience and lifecycle before an assistant can use them responsibly.
Inventory
Identify canonical items, duplicates and restricted material.
Ownership
Assign review, update and retirement responsibility.
02
Make answer, quotation and source easy to distinguish.
The response provides a direct explanation, links to evidence and signals when retrieved material is incomplete or conflicts.
Citation
Open the source and relevant context for verification.
Escalation
Route unresolved questions to the responsible person or process.
03
Respect source permissions at retrieval time.
The assistant searches only material the authenticated user may access and does not reveal restricted titles or snippets through error messages.
Identity
Carry the current user and role into the retrieval decision.
Boundary
Separate public, customer, staff and sensitive collections.
04
Evaluate answers against real questions and refusal cases.
A test set covers answer support, citation relevance, permission leakage, outdated sources, ambiguity and prompt manipulation.
Quality
Score whether the answer is supported, useful and appropriately cautious.
Feedback
Let users flag an answer and route the issue to source or system owners.
05
Launch with a narrow collection and visible support path.
We prepare sources, index them, test representative questions and monitor use before adding more audiences or content.
Pilot
Use a defined group and publish clear purpose and limitations.
Operate
Review source changes, failed queries, feedback and access regularly.
Questions worth answering
What teams ask before they commit.
Does the assistant train itself on our documents?
Not automatically. Retrieval and model-provider data use are designed and documented for the selected architecture; source updates remain an owned publishing process.
What happens when documents disagree?
The assistant should surface the conflict, cite both relevant sources and direct the user to an authorised owner rather than invent a resolution.
Begin with the source library
Test an assistant against the questions your team actually asks.
We can audit a bounded collection, its permissions and the evaluation needed for a responsible pilot.