Applied AI
Designing a governed AI knowledge operations assistant
A secure retrieval and workflow-assistance architecture that helps enterprise teams find, interpret, and act on approved knowledge with human oversight.
The context
The operating problem we designed around
This reference architecture focuses on turning fragmented organisational knowledge into a controlled assistance layer. Answers need to respect document permissions, identify their sources, and remain useful when evidence is incomplete.
The solution also needed a practical path from experimentation to production: measurable quality, controlled model changes, cost visibility, and clear fallback to established human workflows.
Teams were spending time searching across policies, operating procedures, product documentation, and case histories. A generic chatbot would not provide the access control, source traceability, evaluation, or workflow integration required for dependable use.
Delivery approach
From ambiguity to an operable system
Start with bounded workflows
Selected high-value questions and tasks with known source material, owners, and escalation paths.
Design retrieval as a product
Built ingestion, classification, permissions, chunking, indexing, citation, and freshness controls together.
Evaluate before automating
Created representative test sets and reviewed relevance, faithfulness, refusal, and workflow usefulness.
Keep people accountable
Positioned the assistant as decision support, with review points for consequential actions.
The solution
Capabilities designed as one operational system
Governed knowledge ingestion
Connectors synchronise approved sources with ownership, classification, version, and access metadata.
Permission-aware retrieval
Search and generation operate only over content the requesting identity is authorised to access.
Cited assistance
Responses link claims to source passages and communicate uncertainty when evidence is weak or conflicting.
Workflow integration
Assistance is embedded into cases, service operations, and escalation paths rather than isolated in a chat window.
System architecture
Clear boundaries between experience, domain, and platform layers
LAYER 01
Source connectors
Repository-specific adapters normalise documents, metadata, permissions, and change events.
LAYER 02
Knowledge pipeline
Parsing, classification, chunking, embedding, indexing, and quality checks run as observable jobs.
LAYER 03
AI orchestration
Policy checks, retrieval, prompt assembly, model routing, citations, and guardrails execute server-side.
LAYER 04
Evaluation and audit
Versioned test sets, feedback, traces, and release records support controlled improvement.
Advanced capabilities
Intelligence introduced with evidence, controls, and operability
Multi-step task assistance
Guide users through approved procedures while preserving checkpoints and source evidence.
Knowledge-gap analytics
Aggregate unanswered or weakly supported questions to improve documentation and training priorities.
Model routing
Select models by task sensitivity, quality needs, latency, and cost under centralised policy.
Potential outcomes
What the engineering approach enabled
- Faster access to approved organisational knowledge
- More consistent use of current procedures and source material
- A measurable path from AI experimentation to governed operation
- Clear human ownership for review, escalation, and improvement
Engineering lessons
What carries into the next platform
- 01Retrieval quality depends more on content ownership and metadata than model novelty.
- 02A useful refusal is better than a confident answer without sufficient evidence.
- 03Evaluation must reflect real user tasks, permissions, and failure consequences.
Services applied
Technology landscape