Designing a trusted enterprise AI answer layer
Turning fragmented content, product data and enterprise systems into one governed intelligence layer for people, search and AI.
- Type of work
- Proposed direction
- Context
- Multinational manufacturer · Client anonymised
- Role
- Product strategy, AI solution design and enterprise architecture
- Disciplines
- AI architecture, Enterprise knowledge, Governance, Customer experience
Problem
Useful product and brand knowledge was spread across web content, product documents, structured records and live services, with no single governed layer. A chatbot over the website would have answered from a fraction of it.
Context
Product and brand information lived in several places at once:
- Web content in the content management system
- Detailed product documents
- Structured records and spreadsheets
- Recommendation services
- Enterprise platforms
Each source served its own purpose. None had been designed to answer questions, and nothing decided which of them an answer should come from.
Key decision
Design the answer layer, not the chatbot.
A shared layer normalises approved knowledge once and serves every experience that needs it: a conversational assistant, site search and machine-readable outputs for other systems. Because each experience asks the same layer, none becomes its own source of truth, and the interface can change without rebuilding what sits behind it.
System model
Many sources, one governed layer, many experiences. Select a layer to see what it does.
Trust model
Not every answer carries the same risk, so not every answer is treated the same way.
-
Strictness 3 of 3: High-risk facts
Specifications and other facts that must be exact
Served from governed structured records and cited. The model can phrase the answer, not change the fact.
-
Strictness 2 of 3: Explanatory content
How things work, how to choose, how to use them
Generated only from approved sources, grounded and cited.
-
Strictness 1 of 3: Descriptive content
Brand and product stories
Generated from approved content, with lighter checks.
When the layer cannot answer reliably, it escalates rather than guesses.
Decisions
-
One knowledge layer, multiple experiences
The assistant, site search and machine-readable outputs use the same governed knowledge, instead of each channel growing its own source of truth.
-
Hybrid retrieval
Exact facts and open questions need different retrieval behaviour. Keyword and structured lookups give precision; semantic retrieval gives meaning.
-
Structured facts where precision matters
Specifications and other important product facts come from governed structured records, rather than asking a language model to recreate them.
-
Approval before availability
Content is extracted, validated and owned by someone accountable before it becomes answerable.
-
Tiered guardrails
High-risk facts get stricter treatment than explanatory or descriptive content. One rule for every answer would be too loose for facts or too heavy for everything else.
-
Evaluation is part of the architecture
Groundedness, factual quality, latency, escalation and adoption are measured continuously, not checked once before launch.
Proposed direction
A target architecture and product direction for one governed answer layer, shared by an AI assistant, site search and machine-readable outputs.
- One governed knowledge layer instead of a source of truth per channel
- Ownership and approval designed in before content becomes answerable
- Trust and evaluation treated as architecture, not as a test phase
What I learned
The interface was the smallest part of the problem. The real product was the trusted knowledge and orchestration layer behind it.
The client is not named. Brands, products, commercial details and internal terminology are left out, and the architecture is shown at the level of principle.