AI-assisted B2B product discovery
Professional buyers know the job they need done, not every product it takes. Designing discovery that starts from intent, with AI assisting expertise rather than replacing it.
- Type of work
- Proposed direction
- Context
- B2B commerce · Client anonymised
- Disciplines
- AI product discovery, Hybrid retrieval, Human-in-the-loop, Commerce
Problem
A professional customer types “I need to build a 40m² drywall partition.” They know the outcome. They do not know every product, size and quantity it takes, and traditional search cannot help them: it expects product names.
Context
Professional buyers often purchase for a project rather than for a product. Their questions describe a job to be done. The catalogue, meanwhile, is organised by product: categories, attributes and article numbers.
That gap is usually closed by people. Customers work out the list of materials themselves, or they ask someone who knows.
Why it mattered
When search only understands product names, the customer does the translation. Anything they miss becomes a second order, a delay on site, or a sale that goes to whoever understood the job.
It was also a clear test of where AI belongs. The question was never “can a language model answer this?” It was “how do we combine AI with product data and real expertise so that the answer can be trusted?”
Approach
We modelled discovery around intent rather than around the catalogue.
First, interpret the request: what is being built, at what size, to which requirements. Then retrieve candidates with hybrid retrieval: keyword matching for exact attributes, semantic matching for meaning. A product graph adds the relationships between products: what is needed together and what is compatible.
AI reasoning assembles this into a proposed solution, and an expert validates it before it reaches the customer as a recommendation inside the normal commerce experience.
System model
-
Customer intent
“I need to build a 40m² drywall partition.”
A job to be done, not a product name.
-
Intent interpretation
What is being built, at what size, to which requirements.
-
Hybrid retrieval
-
Keyword retrieval
Exact attributes, standards and article numbers.
-
Semantic retrieval
Meaning, synonyms and descriptive language.
-
-
Product graph
Which products belong together, and what is compatible, grounded in product data.
-
AI reasoning
Assembles candidates and relationships into a proposed solution.
-
Expert validation
A specialist confirms or corrects the proposal before the customer sees it.
Human in the loopAI assists expertise; it does not replace it.
-
Recommended solution
A validated set of products and quantities, inside the commerce experience.
Decisions
-
Hybrid retrieval, not vector-only
Exact attributes such as dimensions, standards and article numbers need keyword precision. The customer's intent needs semantic understanding. Neither is enough on its own.
-
Model relationships explicitly
Which products belong together is captured in a product graph, not left for a language model to guess.
-
Ground the reasoning in product data
The AI reasons over retrieved products and their relationships, not over its general training data.
-
Keep an expert in the loop
A recommendation for technical work needs someone accountable. The AI drafts the solution; an expert validates it.
Proposed direction
A solution architecture and product direction for intent-based discovery, designed to fit the existing commerce experience.
- Discovery designed around the customer's job instead of product names
- Hybrid retrieval and a product graph as the knowledge layer
- Expert validation designed in as a feature, not a fallback
What I learned
In B2B commerce, the most valuable thing AI can do is often translation, from the customer's job to the catalogue's language.
The client is not named. The architecture is shown at the level of principle, not implementation.