“I want a workshop with room for two bikes and good daylight” is a useful beginning. It is not yet a configuration.
An AI interface can help translate that intent into questions and candidate choices. The durable product remains a structured model that the application can validate, render, price, and explain.
Translate into a proposal
Give the model a constrained vocabulary: available styles, supported dimensions, option identifiers, and fields for unresolved questions.
The output should be a proposed change, not an instruction to mutate arbitrary application state. Validate it against a schema and then against product rules.
A syntactically valid width can still be unavailable. A valid door identifier can still be incompatible with a selected wall. Schema validation and business validation solve different problems.
Ask about the missing facts
The phrase “good daylight” may suggest windows, orientation, or a different layout. It does not establish the site’s orientation or the user’s privacy needs.
Ask a small, useful follow-up question. Show why it matters. Avoid guessing a site condition simply to keep the conversation moving.
For consequential decisions, retain an explicit unknown. A blank field that prompts review is more useful than a confident invented answer.
Retrieve approved product information
Ground product explanations in a versioned catalog and approved documentation. Record the document and revision that support a response.
Retrieval does not automatically make an answer correct. The model can combine an outdated warranty with a current product or overlook a condition in the source. Check the final claims that affect a purchase.
Separate public product knowledge from tenant or customer records. A similarity search should not cross an access boundary.
Treat images as observations, not surveys
A site photograph can help identify questions about access or obstructions. It does not reliably establish all dimensions, ground conditions, buried services, or legal boundaries.
Show any inferred geometry as inferred. Ask for measurements and route information from the responsible people. The delivery operator or design professional still needs the evidence appropriate to their decision.
An attractive reconstruction is not a substitute for knowing what the image could not show.
Keep the quote deterministic
The AI can explain a price breakdown, but the accepted configuration and server-side rules should determine the authoritative quote.
Record the catalog revision, currency, calculation assumptions, and any exclusions. If an AI suggestion changes the product, show the difference before asking the user to accept it.
Do not silently personalize pricing from a model’s guess about a customer’s willingness to pay. A transparent product relationship is easier to maintain and review.
Validate meaning after validating shape
A response can be valid JSON and still request an impossible building. Use a versioned command schema with approved option IDs, explicit units, a base configuration revision, and typed operations such as adding or moving a feature. Reject unknown fields where silent interpretation would be dangerous.
Then validate relationships: an opening must fit its host surface, a material must belong to the selected catalog, and a requested operation must be permitted for this user. A model should propose the change; deterministic application rules decide whether it becomes accepted state.
Return structured rejection reasons that the interface can explain. A bounded repair attempt can address a missing dimension, but it must not repeatedly guess until something happens to pass.
Score the resulting configuration
Anthropic’s January 2026 agent-evaluation guidance distinguishes an agent’s trace from the environment outcome. Apply that distinction by checking the saved configuration, not just whether the assistant described a plausible change.
Build conversation fixtures for ambiguous units, a changed requirement, an unavailable option, a stale revision, and a cancellation. Define which facts require clarification and which defaults the product authorizes. Score both unnecessary questions and unsupported assumptions.
After each conversation, compare the accepted configuration with the expected constraints and verify that cancellation produced no mutation. For a quote, independently recompute the price from the accepted configuration and catalog revision.
Keep the evaluator outside the model’s narrative. A fluent explanation that says “the door now fits” cannot override a geometric check showing that it crosses the wall boundary. This turns language assistance into a measurable editing workflow.
Evaluate the whole conversation
Test ambiguous requests, unavailable options, incompatible placements, interrupted sessions, and attempts to request unsupported actions. Include prompts that contradict earlier choices.
Measure whether the resulting configuration is valid and whether the user understands it. A fluent response is not the completion criterion.
OWASP’s input validation guidance supports using explicit allowed structures and values. The same discipline helps turn language into a dependable interface.
The strongest AI configurator knows when to suggest, when to ask, and when to leave a decision with a deterministic rule or a responsible person.
Keep a good idea close.
Follow Signal Tower
