← All experiments
Experiment 004Architecture

Business Record Agent

Stage
Concept / Research
Last updated
2026-07-18

Research question

What could an AI agent do if it had access to the structured, authoritative record behind a business?

Hypothesis

Agent usefulness tracks record completeness more closely than it tracks model capability.

Hypothesis — not a conclusion

Why it matters

An agent that can read an authoritative record can answer operational questions a marketing website never states clearly.

Architecture

  1. Business Record
  2. Agent context
  3. Task
  4. Response
  5. Action

Prototypes

What we're testing

  • Which tasks improve most when a record is present
  • Whether missing fields produce refusals or invention
  • What actions an agent should be allowed to take on a business's behalf

Observations

Agent answers degrade in predictable ways when a record is incomplete rather than absent.

What we learned

Testing has not yet produced documented results.

What would change our mind?

  • General models answer just as well without a record
  • Businesses reject agent-initiated actions
  • Record freshness cannot be maintained at the pace agents assume

Open questions

  • Where should the boundary between reading and acting sit?
  • How should an agent express uncertainty about a stale field?

Next experiment

Prototype agent tasks that read a record and answer questions a website cannot.

Experiment 005AI Visibility
Method

Question → Hypothesis → Prototype → Observation → Evidence → Conclusion. We never present a hypothesis as a conclusion.