The Context Problem

Acme’s CRM record says account_health = Green and renewal_risk_score = 82. Ask a model whether the account is ready to renew and it can read both values perfectly. It still has to guess what they mean.

Access gives the model values

A connector can return the right record, field names, and stored values. None of that explains whether 82 is good, whether Green is maintained, or which field wins when they disagree. The model fills that gap with familiar language. “Green” sounds healthy and a high score sounds positive.

That answer is plausible. It is also backward. Acme’s risk scale runs from zero to 100, and a higher number means more renewal risk. The Green field belongs to an abandoned manual process.

Make the failure repeatable

Run two or three questions in the lab. Watch the trace under the answer. The agent reads the correct values and invents the interpretation because no business definitions were loaded.

Try this in the lab: ask “What does a risk score of 82 mean?” Then ask whether Acme is ready for renewal.

The missing layer is organizational context

The CRM record is current record context. Field definitions, score thresholds, deprecation rules, and reporting exclusions are organizational context. They change slowly and apply across many questions. A context layer stores those rules and loads the relevant ones before the agent plans its answer.

Chapter 2 adds the first definitions. You’ll rerun the same question and inspect exactly why the conclusion changes.