Why isn't a written approval policy enough to train an agent?

A policy document states intent. It rarely captures the state needed to apply that intent.

Three gaps show up quickly:

- **Authority is dynamic.** Who currently holds a role, who is absent and which delegations are active change daily. A policy says the head of operations approves; it does not say who that is this week. - **Limits interact.** An amount band, a risk flag and a product rule can each point to a different approver. The correct answer is usually the highest requirement that applies, and the policy may not say so. - **Exceptions are procedural.** Urgent approvals, verbal pre-clearance, split purchases and approver conflicts are often handled by habit and handover notes, not by clauses.

Training only on policy text teaches an agent to quote rules. Training on linked records teaches it to look things up, compare dates and amounts, and decide whether to proceed, route or stop.

A dataset or simulated environment for this workflow can pair policy wording with the authority tables, delegation records and request histories that the wording depends on. It can also include cases where the text and the tables disagree, so an evaluation can test whether the agent notices.

Because the material is generated, ask for a sample and compare its structure with your own process. Check that the disagreements it contains are plausible rather than arbitrary, and that nothing in it reproduces material you already hold.