Write Docs That Teach Your AI When Rules Bend

Write Docs That Teach Your AI When Rules Bend

TL;DR

Most AI hedging comes from docs that state a policy but never define its exceptions. Write a 'condition block' in every article that maps each trigger to a specific outcome, and your agent stops saying 'it depends' and starts giving the answer the customer actually needs.

The Real Reason Your AI Keeps Saying 'It Depends'

You publish a refund policy. It says refunds are available within 30 days. A customer asks whether they can get a refund on day 32 because their account was frozen by a billing error on your side. Your AI agent reads the doc, finds '30 days,' and either says no or hedges into oblivion.

The doc wasn't wrong. It just never taught the agent what to do when the rule bends. That gap is a writing problem, not a model problem.

Why 'It Depends' Is a Structural Failure

Support agents — human or AI — answer conditionally all the time. The difference is that a human carries years of unwritten context. They remember the last time a billing freeze caused a chargeback dispute, and they know the team makes exceptions in those cases. That knowledge lives nowhere in your docs.

When an AI agent hits a question where the answer varies, it has two moves: pick the most literal interpretation of what's written, or flag uncertainty. Both responses frustrate customers. The fix isn't prompt engineering — it's writing docs that make the conditions explicit.

Anatomy of a Conditional Doc

Every policy or procedure that has exceptions needs a condition block: a short section that maps the variable to the outcome. Think of it as a decision table written in plain prose.

Here's the structure:

  1. State the default rule clearly. One sentence, no qualifiers yet.
  2. Name each condition that changes the outcome. Be specific about what triggers the exception — not 'special circumstances' but 'account was inaccessible due to a platform error on our side.'
  3. Assign a concrete action to each condition. Don't write 'may be eligible for review.' Write 'issue a full refund and log it under billing-error exceptions.'
  4. Define who decides when the AI can't. If a condition requires judgment the agent shouldn't make alone, say so explicitly: 'Escalate to the billing team with the error timestamp attached.'

This structure works because an AI agent reading it doesn't have to infer intent — the doc has already done the reasoning.

A Before and After

Before (vague):

Refunds are processed within 30 days of purchase. Requests outside this window may be considered on a case-by-case basis.

An agent reading this knows nothing about what 'case-by-case' means, who considers it, or what outcome to expect.

After (conditional):

Default: Refund requests are approved if submitted within 30 days of purchase. Issue the refund immediately through the billing dashboard.

If the account was inaccessible due to a platform error: Approve the refund regardless of when the request arrives. Attach the incident reference number to the refund record.

If the customer is outside 30 days with no platform error: Decline and offer a one-month credit instead. Escalate to billing only if the customer disputes the credit offer.

The second version gives the agent a path for every scenario your team actually encounters.

How to Discover Which Conditions to Document

Pull your last 90 days of tickets and filter for any resolution that deviated from the written policy. Those deviations are your undocumented conditions. If your team approved three late refunds last quarter, each approval was a condition your doc hasn't captured yet.

In Chattering's shared inbox, you can tag tickets with a custom label like policy-exception as you close them. After a few weeks, that tag becomes a backlog of conditions waiting to be written into your help center.

Formatting Rules That Help the Agent Parse Conditions

MD headers and bold labels matter more for AI reading than for human skimming. When Chattering indexes a doc, structural signals help it locate the right passage for a given query.

  • Use **If [condition]:** as a consistent lead-in phrase. Consistency makes it easier for the agent to pattern-match.
  • Keep each condition block self-contained. Don't make the agent stitch together two paragraphs from different sections.
  • Avoid nested exceptions where possible. If your policy has exceptions to exceptions, that's a signal to split it into two separate articles.

What to Do With Conditions You Can't Document Yet

Some judgment calls are genuinely too nuanced to reduce to a rule — they depend on customer history, contract tier, or context that isn't in the ticket. Document that explicitly: 'If none of the conditions above apply, collect the customer's account ID and escalate to the support lead. Do not attempt to resolve independently.'

That instruction is still far more useful than silence. It stops the agent from guessing and ensures the handoff to a human includes the right context — which Chattering passes through automatically when the escalation triggers.

The Maintenance Loop

Condition blocks go stale when policies change and nobody updates the exceptions. Build a lightweight review into your release process: any time a policy changes, the person making the change is responsible for auditing the condition block in the corresponding help article. Five minutes of editing prevents weeks of miscommunications at the agent layer.

Your AI agent is only as precise as the decisions your docs have already made. Write the conditions down, and the hedging disappears.

Frequently asked questions

Does this only apply to refund or billing policies?

No — any policy with exceptions benefits from a condition block. Common candidates include account access rules, feature availability by plan, escalation triggers, and SLA commitments that vary by tier.

Won't writing out every exception make the docs too long for customers to read?

Condition blocks are written for the AI first and customers second. You can keep the customer-facing summary brief at the top and move the full condition table under a collapsible section or a separate internal doc that Chattering indexes but doesn't surface directly.

What if a customer asks something we genuinely haven't documented a condition for yet?

Your doc should include a catch-all escalation instruction — something like 'If no condition above applies, collect the account ID and escalate to a human agent.' That prevents the AI from guessing and ensures the handoff includes useful context.

Keep reading

Answer it once. Chattering remembers.

Free trial · no credit card required

Write Docs That Teach Your AI When Rules Bend | Chattering.ai