parley
DocsPlatform
Resources
JournalIdeas on running AI supportCustomer storiesHow teams use ParleyGuidesStep-by-step playbooksChangelogWhat shipped, by versionHelp centerAnswers and support
CustomersPricing
LoginGet started
DocsPlatformCustomersPricingJournalCustomer storiesGuidesChangelogHelp center
Get startedLogin

Take charge of your AI customer experience

Get startedTalk to sales

Platform

Platform overviewChannelsPlaybooksIntegrationsAnalyticsPricing

Resources

DocumentationGuidesCustomer storiesJournalChangelogHelp center

Company

AboutCareersContact salesContact supportTerms of servicePrivacy

Security

Trust centerSOC 2ISO 27001GDPRHIPAAData processing
parley

© Parley. All rights reserved.

Journal · Knowledge

Writing support policies an AI agent can follow word for word

Policies written for experienced people lean on judgement. Agents need the judgement written down. A short guide to rewriting them.

RARuth AdeyemiSolutions Lead · 6 min read

Most support policies were written for people who already know the business. "Use discretion on late returns" makes sense to someone with two years on the team. To an agent, it is an invitation to guess.

The fix is not a longer policy. It is a policy where every decision has a rule, every rule has an owner, and every exception says what to do instead.

One decision per paragraph

Split long policies into short sections that each answer one question: can this item be returned, how is the refund paid, what happens if the item is damaged. Agents retrieve paragraphs, not documents, so a paragraph that answers two questions will be cited for both and fully right for neither.

Replace judgement words with numbers

  • "Recently" becomes "within 30 days of delivery"
  • "A small amount" becomes "under £40"
  • "Use discretion" becomes "offer store credit; hand over if the customer declines"

If the team genuinely needs judgement in a case, say so and make it a handoff rule. That is a decision too.

Write the exceptions as instructions

"Final sale items cannot be returned" tells the agent what not to do. "Final sale items cannot be returned; offer to check the warranty instead" tells it what to do next. Every no should come with a next step.

Every no should come with a next step.

Give every policy an owner and a version

When a customer is told something wrong, you need to know which version of which policy said it and who can fix it. In Parley each knowledge source shows its owner, its last sync and the conversations that cited it, so a fix is one edit away.

Test with real questions

Take fifty real customer questions about the policy and run them as simulations before you publish. The failures will tell you which paragraphs are still leaning on judgement.

RARuth AdeyemiSolutions Lead, ParleyAll posts

Keep reading from the journal

All posts
Voice5 min read

Writing for the ear: what changes when your agent picks up the phone

Read
Engineering6 min read

Rehearse before you ship: using simulations as a release gate

Read
Operations5 min read

One agent, every language: localising support without a second build

Read

Try it on your own conversations

Start building an agent in minutes, or talk to us about your support volume.

Get startedTalk to sales