Inside a Playbook: from one demonstration to a governed workflow

The Playbook is the heart of Autofil. It is the artefact your process owner reads and approves, your risk team bounds, and your agents execute. This post walks through how one is made — using an illustrative claims-intake workflow.

Step 1: Show

A senior claims handler opens the Autofil recorder and processes three first-notice-of-loss emails exactly as she normally would. As she works, she talks:

“I always check the policy was active on the loss date, not today’s date. And if there’s any mention of injury, I don’t file it — it goes to a senior adjuster.”

That narration matters. It turns tacit knowledge — the stuff that never makes it into an SOP — into explicit rules.

Step 2: Compile

Within a few minutes, Autofil produces a draft Playbook. It is written in plain language, with each step tied to the evidence it was learned from:

playbook: claims-intake  v1 (draft)
trigger:  mailbox "fnol@" · new message with attachment

1. Extract claim facts from email + attachments
   → policy_no, loss_date, loss_location, description, injury_flag
2. Look up policy in PolicyCenter
   → verify status = active ON loss_date        # from narration
3. IF injury_flag → route to senior adjuster     # from narration
   guardrail: human decision required
4. Create claim in ClaimCenter with extracted facts
5. Reply on email thread with claim number
   → verify claim number matches ClaimCenter

Notice what is not there: coordinates, selectors, or anything a business owner can’t read.

Step 3: Review and bound

The process owner reviews the draft like she would review a new team member’s notes. She adds a variant — “claims from brokers come with a different form” — and Autofil asks for one more demonstration to learn it.

Then risk adds the guardrails, which live in Retypd Governance rather than in the Playbook itself:

• Claims with estimated value over $25,000 require approval
• The agent may read but never edit policy records
• Personal health information is masked in all recordings

Step 4: Test against history

Before the Playbook touches production, Autofil replays it against the last 500 real claims in simulation mode — reading the real screens, but writing nothing. The output is a report: where it agreed with what the humans actually did, where it differed, and why.

Say it matches on 96% of claims. Most of the differences will be legitimate edge cases, which become new branches in the Playbook. A few may be errors in the human processing — which is worth knowing too.

Step 5: Promote, carefully

The Playbook goes live at Level 1 (shadow mode) for a week, then Level 2 (suggest), and reaches Level 3 (supervised production) in week four. Every change from here on is versioned, diffed and re-tested against the replay suite before it ships.

Why this structure matters

A Playbook is a contract between three parties: the people who know the work, the people accountable for its risk, and the agents that perform it. Keeping it readable, testable and versioned is what makes it possible for all three to trust it — and that trust is what lets the workflow scale.

Want to see a Playbook built from one of your own workflows? Talk to us.