Personal Agent Protocol (PAP), aka Poppy 1.0, is here. Poppy was initially developed by Sierra and Meta and has since attracted more than 100 design partners, including major B2C companies such as Bank of America, Cigna, SiriusXM, Target, and United Airlines. The continued partner involvement is beginning to translate into implementation: Shopify has enabled Poppy across its merchant platform, while Synchrony, Mercury, and Turkish Airlines are among those going live. The protocol defines how personal AI agents can act on customers’ behalf with companies, including: identifying themselves, getting authorized account access, and interacting through APIs, websites, or company AI agents. It also supports human handoff.

Just as important is what Poppy doesn’t do:

  • It doesn’t give personal agents carte blanche. Companies control what account access customers can grant their personal agents and what those agents can see or change.
  • It doesn’t make your service policy decisions. Poppy is the protocol. You still decide what customers should be able to delegate to their AI agents.
  • It doesn’t force personal agents to use Poppy. Personal agents can still use other routes to interact with your company. Poppy gives them an official path, but it doesn’t close the unofficial ones.

Don’t Treat Poppy As An All-Or-Nothing Decision

Doing nothing may feel like the cautious choice, but it gives you less control, not more. Personal agents can still reach your company without Poppy. Publishing Poppy gives those willing to follow the protocol an authoritative source for dealing with your company, while letting you decide what they can access, change, or do.

  • Start with the customer’s task, not the protocol. Identify tasks where delegation removes meaningful customer effort, and you’re comfortable with an AI agent completing the work.
  • Be explicit about what’s off-limits. Decide which tasks you won’t allow personal agents to perform through Poppy. The v1.0 protocol spec uses sending a wire transfer and closing an account as examples of excluded tasks.
  • Don’t just let the agent fail. If you don’t support a task through Poppy, use the protocol to point the agent toward an appropriate alternative rather than leaving it to hit a wall or hunt for another (likely unsanctioned) route.

Make Sure Customer Service Tasks Are AI-Ready

Poppy gives personal agents a cleaner way to reach your service operation. It doesn’t fix what they find when they get there. Before making a customer service task available to personal agents, scrutinize what it takes to complete it.

  • Test the task end-to-end. Don’t stop at whether an AI can authenticate or reach an API. Can the agent complete the return, cancellation, booking change, or account update without an employee finishing the work elsewhere?
  • Map the skills the task requires. Identify the cognitive skills an AI agent needs to complete the work, such as interpreting results, assessing policies or exceptions, and reconciling information across systems.
  • Look for the human glue. Employees routinely compensate for disconnected systems, undocumented exceptions, ambiguous policies, and manual workarounds. Employees still must interpret, reconcile, or work around those gaps to get the right outcome; Poppy won’t eliminate that dependency.

Start with tasks where the path to resolution is already clear: the data is reliable, the business rules are explicit, and the necessary system actions can be executed consistently. For the rest, improve system access, make ambiguous policies executable, and identify the tacit knowledge that AI will need before personal agents can reliably complete the work.

Forrester clients can book a guidance session with us to understand what this will mean for their customer service operations.

Share