Your customers will send their agents: what the Personal Agent Protocol means for the board

At a glance: On 6 October 2026, Meta and AI customer-service company Sierra proposed the Personal Agent Protocol (PAP), an open standard for how consumers' personal AI agents sign in to, and act with, businesses. Built on OAuth, it lets customers grant agents read or write access and lets companies decide which channels agents may use. Shopify, Stripe, Walmart and Genesys are among the named partners. For boards, it turns agent readiness into a customer-experience and brand question.

The news: a rulebook for agents that act for customers

For two years most corporate AI effort has gone into agents that companies deploy: service bots, sales assistants, internal copilots. The Personal Agent Protocol addresses the other side of the counter. It sets out how an AI agent working for a consumer, rather than for the company, should find a business, prove it has the customer's permission and get a task done.

Sierra announced the proposal on 6 October 2026 in a post by its co-founders, Bret Taylor and Clay Bavor. It describes PAP as an open standard that Meta and Sierra are developing with industry partners, open for anyone to implement. The named partners are Genesys, Instinct, Rocket, Shopify, Stripe and Walmart. Meta published its own announcement for businesses alongside.

Sierra said a v0.1 specification would follow later in October, with design workshops for interested parties and a reference implementation for developers. A draft specification, nicknamed "Poppy", has since appeared on the protocol's own website. It is marked as a draft, released under the Apache 2.0 open-source licence, with design partners working towards a version 1.0.

How it works, in plain terms

The design is deliberately familiar to anyone who has connected one app to another.

• Discovery. An agent starts at a company's website, where a standard file tells it what the company offers and how to start a session and sign in.
• Guest or signed-in sessions. An agent can act as a guest, for example to check availability. For account tasks, the customer signs in on the company's own page or uses credentials already set up with their agent.
• Permissions set by both sides. The customer chooses whether the agent gets read-only or write access. The company sets limits on what agents can do.
• One session, many routes. Sessions are built on OAuth, the widely used authorisation standard, and carry across channels. A question asked before sign-in and an order change made later belong to the same visit.
• Company choice of channel. Agents can use the company's ordinary web pages, its APIs (through interfaces built on MCP and OpenAPI) or the company's own agent for conversational tasks such as a warranty claim. The company decides which routes to offer.

Sierra lists likely extensions: finer-grained permissions for specific actions, push notifications such as flight delays or shipping alerts, and payments extensions so an agent can complete a purchase without sharing card details.

Sierra frames the benefits for three groups. Consumers get speed, dependability and trust. Brands get visibility into agent activity and control over what agents can do. Agent builders get a direct, consistent way to work with participating companies.

Why this belongs on the board agenda

Two quotes from partners frame the commercial stakes. Tony Bates, chairman and chief executive of Genesys, said: "Personal AI is creating a new front door to the enterprise." Kevin Miller, head of payments at Stripe, said: "A customer relationship doesn't start or end at checkout."

The WCRC analysis is that PAP matters less as a piece of plumbing than as a signal. Two very large platforms and a group of commerce, payments and contact-centre suppliers are betting that consumers will delegate routine dealings with companies to their own software. If that bet is right, four things change for business leaders.

1. The customer journey gains a new intermediary. Today, a customer who wants to change a delivery, dispute a charge or rebook a flight deals with the company's website, app or staff. Under an agent model, much of that interaction happens between two pieces of software. The company still serves the customer, but it may never see the customer's screen, hear their voice or show them a banner.

2. Brand experience has to work for machines as well as people. Much brand investment goes into how things look and feel: page design, tone of voice, the in-store moment. An agent cares about whether information is clear, accurate and reachable, whether actions complete and whether limits are explained. Companies whose service depends on ambiguity, friction or fine print may find agents expose it quickly.

3. Loyalty is tested on performance. If an agent can compare, book and complain on a customer's behalf, it is reasonable to expect switching to become easier. In this analysis, loyalty built on habit and inconvenience looks weaker than loyalty built on reliable service and fair terms. Programmes that reward the customer, not the effort of staying, are better placed.

4. Control is negotiated, not assumed. PAP gives companies a say. They choose which routes agents can use and what agents may do. But the standard starts from the premise that agents will arrive. Companies that refuse to engage still face agents that browse their public web pages; they simply lose the structured visibility and control the protocol offers.

What is still open

This is an early proposal, and boards should treat it as such.

It is a draft. The published specification is labelled a draft, with a path to 1.0 still to run. Details can change, and early adopters carry the cost of rework.

Governance is thin so far. Sierra's announcement does not describe a formal standards body, voting process or independent steward. For a protocol meant to sit between consumers and companies across many sectors, how decisions are made will matter to regulators and to companies wary of handing influence to the two companies that started it.

Competing approaches exist. PAP is not the only effort to define how AI agents work with businesses, and it builds on existing standards such as OAuth, MCP and OpenAPI rather than replacing them. Which approach, or which combination, wins adoption is not settled. A wider list of supporters has circulated but has not been confirmed by the companies, so this article names only the partners Sierra lists.

Liability and consumer protection are unresolved. The announcement deals with permissions and sessions. It does not, and could not, settle who is responsible when an agent makes a mistake, buys the wrong thing or agrees to terms the customer would have refused. Those questions will fall to contracts, consumer law and, in some markets, AI regulation.

Payments are not yet in. Purchases without sharing card details are listed as a possible extension, not a finished feature. For retailers, the commercial value of agent commerce depends on that piece.

Where companies stand today

Most large companies already run some of the building blocks PAP relies on. Many offer OAuth sign-in for partner apps, publish APIs and operate service chatbots. What is new is the idea that the counterparty on the other end is a customer's own agent with delegated authority, arriving at scale.

The practical gap, in this analysis, is less technical than organisational. Customer-experience, digital, security, legal and brand teams usually own separate parts of the customer journey. An agent-facing channel cuts across all of them. Someone must decide which tasks agents may perform, what an agent session looks like in fraud monitoring, how agent-initiated complaints are counted in service metrics, and how marketing reaches a customer whose agent filters what they see.

Contact-centre economics change too. If agents absorb routine requests, human service teams handle a smaller but harder mix of cases. That shifts hiring, training and performance targets. Genesys's involvement suggests contact-centre software suppliers are planning for that mix.

A board-level checklist for agent-facing commerce

The following steps are WCRC's suggestions, not requirements of the protocol.

• Map the top ten customer tasks by volume and value, and decide which you would be comfortable letting an authorised agent complete, read-only or with write access.
• Audit your public information for clarity: prices, policies, stock, service terms. Assume a machine will read every line and act on it.
• Bring security and fraud teams in early. Agent sessions need to be distinguishable from human ones and from malicious bots. Decide how permissions are granted, logged and revoked.
• Revisit loyalty economics. If switching gets easier, test which parts of your proposition hold up when a customer's agent compares them line by line.
• Track the specification. Assign an owner to follow the draft through to 1.0, attend design workshops if relevant and decide when, not whether, to run a pilot.

What it means for leaders

• Treat personal agents as a new customer channel with its own owner, metrics and risk controls, not as an IT integration.
• Decide your agent policy before agents decide it for you: which tasks are open, at what permission level and through which routes.
• Make product, price and policy information complete and machine-readable; ambiguity that people tolerate, agents will expose.
• Re-examine loyalty schemes that rely on friction; reward behaviour that survives easy comparison.
• Watch governance and payments extensions closely; they will determine whether PAP becomes a durable standard.

Quick answers

Q: What is the Personal Agent Protocol? A: An open standard proposed by Meta and Sierra on 6 October 2026 that defines how consumers' personal AI agents discover businesses, sign in with the customer's permission and complete tasks through websites, APIs or a company's own agent.

Q: Which companies support the Personal Agent Protocol? A: Sierra names Genesys, Instinct, Rocket, Shopify, Stripe and Walmart as partners, alongside Meta and Sierra. Other reported supporters have not been confirmed.

Q: Is the Personal Agent Protocol ready to use? A: Not in final form. A draft specification has been published under the Apache 2.0 licence, and its authors say design partners are working towards version 1.0.