AI can find your brand. Can it see why customers should choose you?

Businesses are being encouraged to prepare for customers who arrive with personal AI agents. Make products discoverable. Publish structured information. Enable agent-led transactions.

All sensible work. But the more demanding question is this:

Can an agent establish why your offer is right for this customer, and help them actually receive the value you promise?

On August 19, we wrote about the Agent-in-the-Middle: the customer's AI increasingly standing between a brand's message and its intended audience. In September, The Customer Journey Without the Customer examined what happens when customers delegate the work itself.

Now businesses are establishing how those agents can identify themselves, access accounts, and act.

The customer's agent is at the door. Businesses are deciding how to answer.

Sierra and Meta announced Personal Agent Protocol on October 6 with partners including Genesys, Instinct, Rocket, Shopify, Stripe, and Walmart. Sierra's October 9 update included NiCE among the earlier participants and added 35 design partners. The draft protocol is known as Poppy (Sierra's announcement, draft and partner update).

The example Sierra showcased made the ambition tangible. At Sierra Summit, Rocket's Alex McGillis asked Meta's personal assistant, Muse, to obtain mortgage pre-approval. According to Sierra's account, Muse discovered Rocket's agent, requested permission to connect, and McGillis signed in to share the necessary information. The agents worked through financing details and produced a pre-approval letter. This was a demonstration reported by the company promoting the protocol, but the potential value is clear (Sierra's demonstration account).

A customer states a goal, grants permission, and agents handle the work across organizational boundaries.

There are two visible approaches to that future.

Sierra, Meta's Muse, Walmart, and the wider Poppy group are building controlled ways for agents to represent customers. Walmart's October 21 announcement adds concrete mechanisms: planned Agent Profiles, approval and spending controls, and an Agent Gateway beginning with a developer beta (Walmart's announcement).

Others are restricting access while those arrangements develop. Amazon blocked Muse in September, before Poppy's announcement, after Meta declined its request to remove the bot (Bloomberg's reporting). Axios reports restrictions at JPMorgan, Bank of America, and Citi, and blocking at Morgan Stanley (Axios's banking report).

Bank of America is also a Poppy design partner. That apparent contradiction is instructive. Supporting controlled delegation and restricting unapproved access can be entirely consistent.

The question is which agents may act, under whose authority, through which channels, and with what accountability.

But access only gets us partway.

Three experiments beyond the promise

We ran three experiments to examine the rest: shopping for hiking shoes for a trip to Yellowstone National Park in the United States using Muse and Tasks in Gemini, then evaluating a switch from Google Fi using Muse.

The shopping context was straightforward: the trip was next week. A useful recommendation needed to account for that timing. The mobile task was to find a worthwhile alternative to an existing service, considering actual usage and connected devices.

We wanted to understand what initiatives such as Poppy could enable, and what businesses and personal agents would still need to work out between them unless shared standards evolve to cover it. These were guided consumer exercises, not tests of Poppy integrations.

A clear price. An unresolved purchase.

Muse took the delivery requirement seriously. It insisted on the costlier Salomon option to preserve arrival timing, without first establishing whether express shipping could make the cheaper Columbia option a better purchase.

The Columbia shoes were listed at $129.99 against $175 for the Salomon alternative. That left a $45.01 product-price advantage that could potentially absorb an express-delivery charge. Whether it actually would remained unverified.

Muse also compared the displayed product prices only. It did not account for sign-up, promotional, or member offers shown on the retailer’s site, which a human shopper would typically weigh. Their eligibility and effect on the final price still needed checking.

The missed step was comparing the complete offers: exact shoes, delivery options, arrival dates, available offers, and the final cost.

For businesses opening up to personal agents, through Personal Agent Protocol, their own MCP servers, or commerce protocols, this means going beyond listing-level information. Product price, sign-up and promotional offers, express-delivery charges, destination-specific arrival estimates, and applicable member benefits need to be connected to the same offer, in a structured form, with their conditions explicit. That gives an agent a stronger basis for comparing what each business can actually deliver for this customer, and at what final cost.

Otherwise, an advantage the business has invested in may contribute nothing to the recommendation.

The promise has to hold from listing to checkout

Tasks in Gemini exposed a different weakness: the delivery promise changed at checkout.

A shopper on a website who sees a later date at checkout can go back and look for a pair that arrives sooner, often on the same site. An agent comparing several sellers may not. A changed delivery date creates a risk that it discards the offer rather than exploring alternatives with the same retailer. With agents, a business may not get a second chance to be chosen.

Delivery estimates can legitimately change with location, inventory, or order cutoff times. The point is not that estimates must never change. It is that an agent needs to know when they do.

For businesses, the information should stay consistent as the purchase progresses. Where an estimate changes, the business should make the change explicit and offer current alternatives that are easy for the agent to retrieve.

Ask first, recommend second

The Google Fi exercise exposed a weakness on the agent's side.

Muse recommended a lower-data Fi plan before it knew how much data was actually being used. Only after reviewing a bill screenshot did it correct its estimate from roughly 20 GB to about 37 GB, which ruled out its first recommendation. It also separated benefits already included in the current plan from premium features the customer didn't have.

Muse recovered well. It broadened the comparison and recommended US Mobile Unlimited Starter on an annual plan, using the AT&T network, after the remaining Fi promotional month. It proposed checking the Galaxy Watch and said it would not cancel Fi or port the number before final checkout approval.

But the order was backwards. There is little value in a long, detailed recommendation that ends by asking for usage, devices, and current benefits. Those questions belong at the start.

Businesses can help. An incumbent that gives an authorized agent a current view of the plan, benefits, and post-promotion pricing saves the customer from reconstructing it. But no business data fixes an agent that recommends before it asks.

Getting connected is only the beginning

Poppy's published 0.2 specification addresses discovery, identity, authorization, access to business interfaces, and conversations. It leaves those interfaces to define their domain-specific information. Recognizing an agent does not establish whether a shoe will arrive in time or a mobile plan supports a particular watch (Poppy specification).

Other standards already represent parts of the value proposition. Universal Commerce Protocol, developed by Google with industry partners including Shopify, documents fulfillment, discounts, and loyalty. Agentic Commerce Protocol, developed by OpenAI and Stripe, provides structured commerce interactions (Google on UCP, UCP Loyalty, ACP overview).

The gap is not an absence of fields. It is whether the benefits, conditions, evidence, and actions form a dependable offer for this customer.

For marketing, differentiation must be something an agent can substantiate. For product, the offer must survive checkout, with changes flagged before commitment. For contact centers, the promise and permissions must carry into recovery.

If the agent buys the wrong item, who can initiate the correction? If delivery misses an agreed requirement, what remedy applies? Can the authorized agent resolve it or reach a person without making the customer explain everything again?

Permissions to research, purchase, and cancel remain distinct. So do completing payment and delivering value.

Two purchases. Both need to earn their cost.

Things are changing fast. Sierra's mortgage example shows useful work being removed from the customer. Walmart is beginning to design explicitly for delegated shopping.

But early-adopter enthusiasm can obscure the distance between an impressive interaction and an experience worth using repeatedly.

The Fi exercise is a small example of that distance. The first answer was long and confident. The useful one arrived only after the customer supplied details the agent should have asked for at the start.

An ordinary customer makes two purchases: the personal-agent subscription and the purchases delegated to it. Both need to earn their cost.

Amazon provides an interesting counterpoint. It blocked Muse, while Subscribe & Save lets customers schedule recurring deliveries with discounts on eligible products, advance reminders, and options to adjust or skip. It handles a narrower job, but a useful one (Amazon's consumer explanation).

Personal agents promise to handle far more complicated jobs. To earn a place in a customer's routine, they have to remove more work than they create.

Until they do, the smartest shopping assistant in many homes may still be the one that never says a word: the automation quietly restocking the cupboard.

---

About the experiments: These were three individual, guided experiments, not a controlled comparison or protocol interoperability test. The account combines recoverable screenshots with the experimenter's observations. Historical prices and carrier claims are not current verified offers. Completed purchases, watch verification, and a carrier port were not established in the reviewed record.