What personal AI agents mean for ecommerce customer service
8 min read
In 2023, Gartner predicted that by 2026, 20% of inbound customer service contacts would come from “machine customers” (Gartner, 2023): software acting on a person’s behalf. It’s now 2026. Whether or not that exact number lands, the direction is clear, and it’s picking up speed.
Personal AI agents are now mainstream consumer products. They can email your support team, sit in your chat widget and, increasingly, phone your contact centre. They don’t mind waiting, and they don’t get tired of asking.
For ecommerce brands that raises two problems: more contacts, and more persistent ones. Most of the advice so far is about spotting and blocking agents. We think that’s the wrong place to start. The brands that cope best will be the ones whose policies give the same answer whoever, or whatever, is asking.
What are personal AI agents?
Personal AI agents act on behalf of a consumer. They run in the background across email, messaging apps, the web and, more recently, the phone, completing tasks such as booking, cancelling, chasing bills and making purchases. Meta, OpenAI and Google all launched one in 2026, and they are starting to show up in customer service queues.
They are different from the AI agents that brands put on their own websites, although both get described as “agentic commerce”, which doesn’t help. This article is about the consumer’s agent: the one that turns up in your queue.
Here are the main ones so far:
**Muse (Meta, September 2026, US):** sends emails, fills out forms, lowers bills and makes purchases. Phone calls to businesses were added a week later.
**dots (OpenAI, September 2026):** always-on agents in ChatGPT, connected to more than 4,000 apps, rolling out to ChatGPT Pro and Business Premium users first.
**Gemini Spark (Google, May 2026, US):** sends emails on the user’s behalf and makes purchases, with approval required for high-stakes actions. Google AI Ultra subscribers got it first.
Instinct Concierge (Instinct, September 2026, early access): phones businesses to resolve billing issues, among other tasks.
**Pine (Pine AI, founded 2024):** built for customer service, handling calls, emails, complaints, refunds and disputes.
**OpenClaw (open source, went viral in early 2026):** runs through messaging apps such as WhatsApp and iMessage, and passed 100,000 GitHub stars by February 2026.
There is serious money behind this. Instinct raised $350 million at a $2.5 billion valuation in August 2026. Pine was built specifically for customer service: it negotiates bills, files complaints and disputes charges, and says its users save around $400 per interaction on average, although that is the company’s own figure.
Ecommerce is an obvious target for these personal agents. By connecting to a customer’s inbox, an agent can find the order number, chase a refund, start a return or check on an order’s status with very little effort from the customer.
Will personal AI agents increase customer service volumes?
Yes. Contacting customer service costs the customer time and effort: finding the order number, waiting on hold, writing the email and following up. Personal agents take that cost away, so issues that weren’t worth the effort (a late parcel, a small fault, a price drop) start to become reasons to get in touch. Gartner found 51% of customers would be willing to use a GenAI assistant for customer service on their behalf (Gartner, 2025).
Plenty of customers contact you regardless. But for some issues, the time involved decides whether a ticket ever exists. Remove the barrier of time and those tickets will start to appear.
Agents can also chase several brands at once and follow up on a schedule. Contact-centre analytics vendor MiaRec has forecast that volumes could rise three to five times as assistants lower the effort of seeking help (reported by CX Today, 2026).
Most of the extra contacts will be routine, such as where is my order, can I return this or can I change my address. If your own AI agent resolves these routine requests end to end, you may not notice the extra volume in your day-to-day work, only when you look at the dashboard.
What happens when persistence costs nothing?
A person might give up after being told no once, or maybe twice. An agent can rephrase, re-ask and escalate indefinitely, at virtually no extra cost to the customer. Brands need to make sure their own AI agents keep saying no when the situation calls for it.
The risk is greatest when a brand’s AI agent runs on prompts alone. Those agents are designed to be helpful to the user, and a helpful system can be argued into a refund it shouldn’t give. We’ve written before about why prompt-only AI agents fall short in ecommerce, and personal agents make that weakness more expensive.
Persistence can also lead somewhere less obvious. A personal agent that has been asked to get a refund, and keeps hearing no, may start adding details to strengthen the case: a parcel left out in the rain, a fault that appeared on first use, a promise from an earlier conversation that nobody can find. The agent may not be setting out to deceive anyone. Language models fill gaps with plausible detail, and an agent working towards a goal has every reason to fill those gaps in its user’s favour. If the brand’s AI judges a claim by how convincing the conversation sounds, invented details will work.
None of this starts from zero. Around 9% of all US retail returns are already fraudulent, out of $849.9 billion of returns in 2025 (NRF and Happy Returns, 2025). Personal agents make claims cheaper to make and easier to repeat, which matters most for brands whose checks rely on judgement.
Most people using personal agents will be honest customers trying to save time. A response that treats every agent as a fraudster would punish the customers you most want to keep.
Can brands tell when an AI agent is contacting them?
Not reliably. Identification standards exist, such as Visa’s Trusted Agent Protocol and Mastercard’s Agent Pay, both built on Cloudflare’s Web Bot Auth. However, agent builders have to opt in, and the standards were designed for checkout traffic, so they don’t cover service conversations. An agent that doesn’t want to be identified usually won’t be.
The Cloudflare announcement is a sensible step: agents can authenticate themselves so merchants can see “the identity and intent of agent traffic” (Cloudflare, 2025). But it depends on agent providers adopting it, and nothing obliges an agent that is sending an email or calling a helpline to say what it is.
Some providers won’t. In its case against Perplexity, Amazon alleged that the Comet browser agent disguised itself as an ordinary Chrome session to avoid detection (GeekWire, 2026). In writing, a well-written message from an agent looks the same as one from a person.
Guessing wrong is also costly. If you accuse a real customer of being a bot, you’ve damaged the relationship over a hunch. The safer assumption is that you often won’t know, and to design your processes with that in mind.
Should retailers block personal AI agents?
Blocking is possible, but it’s costly and legally uncertain. Amazon won an injunction against Perplexity’s shopping agent in March 2026, and the Ninth Circuit vacated it in August. Turning away agents also turns away genuine customers who simply chose to use one.
The March ruling found that Comet accessed accounts “with the Amazon user’s permission, but without authorization by Amazon”. On appeal, the Ninth Circuit vacated the injunction on 4 August 2026, holding that under the Computer Fraud and Abuse Act it is the user, and not the agent’s developer, who accesses the site. The court pointed platforms towards terms of service and technical measures instead (Jones Day, 2026).
For most brands the commercial question matters more. If personal agents become a normal way to shop and complain, refusing to deal with them makes you harder to buy from, and it could become a competitive disadvantage.
Broadly, brands have two options: stop it or handle it. Stopping it only works where you can identify the agent, whereas handling it works either way.
The DigitalGenius view: build policies that hold up whoever is asking
The most reliable defence is a consistent answer. Let AI handle the conversation, and make refund, return and replacement decisions through deterministic checks against order, carrier and warehouse data. That way, persistence and well-argued claims can’t change the outcome.
In practice that comes down to five things.
Let rules make the decisions and the model handle the language. A language model is good at understanding what someone wants and writing a clear reply. It shouldn’t be the part of the system that decides whether a refund is due. In the DigitalGenius AI Engine, specialist AI agents handle the conversation, and deterministic Genius Flows make the decision using data from the order management system and the carrier
Check every claim against the data. When a claim includes details, test them against what you already know. Check the order history, what the carrier recorded on delivery, whether the return has actually reached the warehouse, and whether the same item has already been claimed. For damage claims, ask for a photo: DigitalGenius uses visual defect recognition to check customer images alongside the other checks. These are the same checks that stop fraudulent returns today, and they work just as well when the details come from an agent.
Give the same answer on the fifth request as on the first. If a policy answer changes because a request was rephrased, the personal agent’s persistence pays off. Deterministic checks give the same result however many times they are run.
Watch for behaviour that suggests persistence. You may not know whether you’re talking to an agent, but you can see how a conversation behaves: rapid, repeated rephrasing, refusal to accept a policy answer, or the same claim made across several orders. Flag those conversations and either route them to a person or close them politely, whoever is behind them.
Automate the legitimate volume. Most agent-driven contacts will be genuine and routine. Resolving them end to end, with the order looked up and the action taken, stops a rise in contacts from turning into a rise in headcount.
This approach has trade-offs. Rules need maintaining, and policies have to be explicit enough to be enforced automatically, which often exposes where they are vague today. Edge cases still need people: a high-value customer with a genuine problem the rules don’t cover should reach a human quickly, with the full context. That’s a reasonable price for a customer service operation that behaves the same way whether it’s talking to a person, an agent or a person using an agent.
FAQ
Are personal AI agents legal to use with customer service?
Generally, yes. It’s similar to asking someone to email on your behalf. In Amazon v Perplexity, the Ninth Circuit held that the user, and not the agent’s developer, accesses the site. Brands can still set terms of service, and the law is moving quickly, so take legal advice on specific cases.
Will personal AI agents replace customers contacting brands directly?
Not entirely. Many customers will still want to deal with a brand themselves, especially for high-value or emotional issues. Agents are most likely to take on routine, time-consuming tasks such as chasing a delivery, starting a return or disputing a charge.
How should ecommerce brands prepare customer service for AI agents?
Start with your policies. Make refund, return and replacement rules explicit, enforce them through checks against order and carrier data, and make sure the same request always gets the same answer. Then automate routine contacts end to end so that extra volume doesn’t require extra staff, and flag persistent or repetitive behaviour for review.
What’s the difference between personal AI agents and agentic commerce?
“Agentic commerce” is used for both sides. Brand-side agents help shoppers on a retailer’s own site and answer their service questions. Personal AI agents work for the consumer and contact brands on their behalf. The two increasingly meet in the middle, with a consumer’s agent talking to a brand’s agent.
What changes now
Personal agents reward clarity. A brand with explicit policies, enforced against real order data, will find that agents make its customer service faster and cheaper to run: there will be more requests, but each one will be routine and resolved in seconds. A brand that relies on its team or its AI making judgement calls in the moment will find that judgement tested at a volume, and with a level of persistence, that no team has dealt with before.
For most brands, the most useful work over the next six months is to write down what their policies actually are, and to make sure their systems apply them consistently.
