Agentic commerce has spent the year at the checkout. Last month Walmart laid out Agent Profiles, an Agent Gateway, and house rules for outside artificial intelligence agents that shop on customers' behalf. Amazon is blocking those same agents. Sierra and Meta have drafted a protocol, nicknamed Poppy, for how personal agents introduce themselves to businesses, and a long list of retailers, banks, and payment networks have signed up to help design it. Everyone is arguing about who gets to buy.
That's the wrong end of the transaction to watch. An agent that can buy the vacuum will be asked to return it when it stops charging in week five. Then it will be told to file the warranty claim, chase the refund, and dispute the charge that posted twice. Buying was never much work. Getting something fixed means finding receipts, reading policies, locating the right department, telling the story three times, and following up when nothing happens. That is the errand consumers most want to hand off. Agentic commerce ends up in customer service whether customer service invited it or not.
Here's the uncomfortable part: A good share of service economics depends on customers giving up. Rebates that go unclaimed, disputes that get abandoned, and cancellation paths with one step too many are all revenue kept or costs avoided. A personal agent doesn't tire, doesn't forget, and reads the full terms and conditions.
So when contact volume climbs because agents pursue the issues people used to drop, that isn't a capacity problem. It's the real cost of your friction showing up on the books. The answer is a policy decision, not a staffing plan.
Blocking won't hold, either. In commerce, a blocked agent is a lost sale. In service, it's an existing customer with a legitimate problem who can't reach you through his chosen representative. Customers will notice. So will regulators.
Reachable Isn't Authorized
Gateways, protocols, and Model Context Protocol endpoints all solve the same problem: getting the agent to the business. Connection was never the hard part. Every consequential request still has to clear four questions. Whose agent is this? Does the permission it carries cover this action, on these terms? Is the claim true? Did the remedy actually happen?
Those are separate questions, and service organizations tend to blur them. Permission to ask about a bill is not permission to accept a settlement. A receipt supplied by an agent is a claim to verify, not a fact. And when the case escalates, the human specialist inherits the same limits. Asking for a person doesn't widen what the agent is allowed to see.
Most contact centers will run into this on the phone first. Google began testing a Gemini feature in September that calls a business, works through the menus, and waits on hold. None of the new protocols cover the telephone. Your first personal agent is likely to arrive in the IVR, and the best your systems can offer is a guess that the caller is a machine. That says nothing about whom it represents or what the customer approved.
Two moves are important now, and neither is a new bot.
First, own the record of what your agents decide. Call it a decision ledger: what was decided, under which policy, on what evidence, with whose approval, written when the refund or replacement hits the system of record and held in systems you control. The vendor's transcript store is not that record. Disputes outlast retention windows, and they outlast vendor contracts.
Second, measure and pay for resolution. Containment counts a conversation as handled even when the refund never posts. A human customer in that spot gives up. A personal agent calls back tomorrow, gets counted as new demand, and if you pay per conversation, you're billed twice for one failure. Even pricing that's labeled resolution-based usually judges the outcome from the conversation, not from whether the money moved. Ask your vendor how it knows.
Commerce asked whether an AI agent could be trusted to spend a customer's money. Service asks something harder of the business: Can you show what you promised and what you delivered? Personal agents will keep coming back until the problem is fixed. The only durable response is to fix it and to be able to prove you did.
Derek Top is principal analyst and research director at Opus Research.