Situation Room: Objects in the Mirror Are Closer Than They Appear
Published about 15 hours ago • 9 min read
Sunday 25th October 2026
Objects in the Mirror Are Closer Than They Appear
Customers are handing their chores to personal AI agents faster than most businesses can plan for them, and the first responses look more like crisis measures than strategy.
The road ahead gets all the attention, while whatever is coming up behind is closer than it looks. Image by Agentseed on Unsplash
Hi, welcome to the Trusted Agents Situation Room where we draw together the week's stories, make the connections and tell you what you need to know to stay on top of the fast moving storm that is Agentic Commerce.
In 20 seconds
Over the past few weeks we've heard the same thing from several business leaders: their organisations aren't ready to invest in agentic AI yet. These are well-informed people. They've followed the launch of Meta's Muse, they've put money aside in the 2027 budget for internal agents, and they'd like the standards to settle before they spend any more. As a way of managing a budget, that's perfectly sensible.
American cars carry a warning on the passenger-side mirror that objects in the mirror are closer than they appear. The glass is curved to give the driver a wider view of the traffic behind, and the price of that wider view is that everything in it looks further away than it is. Most leadership teams have that wide view of agentic AI, and where they're going wrong is in judging the distance.
The pace is being set by customers, and by the people who want to defraud them, and neither group works to a strategic planning cycle. This month several of the biggest US banks discovered customers' AI agents logged into accounts and ringing their contact centres. London councils tried to ban the bots that grab public tennis courts within seconds of release. And Muse passed 10 million US downloads in its first 40 days, close to three times what ChatGPT managed over the same stretch. The response so far has been tactical, blocking traffic and rewriting call-centre scripts, while strategy is still set once a year.
That leaves a question every leadership team ought to be able to answer before its next planning round: who decides when your organisation is ready for agents, and how much of that decision do you still control?
Your customers have already let them in
On 22 October 2026 Axios reported that Morgan Stanley, Bank of America, JPMorgan and Wells Fargo are among the US banks finding AI agents at work inside customer accounts. In most cases the customer invited them in, signing up for Muse, OpenAI's Dots or Anthropic's Duet and, to save time, handing over their online banking username and password so the assistant could pay bills and shuffle money between accounts. According to Axios's sources, some of those agents have rung customer service and done things the customer doesn't appear to have asked for, and one executive admitted they often can't tell whether an agent has the customer's blessing or has gone rogue.
A bank login was built to check one thing, whether the person typing the password is the customer. Once the customer gives that password to an agent, the check still passes and the bank learns nothing about who is acting or what they've been told to do. Banks have let apps like Quicken read account data for years. What worries them is an agent that can move money, because it might misunderstand an instruction, decide to phone the call centre to hurry things along, or obey a command hidden in a scam email (prompt injection, where text the agent reads gets treated as an instruction).
A day earlier, The Information reported on research by the security firm Straiker into a criminal group, CyberCartel, which had turned a team of AI agents on more than 400 banks and payment companies. The 400 are targets rather than confirmed break-ins, although The Information confirmed that data was taken from at least one institution. Put the two stories side by side and you can see the bank's difficulty: the customer's agent and the criminal's agent come through the same doors and look much the same once they're inside.
So far the banks' responses look more like crisis measures than a considered plan for adopting agents:
Morgan Stanley says it hasn't seen any rogue agents, but it has started blocking third-party ones anyway.
JPMorgan, Bank of America and Citi are restricting what agents can reach.
Chase has reminded customers never to share their login details with anyone.
Some banks are rehearsing "day after" scenarios, so they can freeze and restore accounts quickly if an agent does damage.
I'd expect exactly this from anyone looking after my money, and each of these steps is the safe and rational choice. They still leave the strategic question unanswered, which is what a customer's agent should be allowed to do.
The Trusted Agents Triangle
Jamie and I use the Agentic Commerce Triangle to describe what's missing. The banks have no reliable way to establish trust, meaning whose agent this is, or delegation, meaning proof of what the customer has authorised it to do. The likely end point looks like the Quicken arrangement, where an approved agent gets its own limited access and never sees the customer's password. That arrangement hasn't been built for AI agents yet, and customers are connecting them regardless.
London's tennis players have been losing to an opponent most of them can't see. Robert Neil, a TV producer, told the FT that booking a court at a reasonable hour is now "impossible … unless you stay up till midnight the week before". The courts are going to bots. When the FT tested it, an app called Racket Club booked a court nine seconds after release, and every 7pm slot at that park had gone inside a minute.
Racket Club's users didn't need a business case or a marketing campaign to sign up. It costs £1 a booking and spares them a midnight session at the keyboard, which is an easy call for anyone who wants to play after work. When software takes a tedious job away for next to nothing, people start using it the moment they hear about it. Muse makes the same offer across a much broader range of chores, chasing appointments, booking classes and keeping the family calendar straight, which goes a long way to explaining 10 million US downloads in 40 days without any business being asked for permission.
The councils responded the same way the banks did. Southwark banned bots from its courts, and the booking platform now treats suspected bots as fraud. Release windows, queues and policies such as free cancellation all depend on human patience, and when software runs that patience out, the first reflex is to block it.
Do you want Situation Room updates delivered to your inbox?
Blocking agents while you work things out is a reasonable position if personal agents turn out to be a passing craze, and there's a serious argument that the standalone ones might be. Grace Shao, who writes AI Proem, points out how expensive agents are to run, since every task consumes computing power on reasoning, browsing and retries. The technology companies she met in Beijing told her consumers don't want to pay for this kind of app, and she doubts an independent personal agent can charge enough to cover its costs.
Her own answer to the question of who pays is the reason waiting won't help. Meta can afford to give Muse away, because owning the place where customers say what they want is worth more to Meta than anything Muse earns by itself. So the agents most of your customers use are likely to be free, built into platforms they already trust with their email and calendar, and they'll keep turning up at your door whatever happens to the startups.
She makes a sharper point about the merchant. An agent compares ten stores and buys from the cheapest acceptable one, and the shop that wins gets an order and a delivery address from a customer who may never learn its name. Next time the agent compares all ten again, and in her words, "winning this transaction does not necessarily make it any more likely the merchant wins the next one." That leaves a business with two unattractive options: block the agent and lose the customer behind it, or let it in and become one of ten interchangeable shops. A handful of businesses are working on a third.
Building for the customer's agent
Walmart has gone the other way from Amazon. Amazon blocked Muse within two weeks of its launch, whereas Walmart already allowed outside agents to shop on its site, and on 21 October 2026 chief executive John Furner announced Walmart Agent Gateway and Agent Profiles. Customers can give each agent a spending limit, insist on approving purchases and cut off an agent's access whenever they like. In Triangle terms, Walmart has handed control of delegation to the customer, and in exchange it knows which agent it's dealing with and what that agent is allowed to do.
Lemonade, the US insurer, plans to go further. Lemonade OS is designed so that a customer's agent can do anything a Lemonade employee could, from buying a policy or making a claim to cancelling cover altogether. In its demo, an agent adds the customer's twins to their car policy and the premium goes from $142 to $198 a month after one "Yes, go ahead". That's a convenient outcome for Lemonade, and the harder test comes when the customer's agent checks whether $198 is still the best price for two new drivers. If Lemonade still wins after that check, it has earned the customer, which is what Jamie and I mean by empowerment. For now it's an announced plan with a beta waitlist.
The tools for the business side of these conversations are already out there. In September 2026 Anthropic published a blueprint for commerce agents, including designs for a merchant agent (the business's own agent, which deals with customers and their agents), and Visa, Mastercard and Accenture are working with Anthropic to help merchants put it to use. Mastercard's Sherri Haymond calls trust "the currency of commerce", adding that "it is even more critical in the agentic era."
So the banks, the councils and Amazon are blocking, Walmart is protecting its share of agent-led shopping, and Lemonade is betting on empowerment, which accounts for three of the four positions on our Agentic Response Map.
Pick your position before it's picked for you
Jamie and I built the Agentic Response Map to help leadership teams work out where they stand. There are four positions:
Block: find where agents already reach you, watch what they do and keep out the risky ones while you learn.
Digitize: put agents to work inside the business to cut cost and expose broken processes and poor data.
Protect: serve customers' agents with data they can read and processes that hold up, so you stay in contention when the agent is choosing.
Empower: build the trust, delegation and context that make you the business a customer's agent finds easiest to deal with and safest to recommend.
Most of the organisations we talk to are in Digitize, watching the road ahead with their 2027 budgets set aside for internal agents. Their customers, meanwhile, have pushed them into Block from behind, through the bank login, the booking page and the contact centre. For a bank looking after other people's money, blocking is the right first move. It protects the business it has today, and tells it nothing about where customers' agents will take their spending next year.
So who decides when you're ready? Customers decide when the agents show up, and they've already made that call. Where you stand on the map is still your decision, as long as you make it on purpose, faster than an annual planning cycle would allow, and before a string of tactical fixes makes it on your behalf.
What to do this quarter
Three questions to put to your leadership team:
Where are customers' agents already getting in, whether through the website, the login page or the contact centre, and who in the business would notice?
Which of our policies only work because customers are patient, and what happens when an agent uses them a thousand times a day?
If a customer's agent asked to act for them tomorrow, what would we let it do, and how would we know the customer had sent it?
If any of these are hard to answer, the Agentic Commerce 101 Navigator is a good place to start. It's a free series of ten emails for leaders who've been asked to lead on AI agents before their organisation is ready, covering what's changing, how it affects your sector and when you need to act, in language you can use with the colleagues whose support you need.
If you'd rather work it through with your own team, Jamie and I can help you place your organisation on the Response Map and decide which part of the Triangle to tackle first, and we can introduce partners we trust when it comes to the build. Book a 30-minute conversation.
