What Personal Agents Do to the Front Door of Finance
Oct 12, 2026 | Kanishk Walia | MiddleGame Ventures
It is a Sunday evening in late September and Camille’s motor insurance renewal has arrived by email: up 23% on last year, for a car that has not moved house, changed driver or acquired any points. She forwards the email to a WhatsApp contact with a one-line instruction (“find me the same cover for less and switch me if it is worth it”) and goes to bed. By the morning the contact has read the policy schedule, obtained quotes from eleven carriers on identical terms (same excess, same named drivers, same protected no-claims bonus), found one that is €412 cheaper, declined the renewal and bound the new policy from the renewal date. It has also noticed that she has been paying €9.99 a month for a streaming service she last opened in March, and asks whether she would like that cancelled as well.
The contact is not a broker. It is a personal AI agent (Instinct, in this example, which lives inside WhatsApp and operates a virtual desktop on the user’s behalf) and the scenario above is not a thought experiment. One user of Muse, Meta’s equivalent, reported saving $3,500 a year moving from Liberty Mutual to Progressive on matched terms without visiting an aggregator. Muse launched on 8 September, passed 2.8 million downloads inside a fortnight and displaced ChatGPT at the top of both US app stores. Instinct went from a c.$500M valuation to c.$2.5B in the second half of August and was reported in September to be raising at c.$10B.
The question we have been asking ourselves at MGV is what this does to financial services, and we would posit that the answer is larger than the current debate (which has focused mostly on insurance renewals) suggests. Several questions arise. Which parts of a bank’s, an insurer’s or a wealth manager’s profit depend on the customer not acting? Which parts depend on the customer coming to the firm’s own app, portal or branch in order to act? What happens to each when the customer delegates both the decision and the execution to an agent that never tires and feels no loyalty? And who, in that world, owns the customer?
Lets start with the last question because the market has already begun to answer it in some ways. In the last week of September, Expedia agreed to let Muse users search and book hotels inside the agent, with Expedia remaining merchant of record but no longer the interface, and its shares fell 7.7% the following day. Over the same weekend Amazon blocked Muse from its store, citing its conditions of use (c.$68B of Amazon’s revenue comes from sponsored listings, which only work if a human is looking at the page). Financial stocks were caught in the same move: JPMorgan and Wells Fargo each fell more than 3%, Schwab close to 5%, and the insurance brokers Marsh and Arthur J. Gallagher 2.4% and 4.6% respectively (Reuters).
The pattern is one we have seen in every platform shift. Whoever owns the point at which the customer takes the action (what we would call the interaction layer) captures most of the profit, and the systems that owned that point before the shift are pushed down a layer into middleware, where they remain necessary and continue to be paid, but on a supplier’s margin rather than an owner’s. Travel offers the best precedent, since the GDSs owned the booking until the OTAs took it from them in the 2000s, and Booking Holdings is today worth several times the GDS operators combined. If the next generation of travelers describes a trip to an agent and the agent sources the room from Expedia, Expedia is in the position Sabre was in twenty years ago.
Finance is already layered in this way. The card networks, core banking systems and clearing houses are plumbing; the customer relationship sits with the bank’s app, the investment platform, the comparison site, the mortgage broker (who by some estimates arranges c.90% of new UK mortgages) and the insurer’s website. We argued in our February letter that the question for banks is not whether they survive but where they end up in the stack, and the agent is what moves them down it. We already know what a bank earns when it sits behind someone else’s interface, because Banking-as-a-Service demonstrated this: a thin margin and most of the regulatory risk. It may be the case that the mobile app, the industry’s best move of the last decade, turns out to have been the last front door banks owned outright.
It is worth being clear about who the new owner is. Muse belongs to Meta, as does WhatsApp, inside which Instinct runs; OpenAI has launched Dots, and Apple and Google will surely follow. The front door of finance is consolidating around the companies that already own the customer’s attention, not around a new generation of FinTech, and their relationship with a bank is likely to resemble an app store’s relationship with a developer.
How incumbents respond is fairly predictable. The dominant players will do what Amazon has done and block, arguing that an agent acting on an account is not the customer; challengers will do what Expedia has done and open up, because they have share to gain and nothing to defend. Food delivery and other markets have followed the same sequence: challengers embrace, incumbents resist, and incumbents eventually give way on worse terms than they could have negotiated at the outset. Two things change the timing in finance. Regulators may remove the option to block, since open banking in the UK and FIDA in the EU both rest on the customer’s right to hand access to a third party of their choosing. And opening up has a cost of its own: a bank that lets agents shop its rates will win that business and, as Expedia found, be valued as a supplier for doing so.
Losing the front door results in a loss of distribution margin, but the larger loss is the margin that sat behind it. A material share of profit in finance exists because customers do not shop around, switch, claim, cancel or renegotiate: the savings rate on an account opened three years ago vs the one advertised today, the renewal premium vs the new-customer price, the mortgage left on the standard variable rate, the card balance at 25% while a 0% offer goes unused, the claim that was under-paid and never contested. Only one regulator has tried to measure this. Before the FCA banned price walking in January 2022, c. 6 million loyal UK home and motor policyholders were paying c. £1.2B a year over the odds, and the FCA valued its remedy at £4.2B over ten years. We would expect the figure across savings, cards, mortgages, investments and pensions to be a multiple of that.
None of this makes customers irrational. Not switching is a trade of certain time against an uncertain saving, and for most people the time is worth more; the subscription economy, among others, depends on it. Agents change the time side of that trade, which is why we think they will succeed where comparison sites and open banking did not. Those reduced the cost of looking, and switching barely moved, because looking was never the hard part. An agent also does the switching (the forms, the direct debit), keeps watching (daily, not at renewal) and reads the small print, comparing like with like, which is the detail Camille’s agent got right and comparison sites never reliably managed.
What does this look like sector by sector? In retail banking, the current account will survive, because the agent needs somewhere to keep the customer’s money and a way to move it, but the reason to open the bank’s own app will not. Underneath that sits the oldest trick in retail banking, which is paying new customers an attractive rate on savings while leaving long-standing customers on a much lower one; the FCA found that the largest banks kept doing this even as interest rates rose. An agent that moves spare cash to the best available rate every day puts an end to that, and banks will have to pay more for the deposits they fund themselves with. They may respond by offering one set of rates to agents and another to customers who have not delegated, but this would be a temporary fix, since the group of customers who have not delegated will decrease with time and regulators will ask why two customers with the same risk are paying different prices. There is also a question for bank supervisors, because money that moves every day in search of a better rate is not the stable funding that the rules governing bank liquidity assume it to be.
In lending, the work a mortgage broker does today (keeping track of when a fixed-rate deal ends and finding the next one) is done by the agent at no cost to the borrower, so nobody drifts onto the lender’s expensive standard variable rate because they forgot to remortgage. Credit cards are more exposed still, since moving a balance to a 0% offer is a saving most people know about and few get round to, and an agent will do it as a matter of routine. There is a benefit for lenders as well: an agent holding a year of verified bank transactions makes a cleaner loan application than any form, which should make lending cheaper to process and open it up to people with thin credit files. The uncomfortable consequence is that the customers left behind, still paying the old rates, will be those who are least engaged and often least able to afford it.
In insurance, the price comparison sites played the role the online travel agencies played in travel, and an agent that gathers quotes and completes the switch does to them what they once did to insurers’ call centres. UK insurers are somewhat better placed than they realise, because the FCA already banned charging loyal customers more than new ones; US insurers have no such rule, which is why the savings Muse users report are so large. Claims will be another vector along which we expect things to change. Insurers benefit every time a customer under-claims, accepts a low settlement, or gives up on a rejected claim rather than spend five hours on hold, and an agent that files the claim, argues it and escalates it removes that benefit. We would expect insurers’ claims costs to rise for reasons that have nothing to do with the underlying risk.
In payments, the card networks saw this coming earlier than most, which is why Visa, Mastercard, Stripe with OpenAI, and Google have each published a standard for how an agent proves who it is and what it is allowed to spend. Whoever controls that permission controls the relationship; if the agent platform does, the card network becomes a pipe. The more immediate effect is that an agent, rather than habit, decides how each payment is made. Points and cashback have little hold on software that treats them as one factor among several, and paying directly from a bank account, which has struggled for a decade to win customers away from cards, becomes attractive the moment the thing choosing the payment method is a machine optimising for cost. An agent also needs a place to hold money and a way to spend it without a person approving each step, and a wallet with a pre-agreed spending limit paying in a stablecoin may be the better alternative than any existing rail. As such one would expect stablecoin payments to grow fastest in transactions initiated by agents rather than in the everyday consumer uses the industry has been chasing.
In wealth management, the reason customers stayed with their investment platform was largely that moving was a chore, and Schwab’s share price fall was a result of the market working out what happens when it is not. Whether an agent is allowed to tell a customer to move their pension, or only to show them the options, is the question regulators will have to settle, and it is the wealth industry’s equivalent of Amazon’s block. Products that are chosen by a machine need to be transparent, cheap and easy to compare, which is bad news for funds that are sold on a story. In the markets themselves the change is smaller, because large investors already trade with each other through software, but a few things follow: brokers will have to prove they got agent-directed customers the best price, regulators will want to understand what happens if many agents sell at the same moment, and assets that settle instantly on a shared ledger are the natural fit for trades that software initiates, which is the second place this thesis meets our Digital Asset thesis.
What does this mean for how we invest? First, the two kinds of profit disappear in different ways. The profit from owning the customer relationship goes quickly, because the market prices it in as soon as the threat is visible (Expedia was re-rated within a day of its announcement), whereas the profit from customers not acting erodes one customer at a time as agents are adopted, which is slower and harder to reverse. We should now ask of every company we look at which of the two it depends on. Second, we do not believe the opportunity for founders and thus venture investors is in building and backing the agent itself. Distribution, technology and money all favour Meta, OpenAI, Apple and Google, and a start-up that sets out to be the customer’s agent for money is competing with WhatsApp for the customer’s attention (Instinct’s valuation suggests other investors disagree, and it is too early to say who is right).
Our view is that the durable returns sit a layer down, in (i) the plumbing that lets banks and insurers deal safely with agents (ways of verifying an agent, standard forms of permission, product terms a machine can read, and compliance tools for the interaction), (ii) financial products designed from the start to be chosen by a machine rather than marketed to a person, and (iii) the settlement layer agents will prefer, where stablecoins and tokenised assets finally have a payer that is itself software. Third, founders should build to be chosen by agents rather than to compete with them for the customer. The successful FinTech of this cycle may never meet its customer, and will be selected again and again, without loyalty, because it offers the best terms on the day and makes them easy for a machine to find.
If you are a founder building in this space or interested in speaking with us further on this topic, do not hesitate to reach out for a conversation or coffee.
