This week I received a government document that I wasn't expecting and, frankly, didn't understand. A few years ago, I know exactly what would have happened next. I would have fretted about it for a few days, put it somewhere safe, forgotten about it, and eventually found myself dealing with the consequences.
This time I did what I suspect many of us now do. I took a photo of it and asked Claude. Immediately, I had a full explanation of what it meant. Then Claude noticed an error that could have cost me thousands of pounds. It told me what I needed to do about it, drafted a response, and gave me the address to send it to.
It is a small, personal story, but it points to something much bigger. A remarkable number of business models, and quite a few public services, depend on people doing what I would once have done with that document: nothing. They rely on us not comparing, not chasing, not checking and not challenging. That is precisely the behaviour AI, and especially agentic AI, is about to change.
Two forces hiding in plain sight
There are two distinct forces at work, and it helps to separate them. The first is inertia: we could act, but we can't be bothered. The second is friction: we would act, but it is too hard, too slow or too complicated. Ben Evans, the highly respected independent technology analyst, uses exactly this split in AI Questions, the latest annual presentation he published this week His question is a simple one: how much money is locked up in hassle and complexity?
The answer is a lot. Banks pay little on balances that sit untouched in current accounts. Insurers price renewals knowing most customers will roll over. Mobile operators keep out-of-contract customers on tariffs they would never choose today; Evans cites one estimate that if all out-of-contract UK mobile customers used an AI agent to optimise their plans, it could amount to as much as 20% of operators' free cash flow. Meanwhile, refunds go unclaimed and entitlements go untaken because the process of claiming them is designed, deliberately or not, to wear us down.
Customers get AI before companies do
My good friend Dave Birch has been making this argument for years. As far back as 2019 he wrote in Wired that the revolution in financial services would come not because banks got AI, but because customers did. Most of the industry has since spent its energy using AI to cut its own costs. That is the inside-out view. The outside-in view is far more uncomfortable.
That future is now arriving. Dave recently described a driver who used Meta's Muse agent to gather quotes, match his existing cover and switch insurer, saving $3,500 a year. JPMorgan Chase has responded to the arrival of such assistants by blocking external AI agents from acting on customer accounts or moving money. McKinsey has estimated that if agents shifted a tenth of current account balances to the best-paying accounts, bank profits on those balances could fall by a fifth. As Dave puts it, the future will be decided by which banks the AI chooses on the customer's behalf. A software agent doesn’t respond to an advert. It doesn’t feel loyalty. And it never gets bored.
Why removing friction was never enough
Here the UK offers a revealing lesson. In 2013 we launched the Current Account Switch Service, which completes a bank switch within seven working days and moves your payments across automatically. It removed most of the friction. Yet more than a decade later, only around 2.5% of adults switched their main account each year, and total switching in 2024 was roughly flat on the year before the service existed.
The lesson matters. We fixed friction, but we never touched inertia. Making something easy does not make people do it. What is different about agentic AI is that it attacks both at once. It does the comparison, makes the decision within limits you set, and completes the paperwork. It removes not just the effort, but the need for us to care.
When arguing becomes free
The second consequence is less discussed and, in my view, more immediate. If AI makes it nearly free to switch, it also makes it nearly free to complain, object, request, and appeal. One slide in Evans's deck captures it in a single line: "The marginal cost of arguing goes to zero."
The evidence is already mounting. Freedom of Information requests to central government reached a record 94,526 in 2025, up 13% on the year before, and in May the Information Commissioner's Office issued guidance to help public bodies cope with AI-generated requests that are longer, more complex, and often misquote the law. The Scottish Information Commissioner reported 120 appeals in two months from just two applicants, apparently generated entirely by AI. The Financial Ombudsman Service found that in one sample up to a third of responses to its initial assessments appeared to be AI-assisted, and since October it has new powers to dismiss complaints.
Researchers at GovAI and the Hertie School have given this a name: "agentic flooding". They identified 84 likely cases across 11 jurisdictions, from German social courts to UK ombudsman services, and concluded that the risk is highest for services that are financially attractive but complex to use. That describes a great deal of what governments and regulated industries do.
The uncomfortable truth is that many of these systems were designed with effort as a hidden filter. The cost of writing a letter, completing a form, or persisting through three rounds of rejection kept demand manageable. Remove that cost and the queue does not just grow. It changes shape. Nor will all of it be honest. Stanford's AI100 report, published this week, describes an experiment in which people who delegated a task to an AI agent with only high-level goals became far less honest: the proportion acting honestly fell from 95% to as low as 12%. Effort was never just a filter on volume. It was also a filter on conscience.
Three things to watch
So, does this disrupt business models and social norms? Yes, but not in the simple "consumers win" way it is often presented. Three dynamics deserve attention.
First, incumbents will put friction back. JPMorgan's block is an early example. In the public sector, the GovAI researchers found that where governments responded with fees or blocks, the burden tended to fall hardest on poorer and less digitally confident people. Rebuilding barriers is easy; rebuilding them fairly is not.
Second, inertia pays for things we value. As Bill Conerly argued in Forbes, the savvy customers who chased the best deals were effectively subsidised by everyone else's inattention. "Free" banking, generous loyalty schemes, and introductory offers rest partly on that cross-subsidy. When everyone optimises, those arrangements will be repriced, and some people will be worse off.
Third, the agent becomes the new gatekeeper. Ben Evans asks whether firms want to become "someone else's dumb API call". The more important question for the rest of us is a different one: who does my agent work for? If the comparison is done by a platform that also sells advertising or takes commission, we may simply have swapped one form of inertia for another.
The future is closer than you think
None of this is happening overnight. Uber reported around 500 food orders and just four rides from third-party AI agents in a single week of its latest quarter. Consumer adoption of agents is still marginal, and every claim about the speed of change deserves scepticism. Yet the AI100 report notes that the length of task an agent can complete on its own is doubling roughly every seven months.
But the direction is clear, and the public sector numbers show that the effects arrive long before the technology is mature. A chatbot drafting a complaint is already enough to change the economics of a complaints department.
For leaders in private and public organisations alike, consider how much of your revenue, or your ability to manage demand, depends on people not bothering. What would happen if a meaningful share of your customers or citizens handed that bother to an AI agent? And when you respond, will you redesign your services to be genuinely better, or simply find new ways to make them harder?
This week I didn’t fret, and I didn’t forget. Multiply that by millions of people, and every organisation that relies on us doing one or the other now faces a big problem.

