I suspect a few retired mainframe programmers allowed themselves a smile on 23 February. That was the day Anthropic published a guide to using Claude Code to modernise COBOL, the language that has run banking, insurance and government back offices for more than sixty years. IBM’s shares fell 13% that day, their worst fall since 2000. Anthropic’s pitch was blunt: modernising a COBOL system “once required armies of consultants spending years mapping workflows”, and teams could now do it “in quarters instead of years”. Investors were pricing in something simple, which is that when a scarce skill becomes easy, the premium attached to it starts to go.

COBOL expertise was valuable mainly because it was hard to find. Few people learn the language now, the ones who know it are retiring, and in April 2020 New Jersey’s governor publicly appealed for volunteers who could program it when the state’s unemployment system buckled under pandemic claims. The premium a COBOL specialist could charge always owed more to how few of them there were than to how brilliant any one of them was. AI only has to become good enough to remove a premium like that, which is a much lower bar than beating your best people.

Most of us who’ve run a specialist business have charged for difficulty at some point. When I started Vertical Leap in 2001, a good part of what our clients bought from us was knowledge they couldn’t easily get elsewhere, and we were right to charge for it, even though the web spread much of it around over the following decade. So it’s worth asking plainly what your own clients are paying you for. If it’s mostly that the work is hard and you know how to do it, this story is about your business as much as IBM’s.

Picture a small practice that has built a good living preparing R&D tax relief claims. Its edge is knowing the rules, which costs qualify and how to write up a project so that it stands up to HMRC’s questions. Reading guidance, matching a client’s spending against it and drafting the technical narrative are exactly the kinds of task AI is getting good at. The practice won’t disappear, because clients still want someone accountable for the claim, but it may find it hard to charge next year what it charges today for the hours it bills.

The same pressure is building on audit teams, specialist engineering firms and the management consultancies I wrote about in September, busy selling AI to their clients while it eats into their own billable work.

The markets have been drawing the same conclusion. Three weeks before the IBM sell-off, Anthropic released a legal plugin for its Cowork desktop agent, built for contract review, NDA triage and compliance checks, and Thomson Reuters, RELX and Wolters Kluwer each fell by more than 10%. Anthropic says plainly that a qualified lawyer should check what the plugin produces. A falling share price shows what investors fear rather than proving anyone’s fees have dropped. I don’t know how quickly this will reach every sector, and some difficulty will hold out longer than the markets expect, though the direction seems clear enough to plan for.

IBM is also a useful guide to what survives. Its mainframes still sit at the heart of the banks and insurers that depend on them, and in July its finance chief put a fall in mainframe sales down to buying cycles rather than customers leaving. Hardware, an installed base and decades of trust are hard for a piece of software to reproduce.

A smaller business has its own versions of those advantages. Relationships are the oldest of them, because a client who trusts you to tell them the truth doesn’t shop around every time a cheaper tool appears. Then there’s data that nobody else holds, such as years of jobs, quotes, outcomes and prices, which sits alongside the knowledge moat of documented processes I wrote about in August and makes it stronger. Permission matters too — the accreditations, licences and regulated status that let you sign, certify or take responsibility for work in a way AI can’t. Speed helps as well, though mostly it buys time, letting the firm that moves first set the new price for a while.

The test I’d apply is simple to describe and uncomfortable to run. List what clients pay you for, mark the parts AI could standardise within a year or two, and look hard at what’s left. If what’s left is thin, the time to move is while today’s margin still pays for the change. That means using AI on your own difficult work and passing some of the saving on before a competitor does it for you. It also means putting more senior time into the client relationships that keep customers for reasons other than scarcity. Over time, it means turning expertise into something the business owns, whether that’s data, a tool or a method, so that it no longer lives only in your best people’s heads.

Difficulty was a perfectly good moat for as long as it lasted, and plenty of us built careers and companies on it. AI is turning scarce expertise into a weaker reason to pay, and the businesses that come through well will be the ones customers would still choose if their work became easy tomorrow.