Someone still digs

TL;DR

Every technology wave eats the grunt work below it and precipitates new grunt work around maintaining the automation. AI eats the translation layer, the L1 and L2 work the IT services industry was built on. What precipitates is assurance work, context plumbing, evals, model migration, exceptions, memory hygiene, audit. The machine does the work. The grunt work becomes keeping the machine worth trusting. At the bottom of the mine sit the two things no machine takes, deciding what the firm wants and holding the pager. The services industry converges on an insurance company with engineers, and the firms that last will know exactly where their mine is.

In Pierce Brown's Red Rising, a boy named Darrow spends his days a kilometre beneath the surface of Mars, drilling for the helium that will terraform the planet for future generations. That is the story he has been told. The truth is above his head. Mars was terraformed generations ago. There are cities up there, gardens, whole oceans of people living easy lives. The civilisation that built starships and carved its rulers into gods still needs someone in the mine. It just prefers that the someone never finds out how nice things are upstairs.

Science fiction keeps arriving at this. E.M. Forster wrote The Machine Stops in 1909. Humanity lives underground, tended by a perfect Machine, until the Mending Apparatus that maintains the Machine quietly fails and the species dies of deferred maintenance. Asimov's empire rots the same way, its reactors tended by technicians who perform their grandfathers' rituals without their grandfathers' understanding. Douglas Adams played it for laughs. Golgafrincham exiled its telephone sanitisers as useless, and the civilisation that stayed behind was wiped out by a disease contracted from a dirty telephone. Different centuries, different registers, one instinct. Writers who imagine total automation keep discovering they cannot make it total. Every shining future stands on a maintenance stratum, and the shine depends on not looking at it.

The lie SaaS told

Enterprise software ran the same plot with less drama. Every SaaS product was sold as self-serve, so intuitive your team would simply use it. Every enterprise deployment ended with a systems integrator, an annual maintenance contract, and a ticket queue sorted into L1, L2 and L3. This was not a failure of the software. A generic product meeting a specific organisation always leaves a residue of integration, configuration, exceptions and upgrades. The residue is grunt work, and it is as structural as the helium under Lykos.

Enterprises outsource this residue for three reasons that have never changed. It is not their business. It is lumpy, since a company needs coverage around the clock but not utilisation around the clock, and the work pools better across an IT firm's client base. And, least admitted, outsourcing transfers blame. An AMC is partly an insurance policy. When the system breaks at 3am, someone contractually not-you is holding the pager.

So the firm splits along a clean line. It keeps the wanting and the answering, the decisions about what it is trying to do and the accountability for how it goes. It rents out the translation, the turning of wants into system states. The business stays on the communication layer. The grunt work goes to the vendor. This split has survived every technology wave since the mainframe, which suggests it is not about technology at all. It is about what organisations are.

The stratum moves

Grunt work is not a fixed category. It is a stratum that migrates upward with every wave of automation. Punch-card operators gave way to assembly programmers, who gave way to sysadmins, who gave way to DevOps, who gave way to the SaaS administrator, a career nobody predicted in 1995. Each wave eats the grunt work below it and precipitates new grunt work around maintaining the automation itself. The mine never closes. It relocates.

Now the AI wave, and this time the meal is the translation layer itself. What does L1 actually do? It reads a human's request, interprets the intent, and mutates a system accordingly. Password resets, ticket triage, report pulls, configuration changes. That is precisely the task large language models are best at. AI does not nibble at the services pyramid. It eats it from the bottom, and the bottom is where the headcount and the labour arbitrage live.

The new mine

If the pattern holds, and it has held for seventy years, AI precipitates its own grunt stratum. I count six rooms in the new mine. Context plumbing, the feeding of the machine, connecting systems, keeping permissions honest, fixing the connector that broke when a vendor changed an API without telling anyone. Eval maintenance, the golden datasets and regression suites that know whether the machine is still right. Model migration, the revalidation of every workflow each time a provider retires a model, which is to say forever. Exception handling, because agents resolve ninety percent of cases and concentrate all the weirdness into the rest, a queue the RPA era already named bot babysitting. Memory hygiene, the pruning of what the organisation's AI remembers and the ruling on what it must never learn. And audit labour, since someone reads the logs, produces the filings, and runs the forensics when an agent does something strange.

The six rooms share one property. None of them is doing the work. The machine does the work. All of them are assurance, the keeping of the machine fed, current, watched and worth trusting. That is the quiet shift underneath the age of agents. In the SaaS era, enterprises outsourced execution. In the AI era, execution goes to the machine, and what gets outsourced is assurance. Grunt work migrates from doing to verifying.

The machine does the work. The grunt work is keeping the machine worth trusting.

The stratum will keep moving, because AI eats its own residue too. Agents will babysit agents. Evals will write evals. Each room shrinks over time and the digging moves on. Which raises the question the science fiction writers kept circling. Does the mine have a bottom? I believe it does, and it is made of the two things no machine takes. Wanting, because someone must decide what the enterprise is actually trying to do. And answering, because someone must absorb blame, sign the audit, and pick up the phone at 3am. Machines can hold the work. They cannot hold the pager, because a pager is not a device. It is a promise that a person can be woken.

Followed to its end, the services industry converges on a strange new shape: an insurance company with engineers. Thin on headcount, heavy on trust. We operate your machines, we carry your pager, we sign next to you. The labour-arbitrage giants are built for the old mine, and the new one does not need three hundred and fifty thousand people. It needs a few hundred very good ones and a balance sheet that can absorb liability. That transition will be the story of the IT services industry for the next decade, and most of the industry has not noticed it is already underway.

There is a warning buried in Forster and Asimov, and it is not about outsourcing too much or too little. The empire fell because its technicians performed ritual without understanding. The Machine stopped because the civilisation forgot that a Mending Apparatus existed at all. The modern version is the enterprise that automates everything, disbands its understanding, and rents assurance from people who perform the ceremonies without comprehension. An organisation can outsource its grunt work. It cannot outsource knowing where its mine is, or it wakes up one morning in the basement of a machine that has stopped.

The Golds' mistake in Red Rising was never that they kept the mines. Every civilisation keeps mines. Their mistake was lying about it, hiding the digging below a surface that pretended to be effortless. The honest version is available to any enterprise that wants it: a mine it can point to on a map, an engine it owns, diggers chosen deliberately, and pager holders paid like the civilisation depends on them, because it does. There will be grunt work in the age of AI, and grunt work in the age after that. The firms that last will be the ones that know exactly where their mine is. The rest will find out where it was when the machine stops.