Centaur or reverse centaur: where AI belongs in your business

Dangerous ideas

I heard Cory Doctorow on the radio last Friday, ahead of his talks at the Festival of Dangerous Ideas here in Sydney, and I haven’t stopped thinking about it since. He’s the person who gave us “enshittification” to describe how online platforms slowly get worse. He was on to talk about AI, and his argument was that the bigger risk isn’t the technology itself. It’s what we’ve collectively decided to believe about it.

His new book is called The Reverse Centaur’s Guide to Life After AI, and the title’s idea is the most useful test I’ve come across for figuring out where AI belongs in a small business.

Centaurs and reverse centaurs

A centaur is a human head on a machine body. You’re in charge. The machine handles the tedious parts, or the parts that need more sustained attention than a person can manage. Think of the sensor that beeps when you indicate with something sitting in your blind spot. You’re still driving. You’ve just got a second opinion from something that never gets tired.

A reverse centaur is a machine head on a human body. The machine sets the pace and the direction, and you’re the pair of hands it uses for the bits it can’t do itself. You’re not deciding anymore. You’re checking.

The examples Doctorow reaches for are the ones you’d expect. Warehouse workers with shelves shuttling in front of them at a rate that wrecks their bodies. Delivery drivers on an algorithmic route with no allowance made for a toilet stop. Developers whose job has turned into reviewing an enormous volume of code they didn’t write. It’s the same technology in each case, just arranged differently, which is why he keeps insisting that asking what a technology does isn’t enough. You have to ask who it does it for and who it does it to.

The copyeditors

I saw a version of this about twenty years ago, long before AI was part of the picture. I worked at Nature Publishing Group at the time, and the business decided to stop employing copyeditors for the suite of medical journals I worked on. That left two options. Either copyediting wasn’t actually important to the quality of what we published, or someone else was going to have to do it.

We did it. The production department absorbed the work on top of our own, because none of us were willing to put out something worse than what we’d put out the month before.

Nobody asked us to. There was no announcement that the work had moved. A decision was made once, well above us, and then it quietly landed on the people who cared enough to catch it.

That’s the arrangement, and notice there’s no technology in it at all. AI didn’t invent this pattern. It just made it cheaper to justify and quicker to roll out.

The test

Most AI advice for small business is about tools. Which one to use, what prompt to copy and paste, which workflow to apply. The centaur test is about structure instead, and structure is what decides whether you get value or just get busy.

Think back over your past week. For every task you handed to AI, ask who set the pace and who decided what finished looks like. If the answer to both is you, and the machine did the grinding, you’re a centaur. If the machine is producing output at a rate you’re scrambling to keep up with, and your role has narrowed to sitting behind it checking, you’ve become the horse.

Nobody makes that switch deliberately. It happens because the output is fast and mostly ok, and mostly ok is a comfortable place to stop paying attention.

The vigilance tax

This is the part that costs people money and almost nobody prices in.

Doctorow makes the point through radiology. AI that flags a possible mass and asks the radiologist to take a second look is a centaur setup: the radiologist is still reading the scan, but now with a backstop that doesn’t get tired late in a shift. Outcomes improve. It also means each radiologist gets through fewer scans a day and costs the hospital more, not less, on top of what the software costs. That version isn’t the one being sold. The pitch to hospitals is to keep one radiologist and have them sign off on the machine’s output at machine speed.

That doesn’t work, and we’ve known it since 1983, when Lisanne Bainbridge published five pages in Automatica called “Ironies of Automation”. She was writing about nuclear plants and industrial process control, and the argument holds up uncomfortably well. When you automate the routine parts of a job and leave a person to monitor the rest, you have handed them a task humans are bad at, because watching for a failure that hardly ever happens is not something anyone can sustain.

The research since has a name for it—automation complacency—and the finding that matters most for a small business is that it shows up in experts and beginners alike and can’t be trained out. The mechanism is attention, not knowledge, so knowing better doesn’t protect you.

Reviewing work is harder than doing it, and reviewing work that’s usually correct is harder again, because the errors are rare, unpredictable, and by definition the ones that looked ok on the way past.

I see this constantly in the kind of work I do. In a five hundred row redirect map telling a new website where to send clicks to links that have changed has four rows with incorrect links, you don’t find out until the traffic those links used to get disappears. Or it could be styles that look great in the builder, but fall over on mobile at one breakpoint. A privacy policy that reads beautifully and cites the wrong jurisdiction. Or it could be a configuration of custom fields that work perfectly except for fields that have no content. This sort of stuff can easily fail without anyone noticing straight away, and it’s knowledge of the potential for the failure that lets me know where I need my fallbacks.

So the rule I use is that if verifying the output takes more attention than producing it would have, I haven’t bought a tool, I’ve bought a liability with a subscription attached.

That sorts the work fairly cleanly. AI earns its keep on tasks where checking is fast and failure is loud, so first drafts you were going to rewrite anyway, boilerplate, format conversions you can spot check, anything where being wrong is immediately obvious. It earns nothing on tasks where checking is slow and failure is quiet, and those are the ones people hand over first, because they’re the tedious ones.

This is why I keep saying AI belongs in your systems rather than your content. Point it at a process, and you’re still the one deciding what finished looks like. Point it at writing, and it’s very easy to end up as the quality control step for something you didn’t shape, usually spending longer fixing it than you’d have spent writing it.

You still sign it

Dan Davies has a useful term for what a lot of AI deployment is really doing. In The Unaccountability Machine he calls it an accountability sink, meaning an arrangement where a decision gets made but responsibility for it evaporates somewhere in the machinery, so there’s nobody left to complain to.

Plenty of large organizations find that useful. A small business doesn’t have that luxury.

If I ship code an AI wrote, and it takes down a client’s booking form, the client does not care where the error came from. Say this out loud to your clients too, because a fair few of them are hoping AI moves the risk somewhere else. It moves the work around, not the accountability.

AI multiplies what you already have

The strongest case for AI is the one that gets talked about least, because it doesn’t justify anyone’s valuation. AI makes people who are already good at something better at it.

That’s the radiology finding, and it holds up in ordinary work. An expert is someone who can look at plausible output and feel the wrongness before they can explain it, and that instinct comes from years of being wrong in specific ways and remembering how it felt. Give the same tool to someone/something who knows the domain, and you get faster work with fewer misses. Give it to someone/something who doesn’t, and you get confident output they have no way to assess, at volume.

This is also Bainbridge’s best point, and the one I see people skipping. The skills that decay while the automation runs smoothly are the skills you need the moment it fails. The judgment that makes you useful was built by doing the work, and if you stop doing the work it goes, slowly enough that you won’t notice until you need it.

There’s a version of this that catches people before they’ve built anything at all. Ask an AI to build a case for your idea, and it will build one, including for a bad idea, because building a case is what you asked for. The AI isn’t going to tell you the premise is broken unless you ask it to.

People have always talked themselves into bad ideas, but there’s something new in the way they arrive now. A bad idea turns up fully dressed, with competitor analysis, market sizing, and a phased rollout, internally consistent and never once tested against a person who would actually pay for it. That doesn’t feel like overhyped enthusiasm anymore. It feels like real research, which is a much harder thing to talk someone out of.

I’ve written up the checks I run before building anything an AI helped me decide to build, including when an outside opinion is worth paying for. The short version is that plausible and validated are different things, and the gap between them can be both expensive and reputationally damaging.

Follow the money

The pressure toward the reverse centaur version is relentless, and it isn’t accidental. Here’s where it comes from.

Doctorow’s argument is that the audience for AI hype was never users; it was investors. He points at the investment bank projections putting AI somewhere north of sixteen trillion dollars in value, and makes the obvious observation that a number that size only works if the technology replaces a massive amount of well-paid human labor. Every story about it therefore has to be told as inevitable and total, and even the doom scenarios help, because they make the thing sound powerful.

The centaur version, where skilled people get better and slightly slower, and costs go up, doesn’t support that number at all, so nobody is selling it to you.

That leaves two points to hold onto here. First, spending a lot of money is not, by itself, proof of fraud. Many of these tools are useful, and I use them every day. If the bubble deflates the models don’t vanish, and Doctorow’s better question is what’s left in the wreckage worth keeping.

The second is that the unit economics really are different from the last tech bubble, which matters for anyone building on top of this. When the dotcom money burned it left behind fiber in the ground, cheap talent, and code that kept working, and the underlying businesses had the right structure, where each additional user cost almost nothing to serve. AI has the opposite structure. Every additional user costs money to serve every time they use it, and the heavy users cost the most. Charging enough to cover that means charging more than most of the market will pay. The per-unit cost will come down eventually, but whether it comes down faster than the money runs out is the part nobody can answer yet.

None of that means we should avoid AI. It means don’t build anything load-bearing on top of an AI vendor whose pricing is currently subsidized by investors who want their money back.

The same logic applies to search. Google has degraded into something that serves itself first, and AI answers are absorbing clicks that used to land on your site, so if you built your visibility on ranking tricks then that ground is moving. What survives is depth and ownership. Website pages substantial enough to show you know what you’re talking about, an email list nobody can take away from you, and a site you control. That was always the more resilient option. It’s just that the alternatives have stopped working well enough to be tempting. For more on what’s happened to search in the age of AI, refer to Sarah Moon’s Post-Keyword SEO.

Which one are you

I’d borrow Doctorow’s own position on this. He uses AI daily and has no dogmatic objection to it, he just refuses to accept the arrangement being marketed alongside it.

So pick something you handed over this week and run it against my copyeditors example. Did anyone decide the work had moved, or did it just land on you? Are you setting the pace, or keeping up with the pace set by something else? If it’s the latter, a better prompt isn’t going to fix it.

Sources

• Cory Doctorow, The Reverse Centaur’s Guide to Life After AI: How to Think About Artificial Intelligence Before It’s Too Late (MCD, 2026). He publishes most of his arguments at pluralistic.net before they reach print, so if you want the free version: the centaur and reverse centaur distinction, why reviewing AI output is harder than producing it, and the radiology economics and the bubble.
• Lisanne Bainbridge, “Ironies of Automation”, Automatica 19(6), 1983. Five pages about industrial process control, and still the best thing anyone has written about what happens when you leave a human to supervise a machine.
• Raja Parasuraman and Dietrich Manzey, “Complacency and Bias in Human Use of Automation: An Attentional Integration”, Human Factors 52(3), 2010. The source for the finding that automation complacency affects experts as much as beginners and can’t be trained away.
• Dan Davies, The Unaccountability Machine: Why Big Systems Make Terrible Decisions (Profile, 2024), on accountability sinks.
• Sarah Moon, Post-Keyword SEO, https://sarahmoon.net. A great intro on how SEO works with AI—in short, your authority matters more than any keyword gimmicks.