Centaur or cyborg: how a small team should split work with AI

Most owners think the first AI question is which tool to buy. Spend an afternoon comparing them and the differences blur, because for the work a small team actually does, the better question is not which model sits in front of you. It is how you and the machine divide a single job. Get that division right and the same tool that felt like a toy starts to earn its place. Get it wrong and you either hand over work that needs a human, or you fiddle endlessly with work the machine could have finished.

There is a clean way to think about this, and it comes from Ethan Mollick, the academic who has done most to popularise these two ways of working with AI. He borrows two images from chess, where human-plus-machine teams beat either alone. A centaur keeps a clear line down the middle: the person does some parts, the AI does others, and the join is deliberate. A cyborg blurs the line, weaving back and forth, prompting and editing in a tight loop. Both work. They suit different tasks. The mistake is using one approach for everything.

The wrong question and the better one

Ask "which tool?" and you are shopping. Ask "which split?" and you are designing how the work gets done, which is the part that decides whether AI helps you at all.

Mollick's own warning is the reason this matters. "On some tasks AI is immensely powerful, and on others it fails completely or subtly," he writes, "and, unless you use AI a lot, you won't know which is which." A demo cannot tell you where your work sits on that uneven edge. Only your own tasks can. So the operating model has to be decided task by task, not bought in a single licence.

Centaur work: a clean handoff

Centaur work has a clear seam. You do the thinking that needs judgement, then hand a well-specified piece to the AI, then check what comes back. The machine never touches the parts that need you, and you never get dragged into the parts it can finish.

A worked example. A four-person agency writes a weekly client update. The owner decides what to say, the angle, the awkward truth to include, the one thing to ask for. That is the judgement. The draft itself, turning six bullet points into a clean, readable note in the firm's usual tone, is handed over. The owner then reads it once, fixes the half-sentence that misses, and sends. The seam is obvious: human sets direction and approves, AI drafts. Mollick describes centaurs as keeping "a strategic division of labor", switching between person and machine depending on which is stronger for that part.

Centaur work suits tasks where you can say what good looks like before anything starts, and where a wrong output is easy to spot.

Cyborg work: weaving

Some work has no clean seam. You do not know what good looks like until you see a few attempts, and each attempt changes what you ask for next. This is cyborg work, where you and the model "intertwine" your effort.

Think of pricing a piece of unusual work, or shaping a proposal for a client who does not fit your normal mould. You throw a rough idea at the model, it offers three framings, one sparks a better thought, you push back, it sharpens. Nobody handed over a finished brief, because there was no finished brief to hand. The value is in the back-and-forth. Simon Willison, a developer who writes plainly about daily AI use, suggests treating the model as "an over-confident pair programming assistant": fast, useful, sitting beside you, and wrong often enough that you stay in the loop. That is cyborg posture. You are not delegating. You are thinking out loud with a fast, fallible partner.

The split test

Here is a one-page way to sort your week. For each recurring task, ask three questions. We call it the split test, because it tells you where the line between you and the machine should fall.

Can you say what good looks like before you start? If yes, the task can be briefed and handed over. That points to centaur work: write a clear brief, let the AI draft, check the result. If you cannot specify it up front, the task wants weaving.

Do you expect to react and push back several times? If a task only gets clear by seeing attempts and reshaping them, it is cyborg work. Sit with the model and loop. Trying to brief it perfectly in one shot will waste your morning.

How costly is a confident mistake? This question overrides the other two. Where a wrong answer is expensive or hard to catch, keep a human checkpoint no matter which posture you choose. The split decides how the work flows. This decides where you refuse to take your hands off.

Decide who holds the pen before the work starts, and most of the confusion about AI goes quiet.

Run your real tasks through those three questions once, on paper. Most teams find their week sorts into a short centaur list (drafting, summarising, formatting, first-pass research) and a shorter cyborg list (shaping, deciding, pricing the unusual). The exercise takes twenty minutes and replaces a month of scattered dabbling with a deliberate way of working.

The honest limits

This is a map, not a law, and the ground moves. A task that needs close weaving today may become a clean handoff in six months as the models improve, and the reverse happens too when you take on work that is genuinely new to you. So the split is worth revisiting, not fixing once. Re-sort your list every quarter.

There is a subtler trap. The cyborg loop is absorbing, and it is easy to spend forty minutes weaving on something a clean brief would have settled in five. Watch for that. If you find yourself in a long back-and-forth on a task you could have specified up front, stop and write the brief instead. The two postures are tools. Reaching for the wrong one is its own kind of waste.

What to do this week

Take fifteen tasks your team did last week, the recurring ones. Put each through the split test and write a single letter beside it: C for centaur, Y for cyborg, or H for human-only where the judgement should not leave a person at all. You now have an operating model on one page: what to hand over cleanly, what to weave on, and what to keep. Pin it up, work to it for a fortnight, then correct it from what you learned.

That one page is worth more than another tool. It is the difference between a team that owns a few good ways of working and a team that keeps buying capability it never settles into. If turning scattered AI use into a few repeatable ways of working is the job in front of you, that is exactly what the AI Advisory for Teams work is built to do, starting from the tasks you already run rather than a vendor's feature list. You can also book a business call to talk it through.

Sources and further reading

• Ethan Mollick, "Centaurs and Cyborgs on the Jagged Frontier," One Useful Thing, 16 September 2023. Where the centaur and cyborg distinction comes from, and the warning that you won't know where AI fails until you have used it a lot. Mollick is an academic writing from his own research.
• Simon Willison, on how he uses LLMs to help write code, 2025. A developer's honest account of daily use, and the source of the over-confident-pair-programmer line.