From Co-Pilot to Colleague: The Operating-Model Shift No One Is Planning For
The work moves faster along a route that was drawn for a slower animal.
Executive Summary
For a little over a year, organisations have adopted capable conversational machines the only way large organisations know how: as co-pilots, licensed by the seat and bolted onto existing roles. The productivity gains are real and the metrics are flattering. Yet a pattern is emerging that should trouble us — functions that are measurably faster at their tasks turn out to be no cheaper, no less congested, and no different in shape than they were before. This essay argues that the co-pilot metaphor, useful as a first move, quietly holds the operating model constant and pours all of the change into a single seat: it optimises the task while leaving the division of labour, the flow of work, and the location of accountability exactly where they were. The harder and far less discussed shift is from co-pilot to colleague — giving the machine a defined role and deliberately redrawing the human roles around it. That is an operating-model change, not a technology one, which is precisely why almost no one owns it and almost no one is planning for it.
The Slide That Declared Victory
The quarterly business review has reached the slide everyone was waiting for. The co-pilot programme — a little over a year from first cautious pilot to enterprise-wide licence — is presented as a triumph, and on the evidence of the slide it is one. Adoption sits above eighty per cent. Average handling time is down twenty-two per cent. The draft that once took a service agent forty minutes to compose now takes twelve. The room nods. Someone asks that the rollout be extended to two more functions before the summer. The assistant has done exactly what was promised: it has made people faster.
And yet if you had walked the floor of that function the week before the assistant arrived, and walked it again the week of the review, you would struggle to say what had actually changed about how the work is organised. The same roles report to the same managers. The same enquiries flow through the same three tiers. The same queue forms at the same escalation point, for the same reasons it always has. The people are quicker. The machine they sit inside is untouched. We have made the crew faster without once asking whether the aircraft, the route, or the roster still make sense.
That distance — between a real productivity gain and an operating model that has not moved an inch — is the most important and least discussed feature of this moment. In the fifteen months since a general-purpose conversational model became something any employee could open in a browser, we have become very good at buying assistance and strikingly incurious about what we are assisting. The essay that follows is about the shift we are not planning for: the one where the machine stops being a co-pilot and starts being a colleague, and where the unit that has to change is not the task but the organisation itself.
Why the Co-Pilot Was the Right First Word
“Co-pilot” was a well-chosen metaphor, and its success is worth taking seriously rather than sneering at. It promised augmentation without disruption. The human stays in command; the assistant sits in the second seat — helpful, subordinate, and above all deniable. Nothing about the metaphor asks anyone to give anything up. It also fits, almost perfectly, the way large organisations actually adopt technology. A per-seat licence lands inside an existing software budget. It requires no reorganisation, triggers no restructuring paper, and threatens no one’s job on the afternoon it is signed. It demonstrates beautifully: a manager watches a blank page fill with competent prose in ten seconds and understands the value instantly, without a business case.
For a first move, this was not naïve. When a capability is new and unproven, bolting it onto existing roles is the sensible way to learn what it can do without betting the operating model on it. The mistake is not starting there. The mistake is mistaking the starting position for the destination, and then, a year later, being surprised that the destination looks so much like where we began.
Because the metaphor is load-bearing in a way we rarely examine. A co-pilot assists a pilot who is flying the same plane, on the same route, under the same roster. The co-pilot framing quietly holds the entire operating model constant and pours all of the innovation into a single seat. It optimises the human without touching the system the human is trapped inside.
Task, Role, and the Thing We Keep Not Changing
There are three quite different things an organisation can change, and we habitually confuse them. There is the task — the discrete unit of work, the draft written, the ticket answered, the code produced. There is the role — the bundle of tasks, authorities, and accountabilities we hand to a person and put on an organisation chart. And there is the operating model — the division of labour across roles, the flow of work between them, the metrics that judge them, and the place where accountability finally comes to rest.
The co-pilot operates almost entirely at the level of the task. It makes a given task faster or easier for the person already doing it. But the roles those tasks sit inside were designed, often a decade or more ago, for unaided humans working at human speed with human constraints. Speed up the task and leave the role and the operating model untouched, and you have paved a cow path. The work moves faster along a route that was drawn for a slower animal.
The unit of transformation was never the task. It is the operating model — the division of labour, the flow of work, and the place where accountability finally rests. A tool that changes only the task changes only the speed of the cow path.
Consider a composite drawn from a pattern that recurs across service organisations this year. A customer-operations function of a hundred and twenty agents, arranged in the familiar three tiers, deploys a drafting assistant to its front line. The numbers are real and good: average handling time falls by roughly a fifth, adoption is high, agents genuinely prefer the work. Twelve months on, the leadership team looks at the function as a whole and finds something uncomfortable. Cost-to-serve is flat. Headcount is unchanged. The queue at the tier-two escalation point is exactly as long as it was. The twenty-two per cent that looked so decisive on one slide has quietly dissolved into the system without moving any number that the business actually cares about.
I have watched this happen more than once, and the mechanism is always the same. The binding constraint on that function was never how fast an agent could type a reply. It was the architecture: work partitioned into tiers by difficulty, cases handed across boundaries with loss at every handoff, a quality process built on sampling because no human could check everything. Make each agent faster and you relieve none of that. You have optimised the one thing that was not the bottleneck, and the operating model absorbs the gain the way a sponge absorbs a cup of water. Nothing spills over into results because nothing about the structure was asked to change.
The Forces That Hold Us in the Second Seat
If the co-pilot frame leaves so much value stranded, why is it so sticky? Not because leaders are foolish. It is sticky because a set of entirely rational local forces all point the same way:
- Procurement and measurement logic. A co-pilot is a licence you can buy this quarter and a time-saving you can measure by Friday. Operating-model change is a programme you must justify, staff, govern, and defend for a year before it returns anything. Faced with a fast, legible number and a slow, contested one, organisations reliably chase the number they can see.
- The comfort of accountability. When the assistant merely drafts and the human signs, the human remains unambiguously accountable for the outcome. That is genuinely reassuring, and it is why the co-pilot frame feels safe. But it is also why it is inert: the whole apparatus of responsibility stays exactly where it was, so nothing downstream is allowed to change either.
- The politics of the role. A role is not only a box on a chart; it is an identity, a career, a claim on status and pay. To redraw role boundaries is to touch all of that at once. It is far easier to give everyone a faster tool inside their existing role than to tell a room of experienced people that the shape of their work is about to be redrawn.
- The absent owner. A drafting assistant has an obvious owner — the function that buys it. The operating model has almost no one. It sits across the seams between operations, technology, finance, and human resources, owned in full by none of them. Things that no single executive owns are things that no single executive changes.
- The seduction of the pilot. A pilot that speeds up a task succeeds on its own terms and then becomes the template for scaling. We scale the thing that was easy to prove rather than the thing that was worth proving, and we call the accumulation of easy wins a strategy.
None of these forces is irrational. That is precisely what makes them dangerous. Each decision is locally sensible, and the sum of locally sensible decisions is an organisation that has spent a great deal of money to run its unchanged operating model slightly faster.
What “Colleague” Actually Means
It would be easy to misread the alternative as a demand for autonomy — hand the work to the machine and stand back. That is not the argument, and the autonomy framing is itself a trap, as I will come to. To treat the machine as a colleague is something more prosaic and more demanding than letting it run free. It is to give it a defined role in the operating model: named outputs it owns, defined inputs it consumes, an explicit place in the flow of work, quality expectations it is held to — and, crucially, a redesigned human role built around it rather than merely beside it.
Return to the same customer-operations function, and imagine it had asked the harder question first. Instead of dropping an assistant into the existing tiers, it treats first-draft resolution as a role the machine holds: routine enquiries are resolved by the system, end to end, and passed not to a queue but to a human whose job has been explicitly redefined as editor of record — accountable for what goes out, but freed from producing the first version of everything. With the front of the funnel restructured, one of the three tiers is no longer needed in its old form. The most experienced agents stop working tickets at all and move to designing the exceptions — the cases the system should escalate, the patterns that signal a policy problem rather than a customer one. The headline metric shifts from handling time, which measures the speed of the cow path, to first-contact resolution, which measures whether the path still needs to exist.
| Dimension | The Co-Pilot Frame | The Colleague Frame |
|---|---|---|
| Unit of change | The task | The operating model |
| The human’s job | The same job, done faster | A redrawn job, built around the machine’s role |
| The metric that moves | Time per task | Structure of the work: tiers, handoffs, resolution |
| Where accountability sits | Exactly where it always did | Deliberately relocated to a named human owner |
| What you must reorganise | Nothing | Roles, flows, spans of control, and reporting lines |
This is a better outcome, and it is also a much harder one. It does not fall out of a licence agreement. It requires someone to decide who answers for a wrong answer once a human no longer writes every reply. It requires collapsing a tier, which means telling people their role has changed. It requires finance, operations, and human resources to move together rather than in sequence. Every one of those is slow, political, and owned by no one in particular — which is exactly why the co-pilot frame, demanding none of it, wins by default.
The Strongest Objection, Taken Seriously
There is a serious case against everything above, and it deserves better than a caricature. It runs like this: keeping the human firmly in command and the machine firmly subordinate is not a failure of imagination — it is responsible adoption. These systems are unreliable narrators. They fabricate with total confidence. The liability for a wrong answer sent to a customer, a regulator, or a court does not disappear because a machine produced it. The autonomous-agent experiments that so excited people last year, the ones that were going to chain themselves into complete workflows unsupervised, mostly collapsed into expensive loops and quiet abandonment. Given all that, the gradualist who says “assist, don’t replace; keep the human accountable” is not being timid. They are being correct.
I agree with almost all of it — and it does not touch the argument. The gradualist case is an argument about autonomy, and I have not made an argument for autonomy. I have made an argument about the operating model. Those are different axes, and the whole confusion of this moment lies in treating them as one.
You can redesign the operating model and keep the human every bit as accountable as before — arguably more so. The editor of record in the composite is not a diminished human standing back while the machine runs; that person is more accountable than the agent-in-a-tier they replaced, not less, because accountability has been named and located rather than diffused across a queue. Redrawing the division of labour does not require trusting the machine with anything you would not have trusted it with under the co-pilot frame. It requires being honest that a tool which changes what one person can produce in an hour must, if it is any good, change how many such people you need, in what configuration, doing what. Refusing autonomy is prudent. Refusing to redesign the operating model is not prudence; it is the absence of a plan wearing prudence as a costume.
“Refusing autonomy is prudent. Refusing to redesign the work is not prudence — it is the absence of a plan wearing prudence as a costume.”
The two questions have been quietly fused, and the fusion is convenient, because so long as “should the machine act alone?” and “should the work be reorganised?” are treated as the same question, a firm and reasonable “no” to the first can be used to avoid ever asking the second.
The Shift We Are Not Planning For
The move from co-pilot to colleague will not, in the end, be a technology event. The capability arrived some time ago and is improving on a curve none of us can see the top of; that is not the constraint. The constraint is organisational, and it always has been. The shift is an operating-model event that the technology merely makes unavoidable, and it is being managed — where it is being managed at all — by procurement teams buying licences and functions running pilots, none of whom hold the authority to redraw a role, collapse a tier, or move where accountability sits.
This is why nobody is planning for it. Not because leaders cannot see the tool; the tool is the most visible thing in the enterprise. It is because the change the tool implies lives in the one place our adoption playbook never looks — the shape of the work itself — and is owned by the one person who does not yet exist on most organisation charts: whoever is accountable for what the work should be, rather than for how fast the current version of it runs.
The organisations that look back well on this period will not be the ones that bought the tools first, or even used them most. They will be the ones that treated the arrival of a capable machine as a reason to ask an older and harder question, the one the co-pilot frame lets us postpone indefinitely: what is the work, who should do it, and who answers for it? We did not buy a faster way to do the work. If we are not careful, we bought a reason never to ask what the work should be. That question was always waiting. The machine did not create it. It has only made it impossible, for very much longer, to keep pretending it isn’t there.