A legal team gets an agent that drafts first-pass non-disclosure agreements from a short intake form. The drafts are fine. Not brilliant, but fine: correct parties, sensible terms, the company's standard clauses in the right order. A month later someone asks how much time the team is saving. The honest answer is not much. Each lawyer has been opening the draft, reading it, and then rewriting it in their own words, clause by clause, until it feels like theirs.
Nobody decided to do this, and nobody would admit to it in a meeting. Asked, they would say they were checking it. But checking a document and rewriting it are different activities, and only one of them was in the business case.
Pride is a line item nobody budgets
Every AI business case contains an estimate of time saved. Almost none contains an estimate of time spent redoing work that was already acceptable. Yet that second number can quietly swallow the first.
The behaviour is easy to recognise once you look for it. The marketer who regenerates the copy six times and then writes her own version anyway. The engineer who reads a generated function, decides he would have structured it differently, and refactors it before the tests have even run. The analyst who rebuilds a model the agent produced because he needs to understand it, which is fair, and then keeps rebuilding them long after he does.
Some of this is legitimate quality control. Some of it is people making the work theirs so they can stand behind it. And some of it is plain pride: a reluctance to put your name on something you didn't make, combined with a quiet suspicion that if a machine can make it, your making it was never worth much.
Why we discount machine-made work
Psychologists have a name for one side of this. The IKEA effect describes how people value things more when they have built them themselves, even badly. A wobbly bookshelf you assembled feels better than a solid one you didn't. Work behaves the same way. A proposal you sweated over feels more valuable than an equally good one that arrived in thirty seconds, and you will tend to judge the second one more harshly.
The harshness is also unevenly applied. When a colleague's draft contains a mistake, it is a slip: they were busy, it happens. When an agent's draft contains the same mistake, it becomes evidence about the whole category, proof that the thing can't be trusted. One error in a machine-made contract can justify months of rewriting every contract that follows, long after the cause was fixed.
Organisations have an older version of the same reflex: "not invented here", the habitual distrust of ideas and tools that came from outside the team. Agent output is the purest case of not-invented-here there has ever been. It wasn't invented here, or anywhere, by anyone you can go for a coffee with.
Then there is identity. A senior copywriter has spent fifteen years getting good at a thing. If the first draft now costs nothing, the obvious question is what those years were for. Rewriting the draft answers that question without anyone having to ask it out loud. It says: my hand is still needed here. Which may well be true. But it should be true because the draft needed it, not because the writer did.
The redo is invisible, which is the problem
Rework is hard to see from outside. The NDA still goes out on time. The campaign still launches. The output suggests the AI project is working, the timesheets suggest it isn't, and nobody connects the two. Managers conclude the tool is underwhelming. So do the people using it, because they have never experienced it without their own rewriting layered on top.
This produces an odd equilibrium. The tool is used, so it isn't killed. The savings don't appear, so it isn't expanded. Everyone is slightly disappointed and nobody can say why. The rework never shows up as a lump that someone could question. It is twenty minutes here, forty there, spread across a team and absorbed into the working day as conscientiousness.
It also breaks the feedback loop. If every draft is rewritten wholesale, nobody tells the agent, or the people configuring it, what was actually wrong. "I redid it" is not a bug report. The drafts never improve, which confirms the suspicion that they aren't good enough, which justifies the rewriting.
Separate fixing from owning
Forbidding edits would be the wrong answer. Editing is often exactly what the work needs, and a team that rubber-stamps machine output has a different and worse problem, the one known as automation bias. The aim is to tell the two kinds of editing apart.
A simple discipline helps. Before changing anything, the reviewer writes one line: what, specifically, is wrong with this? "The indemnity clause is too broad for a supplier this size" is a fix. "It doesn't sound like me" usually isn't, unless voice is the point of the document. If you can't name the defect, leave it alone. That single line also becomes the feedback that makes the next draft better.
Watch where the habit is strongest, because it is rarely evenly spread. Junior staff often accept a reasonable draft with relief. The heaviest rewriting tends to come from the most experienced people, whose hours are the most expensive and whose sense of craft is most tied up in the work. That is where the one-line rule earns its keep.
It helps, too, to agree in advance and in writing what good enough means for each kind of output. A standard NDA has a short list of things that must be right. A board paper has a longer one. When the bar is explicit, people measure the draft against the bar instead of against the version they would have written. We looked at this shift from making to checking in an earlier piece on the verifier's role. The redo problem is what happens when a team is handed the new job but keeps the old instincts.
Give pride somewhere better to go
None of this works if you treat pride as a defect. It is the same instinct that makes people good at their jobs. The trick is to point it at something the agent can't do.
For the lawyers, that might be the unusual contracts, the negotiation, the call with the other side's counsel. For the copywriter, it might be the voice guide the agent works from, and the campaigns where the idea matters more than the words. For the engineer, the architecture and the review standards. When people own the frame, they are less tempted to repaint the picture, and their craft shows up in every draft without them having to retype it.
Managers can help by changing what they praise. If a team only ever hears appreciation for work people visibly made, they will keep visibly making it. Praise the good call instead: the draft that went out untouched because it was right, the hour saved and spent on the awkward case that needed a human. Recognition costs nothing, and people follow it more closely than any policy.
Then run one honest check. For a fortnight, ask people to keep the agent's draft and the version that went out, side by side. Most teams have never looked. The differences will show you, quickly, which edits were the job and which were someone making sure they still had one.
