Two pieces published this month describe the same problem from opposite ends. A manager is likely to meet it in their first month with AI, not their first year.
The problem is that AI makes work look finished before it is.
The skills end
In MIT Sloan Management Review on 14 September, three consultants describe what they call a capability mirage. Melissa Swift runs Anthrome Insight. Teryluz Andreu and Dolores Hernandez work at Axialent. Both firms sell advice on organisational culture, which is worth knowing as you read.
Their argument is simple. When anyone can produce polished work for routine tasks, the quality of the output stops telling you much about the skill of the person who produced it. Colleagues lose the ability to judge each other, trust weakens, and underlying capability can erode while the work still looks fine. AI, in their phrase, is less an improver of work than an amplifier.
One of the people they quote, Stephanie Antonian, chief executive of Aestora, puts the risk plainly. Everyone can now produce work that looks pretty good, she says, but you don’t know what’s underneath it.
The evidence behind the argument is thin. The article cites a joint study by the two authors’ firms, based on interviews with AI experts, and gives no sample size, dates or figures. Treat it as an informed argument from practitioners, not a research finding.
It is still worth taking seriously, because it names a question managers will recognise quickly: how do I know who on my team can actually do this?
The process end
The second piece has stronger evidence. In Harvard Business Review on 16 September, researchers from Erasmus University and Vrije Universiteit Amsterdam, with a colleague at RTL Nederland, summarise a year spent inside RTL’s creative department.
They observed how staff used AI tools day to day, interviewed them, and analysed 207 AI-generated images along with the prompts and decisions around them. The peer-reviewed study appeared in Academy of Management Discoveries in the spring. So the research is not new this week, but the HBR piece brings it to a management audience.
They describe two effects. The first they call process collapse. Brainstorming, sketching, visualising and producing used to pass through several stages and several people. With AI they fold into one person working with one tool. The second is early aesthetic closure: a striking AI image locks in an idea too soon, because it looks so convincing that nobody wants to reopen it.
HBR’s article opens with a case. A senior creative made an image with AI and the client loved it. The production team then had to recreate it with real cameras and real actors, and the image showed something reality could not deliver.
Why this is not about creative agencies
Replace “image” with “draft” and this is every department.
A report that used to go from an analyst to a manager to a director now reaches the director looking finished. A proposal that used to be sketched on a whiteboard by three people now arrives as a polished document written by one. The steps where someone would have said “that number looks wrong” or “we cannot deliver that by March” did not fail. They were skipped, because the work did not look as if it needed them.
And the cost lands downstream, on whoever has to deliver what was promised.
That is the connection between the two pieces. The consultants describe what polished output does to your view of people. The researchers describe what it does to your process. Both come from the same fact. The look of finished work used to be evidence that the work had been through a process. It is not any more.
What to do before the first draft
Both pieces point at decisions a manager can make before anyone opens a tool.
Decide where the checkpoints go. For any piece of work that matters, name the point at which a second person looks at it, and what they are checking for. Write it down, because the natural pressure will be to skip it when the draft looks good.
Separate the look of the work from the judgement in it. When someone brings you a polished draft, ask them to walk you through one decision in it. Why this figure, why this recommendation, what they considered and rejected. Their answer tells you what the document cannot.
Make the review a conversation, not a form. A checklist that asks “has this been reviewed?” will be ticked. A ten-minute conversation in which someone explains the work to a colleague will surface the weak points, because explaining is where people find out what they do not understand.
Keep some work done without the tool. Not as a test, and not everywhere. But if a skill matters to your team, somebody should still be using it often enough that you would notice if it faded.
Warn the client, or the boss, early. A convincing draft invites commitment. Saying “this is a direction, not a design” at the start costs one sentence and can save a rebuild later.
The honest limit
Neither piece tells you how often this happens or how much it costs. The consultants’ study has no figures, and the rework in the RTL case is described rather than measured.
What they give you is a name for something you are about to see, and a reason to set up the checks before you need them rather than after.
This is also a problem that gets harder to fix the longer it runs. A checkpoint added in the first month feels like normal process. The same checkpoint added after a missed deadline feels like a punishment, and people route around it. The cheapest time to decide where the second pair of eyes goes is before anyone has got used to working without one.
It looks pretty good. That is exactly the problem.