Don’t let the robot mark its own homework

It has brought a lot of “use your powers for good” energy into the workplace. 

But, like every superhero story, the important question is not just what power you have. It is who is holding it, how they use it, and whether anyone sensible is checking where the cape is pointed.

AI can help thinking. It should not replace ownership.

Lately the talk around our dinner table has highlighted a few issues creeping into the workflow. 

It goes like this, AI drafts the thing, and if you are using it well there are a couple of iterations here between you and your assistant. You shape the brief, test the idea, push it, pull it apart, ask for alternatives, tighten the logic, and get it to a point where it is useful.
 
Then we send it to someone to review. And they ask AI to summarise it, check it, rewrite it, or help them respond, so that response comes back polished, tidy and professional. Great we are powering. Everyone feels like the work has moved forward.

But has anyone actually read it?

Not skimmed it. Not asked another tool to digest it. Not approved it because it looked neat and had headings in the right places. Actually, read it. Questioned the thinking, asked whether it solves the right problem. Looked for what is missing. Checked whether the recommendation makes sense in the business context. Thought about the audience, the risk, the timing, the consequence.

This is where I get that really nervous feeling in my stomach.  It’s starting to feel like Skynet. Not because AI is evil. Frankly, people, the robots deserve better criticism than that.

The real issue is that we are starting to build workflows where AI creates the thing, AI reviews the thing, AI summarises the thing, and the human becomes a polite forwarding mechanism with a calendar invite.

That is not progress, that is the robot marking its own homework.

The brief is where the thinking starts

AI is at its best when it helps us think. It can cement an idea, finish the thought we were reaching for, give structure to something messy, or act like a thinking partner when we know the destination but need help clearing the path.

But the value does not come from the tool alone. It comes from the quality of the brief and the quality of the human judgement around it. A good assistant needs a good brief. That means context, audience, purpose, constraints, examples, tone, what success looks like, what to avoid and what needs extra care. Without that, AI can still produce something that sounds confident. It may be clear, polished and completely wrong.

This is one of the traps. Because AI can produce an answer quickly, it can make us feel like the thinking has happened. But often the real thinking should have happened before the prompt was ever written. We saw a version of this recently.  

“A well-written wrong answer is still a wrong answer. It just has better shoes.”

A business had a real commercial problem. They needed to increase sales for a particular offer, so they asked AI for a plan. Reasonable, we would do that too as a starting point. But the prompt was thin. It did not include enough context about the audience, the buying cycle, the existing relationship, the offer, the level of trust already built, or the risk of over-communicating.

“AI did what AI does. It answered the question it was given.”

The result looked active and energetic. Lots of touchpoints, lots of momentum, lots of “we’re doing something” energy. But in that business context, it would have been the wrong move. It risked putting too much pressure on the audience, damaging trust, and treating the list like an extraction point rather than a relationship.

That is the danger. AI had not failed. The failure was in the brief. The bigger failure would have been accepting the answer because it looked strategic.

AI has changed the bottleneck

There is another version of this happening in product and innovation teams.

In one tech business close to home, the old bottleneck for innovation used to be resources. Someone would have an idea, then a team would pull together a document or business case laying out the opportunity, the commercial logic, the technical requirements, the risks and the resourcing needed to make it real. Then people would read it, review it, debate it and decide whether the idea lived or died.

It was slow, but the slowness did something useful. It forced prioritisation.

Only one or two ideas could realistically move forward because the business could not afford to spin up every possible concept. The process was clunky, but it created a natural filter.

Now AI can help create the document and support a proof of concept in a day. A team can test whether an idea might work before they throw weeks of effort at it. A concept can fail quickly and quietly, which is often exactly what should happen. No big drama, just a useful little “nope” before everyone books three months of meetings.

And sometimes the proof of concept shows the idea has real legs. But now the business has a new problem. What happens when instead of one or two ideas making it through, ten or fifteen suddenly look possible?

The bottleneck has moved.

It is no longer just, “Can we make the case?” or “Can we build the first version?” It’s now can the business absorb this many good ideas? Can leadership make decisions fast enough? Can the team resource the next stage? Can operations support it? Can customers handle the change? Can the organisation tell the difference between a promising proof of concept and something genuinely ready to scale?

AI has not removed the need for judgement. It has moved the pressure closer to judgement. Faster creation does not automatically mean better decision-making. Sometimes it just means you have created a bigger pile of decisions.
 
That is the shift we need to understand. The best AI workflows do not remove human judgement. They make room for better human judgement by speeding up the first draft, the first structure or the first proof of concept. But only if we use that time to think harder, not just make more stuff. Otherwise, congratulations, we are faster, we are not necessarily better.

Review is not tidy-up

This is the part I think we need to get much clearer about. A review is not just making something sound better. It is not just fixing grammar, smoothing the tone, cutting a few words, or making the headings punchier. A proper review asks harder questions.

Does this solve the right problem? Does the recommendation make sense in context? What assumptions are hiding inside the answer? What has been missed because the brief was too narrow? Is this useful, or is it just well presented? Would this work with the actual audience, not the imaginary audience sitting quietly in the strategy document behaving beautifully?

That is what needs to happen and that is very human work.

Humans understand consequence, we understand the politics in the room, the history with the client, the tone that will produce a negative result, the promise we should not make, the customer who is already tired, the team that cannot absorb another new process, the learner who needs clarity not more information.

AI can assist with review. Of course it can.  It can challenge, summarise, compare, critique and help spot gaps. I use it for that too. But there is a big difference between using AI as part of a review and using AI instead of a review. One strengthens human judgement, The other hides the absence of it.

AI should raise the standard, not lower the bar

The danger is not AI, it is lazy process. I do not think AI has stopped people thinking, but I do think it has made it easier for people to avoid thinking if the workflow lets them.

If the process is generate, polish, approve, ship, then yes, thinking gets squeezed out. The human becomes the person moving the AI-generated thing from one box to another while quietly hoping someone, somewhere, has actually thought about it. That is not a great plan. It is admin with a cape.

But when the process starts with a proper brief, uses AI to generate options, interrogates the assumptions, tests the work against the business problem, and ends with a decision someone is prepared to own, AI becomes powerful.

Not the whole brain of the operation. A powerful part of it.

The people using AI well are not thinking less. They are often thinking harder. They are asking better questions, setting clearer constraints, testing more possibilities and using the tool to stretch their thinking. The people using it badly are outsourcing the hard bit.

The promise of AI is not that we can pump out more average work at twice the speed. Nobody asked for that. The promise is that we can get to the first version faster, so we have more time for the part that matters: thinking, testing, questioning, improving and making sure the work is useful, appropriate and connected to the real problem.

So yes, use AI. Use it properly, use it often, to explore, draft, summarise, prototype, compare, challenge and accelerate. Let it help.

But do not confuse AI review with human judgement. Do not mistake polish for quality, speed for strategy, a proof of concept for readiness, or a confident answer for the right answer.

AI is a brilliant assistant. But it is still an assistant, and somebody needs to be the human brain in the room.