Walk into any gym and you’ll see two people at the bench press. One is lying on the bench, ready to lift. The other stands behind it: the spotter. A good spotter keeps you safe and steps in only if you fail.
Now replace the spotter with a forklift that lifts the bar for you. The weight moves either way. Only one of them builds muscle.
Ted Chiang made the same point in The New Yorker in 2024, when he likened using ChatGPT for coursework to “bringing a forklift into the weight room.”
I see a version of this in my own work. Advising one client on capability investment, I found my contact quietly competing with a peer who held a separate AI budget. When we finally got everyone in the same room, it turned out they were all facing the same choice: spotter or forklift. As I put it to them, buying a gym membership doesn’t make you an athlete.
Many of the AI products pitched to organisations today are forklifts dressed up as spotters.
Forklifts aren’t the enemy
Some work should be forklifted: reconciling invoices, summarising meetings, drafting the first pass of a routine process or policy. If nobody needs to get stronger at a task, lift it and move on.
The mistake is buying a forklift and expecting muscle. When a vendor promises “instant capability” or “faster ramp-up,” ask which one you’re paying for.
What a spotter looks like
The research points the same way. Linking back to my previous blog on the snake eating itself… In the maths study, the AI tutor that gave hints instead of answers left students as capable as those who never had AI. The version that handed out answers left them worse off.
In the customer-support study, the tool suggested responses that agents could accept, edit or ignore. They stayed responsible for the conversation. The AI helped newer agents gain experience faster, and the authors found evidence that it helped them learn, particularly among international agents.
I’ve seen the same thing in practice. A coaching firm I work closely with built AI into its programs with one design choice most organisations skip: they constrained it. Clients know exactly what it will and won’t do. It asks coaching questions and pushes back on weak reasoning. It doesn’t hand over the answer. Its narrow brief is the reason it works.
In practice, three tests separate a spotter from a forklift:
Who keeps the decision. A spotter leaves the judgement with the person. A forklift makes it for them.
What comes first. A spotter asks what you think before offering anything. A forklift goes straight to the output.
Whether the reasoning is visible. A spotter makes people show their working, so someone can check it and coach them. A forklift hides the working inside a polished result.
Sort what you’re already paying for
Put every AI tool and learning product you fund into one of two columns.
| Spotter | Forklift | |
|---|---|---|
| Purpose | Building judgement people will need | Removing work nobody needs to get better at |
| Who decides | The person | The tool |
| What it does first | Asks | Answers |
| How to measure it | Capability with the tool switched off | Time and cost saved |
Two things usually show up before your leadership pipeline breaks.
First, tools bought as development that are really forklifts.
Second, roles where a forklift now does all the lifting that used to build judgement & critical thought, and nothing has replaced it.
Four questions before you sign
Is this meant to build capability or remove work, and has the vendor been honest about which?
Who makes the final call when the tool is in use?
Does it ask before it answers?
If it’s a forklift, where will people now get the practice it took away?
Where Plexity fits
Plexity works where business acumen and people performance meet. I sit as an independent partner, with no product to sell. I help executives decide which investments should spot and which should lift, hold vendors to that line, and design development that endures and prepares people for the future of work.
Coming next: What to measure if you want to know whether it built capability.
References
Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. https://doi.org/10.1073/pnas.2422633122
Brynjolfsson, E., Li, D., & Raymond, L. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044
Chiang, T. (2024, August 31). Why A.I. isn’t going to make art. The New Yorker. https://www.newyorker.com/culture/the-weekend-essay/why-ai-isnt-going-to-make-art