The Vending Machine Problem
By Martin Knudsen - Chief Executive Officer
10/14/2026
A vending machine does one exchange and does it well. You make a choice and the item you chose drops into the tray. It never asks what you're actually trying to get done, and it has no opinion on whether something else would have served you better. It delivers, and it rarely gets it wrong.
Day-to-day AI use has settled into something close to that. A person asks for a draft and gets a good draft, or asks for a summary and gets a good summary. That reliability is what keeps the pattern in place.
Starting with single requests is the sensible move
Asking for one thing at a time is how anyone reasonably starts with a new tool. You don't yet know what it's good at, so you hand it something small and check what comes back: a first pass at a tricky proposal, say, or a tidy version of rough meeting notes.
The results tend to be good. They aren't perfect, but they save real effort and they earn a second try the next day, and the day after that. The people using AI this way haven't misunderstood it or fallen behind. They found something that helps and kept using it, which is the right response to a tool that works.
I'd go a step further. A lot of work genuinely suits a single, well-aimed request. Many tasks are small and happen once, and building anything more elaborate around them would be ceremony. So the habit is sound. What deserves a closer look is what the habit teaches people to expect, because that expectation ends up setting the size of everything they ask for next.
The plateau confirms itself
A short request names the thing you want back. It leaves out nearly everything that makes the work specific to your company and this week, starting with who's going to read it and what your team would actually accept.
Ask for a summary of a long complaint thread and you'll get an accurate one. It won't mention that this customer nearly left last year, or that someone on the account already promised a fix, because neither fact was in the request. The model can't know what it wasn't given, so the answer comes back competent and general. Output quality follows the context and scope that were handed over, and a one-line request hands over very little of either.
Then a quiet loop closes. The person reading that competent, general answer draws a fair conclusion: AI is handy for this kind of thing and not much more. On the evidence in front of them, they're right.
The requests stay small because the answers are fine, and the answers stay fine because the requests are small.
More usage doesn't break that loop. It runs the loop more often. Extra licenses put the same small exchange in front of more people, and better prompting sharpens each exchange without changing its size. A usage dashboard counts all of it as adoption, accurately, while the business operates much as it did before anyone had a login.
So I don't read this pattern as resistance. People who resist a tool don't use it every day. I don't read it as a training gap either, because the habit was learned correctly, from real results. Mandates won't shift a habit people formed for good reasons. The move that would change the results is handing AI something bigger than a single request, and the loop gives nobody a reason to make it.
The requests stay small because the answers are fine, and the answers stay fine because the requests are small.
Speeding up a task and rethinking a week are different moves
Picture one person in one role, someone who runs customer operations for a region, across two versions of the same week. In the first version, she uses AI one request at a time. The monthly account review still gets put together, only faster. The hard email to a supplier still gets written, from a decent first draft instead of a blank page. Her week keeps its pieces and their order, and each one takes a little less time.
In the second version, the week itself was reconsidered before anything in it was sped up. The background she used to re-explain at the top of each request, about her accounts and her team's standards, gets explained once and stays explained. Work that doesn't need her keeps moving while she's busy elsewhere, so the account review is already drafted by the time she sits down to it. The time that comes back goes to the calls only she can make, like which customer gets an exception and what to tell her team when a quarter goes badly.
She brings the same judgment to both weeks and works just as hard in each. The difference is how much of her work was open to rethinking. In the first version it was a task at a time. In the second, it was the week.
The week has to be reconsidered on purpose
The first version of her week has no fault anyone would report. Deadlines get met and the answers are good. Improvement efforts go looking for what's broken, and this week gives them nothing to find. Reconsidering that week means taking apart a routine that pays off every day, for a gain nobody can see yet. I'd bet that's why this so rarely happens by drift.
The person best placed to reconsider a team's week is usually whoever manages that team. They can see how the work fits together, and they have some say over its shape. They're also inside that week, answering to the same deadlines and rewarded daily by the same good answers. That's a hard place to argue with a routine from.
We think it takes a deliberate stop, made by someone with the standing to make it, to look at the week as a whole instead of the next task in it.
The vending machine down the hall is as reliable on Friday as it was on Monday, and the people using it are right to keep using it. It does what it's asked. AI in a working week does the same, request by request, and a single request gives it no opening to say the week was the thing worth rethinking. That part belongs to the people whose week it is.
Chief Executive Officer, Arkane Digital
Martin works with enterprise leaders to define AI operating models that align strategy with execution, helping organizations move from experimentation to scalable adoption.