Written October 2026.

I have recently spent a lot of time on what happens when an organisation gives everyone AI tools. The good part is what you would hope for. People who know their own work closely start building things to make it easier, and some of those things are clever. That is a ground-up use of AI that I think is worth encouraging.

The harder part came when I got close to some of the results and helped take employee-built tools toward something other people could rely on. The distance between “it works on my screen” and “it can run for a team” was larger than I had expected.

What I first believed

I had assumed AI would close most of that gap. If a person can describe a tool and have working software in an afternoon, then the remaining work should be small: put it somewhere shared, tidy it up, done.

That belief is easy to hold because the demo is convincing. The screen renders, the buttons do something, the output looks right. A prototype that works in front of its author gives every sign of being nearly finished.

I should say plainly that the people building these tools were not careless. They were solving a real problem in their own work with the means they had, and they solved it well enough to show me. The gap I am describing is not a gap in their effort or their judgement about their own job. It is a gap in what production asks for, which is rarely visible from where they were standing.

What challenged it

Many of the proofs of concept I saw were half built in the same ways. They had been written for one user, and they could not handle several people at once. They lacked backend functions that a shared tool needs, such as storing data safely and recovering from errors. And security, authorization and data protection had mostly not been considered, because nothing in the experience of building the tool had asked the builder to consider them.

That last point is the one I keep coming back to. A tool that is only used by its author can mostly do without those things, at least until something breaks. The AI did not leave them out by mistake. It built what it was asked for, and what it was asked for was a tool that works.

A short list of what a prototype usually does not yet have:

  • Support for more than one person using it at the same time
  • Backend functions and error handling
  • Authentication and authorization: who may see what
  • Protection for the data it touches
  • An owner who will maintain it

None of that is exotic. It is simply invisible in a demo.

How my thinking changed

I stopped asking “can this be made to work?” The answer was nearly always yes. I started asking “should this be hardened, and what is the evidence?” By hardening I mean adding the security, error handling, capacity and support that a shared tool needs.

Hardening a tool costs engineering time, and that time comes from somewhere else. So a prototype has to earn it. The questions I ended up wanting answered were fairly ordinary:

Are people actually using the tool, or did they only try it? Does it help the team, in time saved, in more room for focused work, in output? And if it were taken further, would that benefit multiply, or is it a good fit for one person’s habits?

These are the questions you would ask of any investment. The difference is that AI has made it cheap to produce candidates, so there are far more of them to ask about. When a working demo takes an afternoon, the demo stops being the evidence. Use and value are.

Reusable help goes some way, and it is worth knowing where it stops. Standard onboarding steps and templates, the kind that bring a prototype up to what shared infrastructure expects, close part of the gap. They make the common cases quicker and more consistent. But in my experience the process still needs engineering time, even with that help. I would be wary of anyone who says it does not. The templates handle what is predictable, and the unpredictable parts of a tool still need a person who understands the system it is joining.

There is a second change in my thinking, and it runs in a more hopeful direction. AI has made it much cheaper to take an idea from the business and test it. The cost of building has fallen enough that I no longer treat “can we deliver this quickly and well?” as the hard question. Increasingly the answer is yes. The question that has taken its place is “is this what we should be focused on, given the organisation’s priorities and where it is heading?”

That is a harder question, because no tool answers it for you. It is a question about judgement and sequencing, and it belongs to leaders.

The principle

Here is the idea I would offer to other engineering and data leaders, in one sentence: AI lowers the cost of making a prototype, but it does far less to lower the cost of making something others can depend on, so the decision to harden belongs to a person and needs evidence.

Put differently, I now think of two separate gates. The first is cheap and should be wide open: let people build, and let them show what they have made. The second is deliberate: a tool crosses it when there is evidence of use, a reason to think the benefit would spread beyond one person, an owner, and a decision that the engineering time is better spent here than elsewhere.

Keeping the first gate open respects the people doing the building. Keeping the second one honest respects everyone who will rely on what comes out.

I offer this as what I have seen, not as a proven model. Different organisations will have different risk tolerance, different data, and different amounts of engineering capacity, and I would expect the right thresholds to differ.

This week

Pick one employee-built tool that people around you have started to depend on. Ask three things about it: who owns it, what data it touches, and what happens if ten people use it at once. If you cannot answer all three, you have found a prototype that is acting as a product, and that is worth knowing before something goes wrong.

If you are working through the same line between tools built outside engineering and production software, I would like to hear how you draw it: connect on LinkedIn.