When AI makes the work easier, what am I still responsible for?

Not long ago I had an AI write a couple of code changes at work, and I asked it to follow the conventions the project already used. What came back looked right. It was neat, it matched the patterns around it, and it would probably have passed a quick glance. I am a product manager who does not really know how to code, and I could not fully judge what I was looking at, so before committing anything I handed the changes to a second model and asked it to review them. Nothing dramatic happened. But sitting with that work, I had the clearest version of a question I keep coming back to: once AI makes the work easy, what part of it is still mine?

The question is new because the hard part of work used to be producing it. A report took days and a prototype took weeks, and the effort itself forced a kind of discipline: nobody spent three weeks building something without first being fairly sure it was worth building, and by the time you finished, you usually understood what you had made. AI has broken both of those links at once, and each one leaves me with a responsibility the machine cannot take.

The first comes before anything is made. When producing something costs almost nothing, the first question is no longer how to make it. It is whether it deserves to exist at all: who it is for, what problem it actually solves, what it will cost to maintain, and what it will push aside. That has always been the heart of product management, and cheap production makes it more important, not less, because it has never been easier to build something nobody needs. I caught myself with an idea for an AI assistant I was genuinely excited to build. Before writing a line of it, I had to ask what was stopping people from simply using the tools I already use every day, and whether what I had in mind was truly a product. The honest answer was that nothing fundamental stopped them. So the plan changed, and before building anything, I am now testing whether there is a version people would actually need.

The same question shows up in smaller ways every week. When a tool offers to build me a dashboard, a new workflow or another layer of automation, I have learned to ask whether it is worth having at all, because everything I make has to be maintained, and a system that needs constant care quickly becomes one more job. Effort that goes into the wrong thing is still wasted effort, however quickly it was produced.

The second responsibility comes after the work is made, and it is harder to see. I can now get a clean draft, a plan or working code faster than I can read it properly, and the danger is not that the output is bad. It is that it looks finished. AI is often confident, well formatted and convincing even when it is wrong, and while a model will sometimes challenge a bad premise or catch its own mistake, I cannot rely on it to do that every time. It can know a great deal in general and still miss the one piece of context that decides whether its answer fits my situation.

A finished looking answer is the most dangerous kind, because it is the one I am most tempted to stop reading.

So the responsibility has moved, but it has not gone anywhere. After deciding what is worth doing, it sits in one more moment: the decision to accept the work. Whatever tool helped, it does not take responsibility for that decision. The people who read my work, use my product or rely on my call need it to be true, to work, and to be something I can stand behind, and none of that becomes easier because the first draft arrived quickly.

I learned what standing behind your work means long before AI. A colleague once pointed out that my team should be more open about the work behind our results, and it made me realise that your work will speak for you, but you also need to speak for it. Show the workings, not only the answer. That has become my test for accepting anything a machine gives me. If I can't explain the decision in my own words, without the output in front of me, then I don't really own it yet, and I am not ready to act on it.

Passing that test takes a few honest questions rather than a ritual. Did I define the problem properly, or did I get a beautiful answer to the wrong question? Have I checked the parts someone will actually act on, the numbers, the names, the dates and the claims that would be expensive if they were wrong? And when it is code or a product change, has it been tested where it will really be used, since a fast first version doesn't remove the work of testing, rolling it out and maintaining it? When I can't answer those myself, as with that code, my job is to find someone or something that can, whether that is a second review, a test or a person who knows more than I do. For anything with real consequences, I make the final call myself.

The quieter risk is what happens to ownership over time. If I keep accepting work I can't explain, I may slowly lose the ability to explain it, and then the work stops being mine in any meaningful sense, even with my name on it. I have felt a small version of this with writing. For years I wrote a weekly email to a small circle of friends, and that habit kept my voice alive. When I stopped, and leaned more on tools, my writing started to feel flatter, less like my own thinking and more like something I had approved. I once read that leaning on AI constantly can change how your own mind works, and whether or not that is fully true, it matched what I was feeling. I have had to admit that I am not immune to the temptation to let the tool do the thinking, and part of why I am writing these essays at all is to make sure the thinking stays mine.

AI makes a great deal of my work easier, and I don't want to pretend otherwise. It means I spend less time producing and more time on the parts that genuinely need me: deciding what is worth doing, judging what comes back, and owning what happens next. The work got easier. The responsibility didn't move, and that part was always mine.