Is it done?

“Is it done?” was a haunting question throughout my career, and most of the time it answered itself, because the budget or the team’s patience ran out before the work did. With AI, for most people, it feels like nothing runs out.

I usually decide something is done by weighing the time I’ve spent analysing a topic against its importance in the bigger picture. It’s more of an internal reflex, and it’s prone to errors. Most of the times it took way longer than what it was worth for.

Part of the reason sits on the other side of the conversation. Work expands to fill the time available for its completion, as Northcote Parkinson wrote in 1955, and with AI it expands to fill the rounds on offer, because a model always has one more to offer. The urge starts with us: we tend to fix things by adding rather than removing, even when removing would work better, a pattern called additive bias.

Models picked it up and amplified it. GPT-4 leaned toward adding even more than people did, especially when asked to “improve” something rather than “edit” it, and training layered verbosity bias on top, because the people rating answers tended to prefer the longer one. It hasn’t faded: the most concise of 76 assistants tested this January was still GPT-3.5, a model from 2023. A model shaped this way answers “more” to whatever you ask it, including whether the work is done, so the feeling of done arrives later than it should.

Past a point, those rounds run into diminishing returns, and then into negative ones: models revising their own work without specific guidance get worse with every round.

Which means the question we ask decides whether the work ever ends. “Improve” is the one word that reliably makes a model add, and adding never finishes anything; finishing is mostly taking away.

So the weighing has to happen before the Nth prompt, while we can still see clearly how much a topic matters at this stage of things, counting the budget in rounds rather than minutes, because minutes slip by unnoticed inside a conversation and rounds don’t.

The bottomless promise of the RSI (recursive self-improvement) agentic world never runs out of yeses. “What would stop us from shipping?” does.

In a team, the same call needs an owner. When everyone works with an agent that always has one more round to offer, the person who can say “this is enough for now” is worth more than another draft.