The first time I woke up to development work that had continued while I was asleep, I found it extraordinary. I had given the AI work to do, gone to bed and come back to completed sprints. Progress no longer depended on me being there for every step.
I had already become used to AI helping me work more quickly. This felt different. I could leave something with it and return to a result I could inspect. That changes how I think about my own capacity, and about the work a business can afford to attempt.
It also creates a new responsibility. If work can continue without me, I need to be clear about what I have asked for, what the system can change and what I will check when I return.
Conversation has a natural ceiling
Most of us started with a conversation. Ask a question, read the answer, refine the request and ask again. That is useful, but it keeps you close to the work. You are still deciding what happens after almost every response.
I now find myself directing several streams of work rather than executing every part myself. Research, development and documentation can move forward across different projects. My judgement still matters, but it is being applied differently.
I increasingly feel like an orchestrator. I do not have to write every document, assemble every piece of research or carry out each development step. I have to decide what is worth doing and give the work enough direction to make the result useful.
Time horizon changes usefulness
I estimate that I am producing roughly three to four times as much as I could before using AI this deeply. That is my personal impression, not a measured productivity study. Different tasks benefit by different amounts, and the quality of the result still needs checking.
The increase is not simply about completing a familiar task faster. I am juggling four or five significant things that I would previously have struggled to pursue at the same time. Some of the work can now continue while my attention is elsewhere.
That can be daunting. More ideas become realistic, which means there are more decisions to make about where to spend my time. But there is also a real sense of fulfilment in seeing something take shape that might otherwise have sat in a backlog.
Delegation requires a different brief
In one recent product-development example, I estimated that the research spike and sprint definition took around 40–45 minutes, followed by roughly another 45 minutes of development. A feature had moved from investigation to implementation in about 90 minutes.
I would be careful about turning that into a universal comparison with a traditional development team. Scope, testing, security and production readiness all affect what the result is worth. A working feature is not automatically a finished system.
What matters to me is that I can make an idea tangible quickly enough to examine it. To make that useful, I need to define:
- The outcome and the constraints.
- The information and systems the agent can use.
- What it can change and what requires approval.
- What it should do when it is uncertain.
- The evidence I need before accepting the work.
Those are familiar management questions. Giving the task to AI does not remove them.
Trust becomes operational
I am more comfortable delegating when mistakes are recoverable. Preparing a change I can review is different from taking an action that is difficult to reverse. The amount of freedom should reflect that difference.
For work with limited consequences, it can make sense to let the agent proceed and inspect the result. Other work needs checkpoints or approval before anything happens. I want to be able to see what changed, ask why and correct it where necessary.
Trust therefore has to attach to a particular task and workflow. A good result in one setting does not justify giving the same system unrestricted access everywhere else.
Management may become asynchronous
There is a useful example outside development. A colleague sent me an approximately eight-minute voice note about an event. I transcribed it and used AI to turn the transcript into a structured brief for the wider team. I estimate the whole process took about 15 minutes.
Previously, either I would have had to structure it manually or the colleague would have needed to write a formal account. Information that could have remained in a voice note became something other people could use.
That matters because businesses leave a lot of knowledge undocumented. If the effort required to capture it falls, documentation becomes worthwhile in situations where nobody would previously have made the time.
The failure mode is invisible drift
The risk is that an agent spends hours pursuing the wrong interpretation of the task. More time is not valuable if it produces more of the wrong work.
I need enough visibility to answer some basic questions:
- What did it do, and which sources did it use?
- Where did it become uncertain or change direction?
- Can I stop the work and reverse its changes?
- Has the task moved beyond what I authorised?
A brief, a reviewable result and a record of the changes make delegation much more useful. They also give me a way to improve the next instruction rather than simply hoping the next run goes better.
Productivity is not the biggest prize
The opportunity I find most interesting is the work that now becomes possible. There are ideas I would previously have dismissed because researching or building them required too much time and coordination.
Some will still be poor ideas. AI does not make them worth pursuing. But I can now investigate more of them before making a large commitment.
I expect my own capacity to change further as the tools improve, although I do not know by how much. The useful test is whether I am bringing worthwhile work to completion, rather than how many tasks I can start.
Learn to delegate before you automate everything
I would start with a bounded piece of work that someone can evaluate properly. Define the outcome, set the permissions, give the system the context it needs and review what happened. Widen the scope when the results justify it.
Waking up to completed work has changed what I believe I can take on. It has also made the quality of my direction more important. My time is no longer the only limit on execution, but deciding what deserves that execution is still my responsibility.
Source: OpenAI to acquire Ona