Behind the builder / 5 August 2026
How much work actually left my desk?
UBIKITOS created a human-versus-system workload baseline. The challenge was separating forecasts, hidden supervision, and actual evidence of work transferred.
The question after the waterpark
The previous day’s waterpark story gave the project a very human reference point. The interesting ambition was never merely to have software that could produce more software. It was to make useful work less dependent on the founder watching every move. On August fifth, that ambition became a measurement problem. A Human versus Autonomous Work Forecast and a founder-versus-system baseline were created, alongside work on consistent identifiers and supporting records. The conversation was moving from “this feels powerful” toward “what can we actually count?”
That shift matters because AI creates a persuasive impression of activity. Replies arrive, documents grow, tasks acquire updates, and the system can explain its progress in polished language. Meanwhile, the founder may still be doing the difficult part: deciding what the request meant, fixing assumptions, checking the result, and restarting work that stopped. If those contributions disappear from the accounting, autonomy looks better on paper than it feels in an ordinary day.
A forecast is not a timesheet
The word forecast needs to keep its meaning. It is a way of estimating what may happen, not a receipt for time already saved. A baseline is also a starting point, not a final verdict. Today’s work made workload transfer more measurable, but it did not establish a reliable total of hours returned to me. That would require repeated observations of actual tasks, including the human effort wrapped around them. The distinction is simple enough to state and remarkably easy to lose in a productivity story.
Consider a hypothetical research task. An assistant produces a useful draft quickly, but a person spends time checking sources, removing unsupported claims, and correcting the brief for the next run. The draft may still represent a real saving. The saving is just not equal to the entire time the original task might have taken. Counting only generation time would miss the work that remains. Counting only corrections would miss the useful contribution. A fair comparison needs both sides of the exchange.
The hidden work around the work
Some founder effort is obvious: writing a specification, reviewing an output, approving a consequential action. Other effort is less visible. Remembering which version is current takes attention. Recovering context after an interruption takes attention. Figuring out whether a task is truly finished takes attention. A system can make one part faster while making these surrounding tasks heavier. That is why the right measure cannot be a simple tally of how many responses or files appeared.
For UBIKITOS, the useful direction was toward complete workflows. What did the human need to do before the system could start? What interventions occurred while it ran? What checking was required afterward? Did the result land somewhere useful? These questions make the measurement less flattering and more practical. They also reveal where the next improvement should go. If most of the burden is in repeated clarification, a faster model may do less good than a clearer intake process.
Why identifiers entered the story
Consistent identifiers sound like a detour from a personal productivity question. They are actually part of answering it. If the same mission appears under different names in different places, it becomes difficult to tell whether a report refers to new work, a retry, or another description of the same activity. Workload accounting depends on being able to follow a task without losing its identity. Otherwise, the system can accidentally count movement as completion.
Imagine a job that is submitted once, retried after a problem, and eventually finished. If every attempt is treated as a separate successful task, the productivity report becomes fiction without anyone deliberately lying. If the failed attempts vanish, the report hides the effort required to reach the result. A consistent identity lets the story remain connected. It helps distinguish the mission from its attempts, and the final outcome from the activity that happened on the way there. That is useful operational housekeeping, not decorative bureaucracy.
The denominator has to be honest
A percentage can look wonderfully decisive while saying very little. “Mostly autonomous” depends on what has been included in the total. Does it refer to tasks, elapsed time, estimated effort, or a selected part of the workflow? Does a tiny automatic update count the same as a difficult decision? Does the human review disappear because it happened in another application? Before a number can guide the business, its denominator needs to be understandable.
This is where a founder-versus-system baseline can become genuinely useful. It provides a place to make the categories explicit and to notice when the work does not fit them cleanly. The purpose is not to force every minute into a flattering chart. It is to see where responsibility and effort actually sit. Some work should remain human-led. Some work should become easier through better tools. Some work may be unnecessary altogether. A good measurement system leaves room for all three possibilities.
Repetition would make the baseline valuable
One baseline can reveal assumptions. Repeated measurement can reveal a pattern. That was the next step: turn the comparison into an operating metric and close the next autonomous mission loop. Repetition matters because a single successful run can depend on unusually clean inputs or generous supervision. A difficult run can also be unrepresentative. The aim is to learn what happens under ordinary conditions, including the inconvenient cases that do not make an attractive demonstration.
The same principle applies to claims about time saved. A useful result would be tied to a defined kind of work and the conditions under which it was observed. It would not need to become a universal promise about replacing the founder. In fact, a narrow, repeatable improvement would be more useful than a sweeping estimate. It would identify a place where the system earns trust through experience, and a place where further investment has a clear reason rather than a general sense of technological possibility.
The ambition was still personal
There is a human reason to care about all this accounting. The point of an assistant is not to create a second job supervising the assistant. It is to make room for better decisions, creative work, and a life that is not organised around checking whether the latest task quietly stopped. The waterpark episode belongs beside the measurement work because it makes that ambition concrete without pretending that one day away proves the entire system is autonomous.
August fifth did not deliver a triumphant number of hours saved. It delivered a more serious way to ask the question. The forecast, baseline, and identity work gave the next missions something to be measured against. That is a useful kind of progress: less dramatic than declaring the founder liberated, and much more likely to lead somewhere dependable. The next task needed to move through the system with its human contributions visible. Then we could begin to see how much work had really left my desk.