The future of work debate has an evidence problem

There is no shortage of numbers about how much work agents will absorb. There is a considerable shortage of numbers that came from observing agents doing work.

This gap is not an accident. Watching real adoption is slow, messy and organisation-specific, while producing an exposure estimate takes an afternoon and a task taxonomy. So the estimates arrived first, and they have been circulating ever since with a precision they never had.

What the estimates actually measure

The usual method takes an occupation, decomposes it into tasks, judges which tasks a model could plausibly perform, and reports the proportion. It is a reasonable method for what it is. The trouble is what happens in translation.

A task in a taxonomy is a clean unit. A task in a job is embedded in context that the taxonomy discarded: who you have to ask first, which exception applies this quarter, what the client said on the phone that never made it into the record. An agent that can perform the taxonomy version has not necessarily touched the real one.

So an exposure score is a statement about tasks as described. It is routinely read as a statement about jobs as performed, and those are different claims.

The part that gets left out

In the organisations we work with, the largest share of the effort is not the task the agent performs. It is everything around it.

Deciding which work is safe to hand over. Establishing who owns the agent. Defining what good output looks like precisely enough to score it. Building the review step. Noticing, three weeks later, that the agent has been quietly doing something almost right.

None of that appears in an exposure estimate, and all of it determines whether the exposure ever converts into anything. This is why adoption curves look nothing like the forecasts: the bottleneck was never model capability.

What we would measure instead

Honest measurement of agent adoption is harder and less quotable, but it is possible. It looks like this.

• Track outcomes, not capability. Not "could an agent do this" but "did one, and was the output accepted without rework".
• Count the review. If a human reads every output carefully, the work moved rather than disappeared. That is still a change, but it is a different one.
• Follow the orphans. Agents that run for months after their purpose ended are a real cost and a signal about governance.
• Measure per agent, not per organisation. Aggregate figures hide the fact that two of forty agents usually do most of the useful work.
• Report the failures. Adoption studies that only count successes describe a world nobody works in.

Why this matters beyond accuracy

Bad estimates are not merely wrong, they are expensive. A leadership team that believes sixty percent of a function is automatable will plan for that and then spend a year discovering the difference between a task and a job. A team that believes nothing is automatable will miss the narrow cases where agents genuinely remove hours from a week.

Both errors come from the same place: a number that sounded specific enough to plan around, produced by a method that was never capable of that specificity.

What we are doing about it

Yaju Labs studies how organisations actually create, use, manage and improve agents, which means watching the operational layer that exposure estimates skip. We publish what we find, including the parts that do not support a clean narrative.

It is slower work than producing another forecast. It also has the advantage of describing something that happened.

Next

The Future of Work pages cover the four areas we study and why the gap between building an agent and operating one is the real subject. If you want to take part in that work rather than read about it, the Open Development Community is free to join.