The State of Work

It's August 2026. Every company has bought AI. The industry is arguing about which software survives it.

We went and looked at the work instead. 128 companies across 9 sector groups. At six of them we timed it step by step: 187 real processes. Who does it, in which tools, how long every step takes, stated by the people who actually do it.

2,546 hours a week of ordinary work.

Here's what we found.

The waste companies keep buying software to fix doesn't live in the software. It lives in the gaps between it. Half of all handoffs are a person carrying work from one screen to another. On conservative math, about a quarter of the time we mapped is recoverable. At the companies that already bought AI, that waste sits inside 34 processes people still run by hand.

Two things before you read it. Six timed companies out of 128 is small, on purpose. I'd rather time six end to end than collect a thousand survey answers. And not everything people flag is worth fixing. Automating everything was never the point. Finding the gaps between your people is.

Everything below comes from timed process documentation: real work, stated step by step by the people who do it. No survey answers, no industry averages, no extrapolation.

Every number in this report comes from one of three groups. We studied 128 companies across 9 sector groups. 13 of the 128 run our Ace Work platform end to end, a measure of platform coverage, separate from the groups that follow. For 14 of the 128, we wrote every process down step by step: the documented group. For 6, we also put a time on every step: the timed group.

Across the six companies we timed, the conservative median is 26% of documented time recoverable. The raw median, before we discount for adoption and ramp-up, is 36%. And the six split evenly, three under 25% and three between 25 and 50%.

Different industries. Different sizes. Same leak.

The top 10% of processes hold 63% of the recoverable time; the top 5% alone hold 45%. "Top" means ranked by weekly recoverable time, not by how often a process runs. So you don't fix everything. You find the right 19 processes and ignore the rest.

13.4% of the processes we timed run under an hour a week. Too small to justify a traditional automation project, so they stay manual. They only get fixed when generating a workflow takes minutes, not weeks.

99% of the steps we mapped are done by hand. 48.9% of handoffs cross from one tool to another. The average process touches 2.9 tools, and spreadsheets show up in 32% of handoffs.

When people describe what's wrong with their work, they don't complain about speed. They complain that they can't see what's happening.

Seven companies in our documented group have already bought AI. It shows up inside 34 of their processes. Those processes are still manual.

The same five kinds of work leak time across the six companies we timed. Reporting shows up at all six; the other four each show up at five.

At the median company we timed end to end, work crosses 52 different tools. The waste isn't inside any of them. It's a person carrying work between them, 15 minutes at a time.

Discounted for adoption, and again for ramp-up, the recoverable time at the average company we timed comes to:

Time figures were stated or estimated by the people who do the work, through structured process documentation and AI-led interviews. Nothing was estimated on their behalf.

Everything in this report is aggregated and de-identified process documentation: figures are published only as medians, averages, and bands across the cohort, no published figure describes a group smaller than three companies, no company is named, and no company can be inferred from any published figure or combination of figures. Sectors with fewer than three companies are merged into an Other group before anything is published.

Automation potential is never assumed to reach 100%. Conservative figures further discount for realistic adoption and ramp-up. These are projections from self-reported time, not stopwatch measurements, and they describe the companies we mapped, not the economy. Six companies is a small cohort. It is also six real ones.

Dollar figures use a flat $100 an hour and a 46-week year. That rate is a round stand-in for fully loaded cost, not a measured salary figure, and it is applied uniformly regardless of who does the work. A company whose recoverable time sits mostly with junior staff should read the figure down; one where it sits with senior staff should read it up.

Documented time is not evenly spread. A small number of continuous, always-on processes carry a disproportionate share of the total, which is why medians appear throughout this report and why the receipt is built on a share of time rather than a count of processes. Where a figure is an aggregate rather than a median, it is marked as one.

The State of Work is published in editions. The cohort grows each time, and every edition remeasures what broken work costs. Leave your email and the next one arrives the day it ships.