In a randomized experiment across 515 high-growth startups, every company got the same AI credits, the same models and the same training. One group was additionally shown how other companies had reorganized the work. That group found 44 percent more use cases and generated 1.9 times the revenue. The variable was not the tool. It was whether anyone had mapped the tool onto the work.
One thing to say before you start. This sheet is my extension of that research, not a finding from it. What the study tested was being shown how other people redrew their work, so the stories in the issue are closer to the tested thing than this page is.
1. Sweep the functions before you pick one
In the experiment the extra use cases were concentrated in product development and strategy, not in the three places everyone starts. Check every function you have actually examined this quarter. The unchecked ones are your map.
- Customer support
- Product development
- Internal comms and email
- Strategy and planning
- Meeting notes
- Pricing and packaging
- Sales and pipeline
- Procurement and vendor management
- Finance and close
- Quality, testing, inspection
- HR and recruiting
- Field ops and logistics
- Legal and compliance
- Research and discovery
Support, email and meeting notes are where the work is already written down, which is exactly why everyone's answer is the same three answers. The parts nobody documented are invisible to the search, and they tend to be where the margin is.
2. Name the workflow
Pick the one that hurts most, not the one that is easiest to describe. Easy to describe usually means somebody already mapped it.
| The workflow | What it costs when it goes wrong | Who owns it end to end, by name |
|---|---|---|
If that last box is blank or holds a committee, stop. That is the finding, and it is a bigger one than anything below.
3. Map it, one row per step
| The step, as it actually runs | Who does it today | What an agent would actually do here | Who still has to be standing in it, and why |
|---|---|---|---|
Do it across functions, not only the one you own. If your map stops at your department's edge, you have rebuilt the control group.
4. What are you about to throw away?
Census Bureau data on American manufacturers found that AI adoption hurt productivity and profitability in the short run, and that among older establishments the abandonment of structured management practices accounted for roughly a third of those losses. The damage did not come from the technology. It came from dropping practices that were already working.
| A check, review or routine this change removes | What was it actually catching? | Where does that job go now? |
|---|---|---|
If a routine is being removed and nobody can say what it was catching, it was not overhead. It was a control whose value nobody had measured.
5. Who has agreed to hold through the dip?
The same research found J-curve returns, short term losses before longer term gains. Measured productivity is supposed to get worse first. The companies that become cautionary tales are usually the ones that quit at the bottom.
| Expected dip, how bad and how long | What we will measure, end to end | Who agreed to hold, by name | What would make us stop |
|---|---|---|---|
A title cannot hold a line. A person can.
6. The two questions
Is this a solution to a problem we already have, or is it a solution looking for a problem? Ask it early, out loud, while the answer can still change what gets built.
And if we built this workflow from scratch for an agent, what would it look like, and who would we want still standing in it?
I should disclose an interest. I am COO of Swa Technology, which works in this category. That is my bias, on the table. Every line on this page stands whether you ever look at what we make or not.
More from Ben Pickett at benpickett.com
© 2026 Ben Pickett · Enterprise Field Notes