Video
Why we built the T-W-O Project
Posted: | Wei Chen | 2:24
AI makes tasks faster, but experts still review in hours what it generates in seconds. This short talk explains why AI pilots stall and why the T-W-O Project starts with workflows and owners.
Key takeaways
- AI generates in seconds while experts still review in hours: the validation tax moves the work onto the reviewer.
- Pilots speed up tasks but leave the workflow unchanged, so the real design question is what to decide up front, what to delegate, and where to review.
- In the team's own spreadsheet benchmark, three human gates (scope, pilot, claims) lifted scores by 30 points with the same questions and the same grader. The reusable asset was the workflow, not the model.
- Durable change works at three levels: task, workflow, and organization, with owners, shared standards and incentives.
AI is making individual tasks faster, but most pilots stall because the work around the task has not changed. This short talk is the reason The T-W-O Project exists.
The gap where pilots stall
AI drafts in seconds. The people who have to trust the result still check it for hours: the numbers, the charts, the claims, the sources. That gap is the validation tax. The model is fast, but the review moves onto the person who has to stand behind the output.
The talk points to the same pattern across organizations: AI use is spreading faster than it is scaling. In McKinsey's mid-2025 global survey, 23% of respondents said their organizations were scaling an AI agent system somewhere, but in no single business function did more than 10% say so. In Anthropic's survey of its own staff (132 engineers and researchers, August 2025), more than half said they could fully delegate only 0-20% of their work to Claude. Both are self-reported snapshots from 2025, and the numbers are moving.
Pilots speed up tasks. The workflow stays the same.
If only the task gets faster, the review, the handoffs and the decisions around it stay exactly as they were. The design question changes: what do we decide up front, what do we delegate, and where do we review? The answer in the talk is three lines: set the rules, delegate the repetition, own the gates.
One real example
The team asked an AI-built benchmark which model to standardize on for spreadsheet analysis. The chart looked finished and pointed to the most expensive models. Underneath were three setup problems: files the agent never received, runs that returned nothing, and a grader that marked correct answers wrong. The team's own workflow mistakes had been counted against the models.
After redesigning the work with three human gates (scope, pilot, claims) and letting the agent handle the volume, the same questions and the same grader produced scores 30 points higher, almost all of it on tasks that needed files. The text-only control barely moved. A cheap model at maximum reasoning matched one that costs up to 40 times more. The reusable asset was the workflow, not the model.
What makes a workflow last
Durable change comes down to three levels. At the task level, AI automates existing work or augments the person doing it. At the workflow level, agents coordinate the steps and humans review at defined gates. At the organization level, owners, shared standards and incentives make the new way of working hold. That is the Task, Workflow, Organization view behind the T-W-O Project.
Where to start
You do not need to reorganize tomorrow. Choose one entry workflow, train the people who operate its gates, and pilot it before you scale. AI changes tasks. Workflows need owners.
A note on the numbers: the talk was recorded in October 2026 and uses 2025 survey data. Its line that most companies are not using AI agents at all, in every business function, simplifies the McKinsey figures cited above. Use the figures in this write-up, and expect them to keep changing. The transcript below is verbatim.
Transcript
[00:00]Has AI ever made you faster? And then you spent longer checking the result?
[00:05]That gap is where most AI pilots stall.
[00:09]Across organizations, AI uses spreading, but scaling is not.
[00:13]In every business function, most companies are not using AI agents at all.
[00:17]Even at Anthropic, more than half of surveyed engineers say they can fully delegate no more than 20% of their work.
[00:24]The reason is a validation tax.
[00:27]AI generates in seconds.
[00:29]Experts still review in hours.
[00:32]AI does not remove the work.
[00:35]AI moves the work onto the reviewer.
[00:39]Pilots speed up tasks, but the workflow stays the same.
[00:43]So the design question changes.
[00:45]What do we decide up front?
[00:47]What do we delegate?
[00:48]And where do we review?
[00:50]Set the rules.
[00:51]Delegate the repetition.
[00:53]Own the gates.
[00:54]Here is one real example.
[00:56]We asked an AI-built benchmark which model our team should standardize on for spreadsheet analysis.
[01:02]The chart looked finished, and it pointed to the most expensive models.
[01:07]But three setup problems were hiding underneath.
[01:10]Files the agent never received, runs that returned nothing, and a grader that marked correct answers wrong.
[01:16]Our own workflow mistakes were counted against the models.
[01:20]So we redesigned the work with three human gates, scope, pilot, and claims, and let the agent handle the volume.
[01:28]Same questions, same grader.
[01:30]Scores jumped by 30 points, almost all of it on tasks that needed files, while the text-only control barely moved.
[01:38]And a cheap model at maximum reasoning matched one that costs up to 40 times more.
[01:44]The reusable asset was the workflow, not the model.
[01:48]What makes a workflow last comes down to three levels.
[01:52]At the task level, AI automates existing work or augments the human.
[01:56]At the workflow level, agents coordinate the steps and humans review it to find gates.
[02:01]And at the organization level, owners, shared standards, and incentives make it durable.
[02:07]You do not need to reorganize tomorrow.
[02:10]Choose one entry workflow, train the people who operate its gates, and pilot it before you scale.
[02:17]AI changes tasks. Workflows need owners.