Skip to content

The TWO methodology

Task. Workflow. Organization.

AI changes tasks. Workflows need owners.

TWO is a way to redesign work around AI at three levels. Start with one real workflow.

The validation tax

AI drafts in seconds. Checking takes hours.

The people who have to trust the result still verify it. We call that gap the validation tax: the model is fast, but the review moves onto whoever stands behind the output.

Most pilots make one task faster and leave the handoffs, the review and the decisions around it unchanged. So the checking becomes the bottleneck.

T-W-O

Redesign the work at three levels.

These are levels of design, not separate boxes. Copilots can support a workflow, and agents can do individual tasks.

Task

AI automates existing work, or augments the human.

What good looks like

A verified output: numbers recomputed, sources named, errors fixed.

Ask

What does a good result look like, and who checks it?

Workflow

Agents coordinate the steps. Humans review at defined gates.

What good looks like

A repeatable process with gates that someone else can run.

Ask

At each handoff, what output arrives, what evidence makes it acceptable, and what happens if the check fails?

Organization

Owners and shared standards. Incentives make it durable.

What good looks like

Standards and skills that a team adopts, with the gates owned.

Ask

For each place AI acts, who is accountable?

OrganizationOwnership, shared standards, incentives. This is what makes the workflow last.

WorkflowSteps, handoffs and review gates. Agents coordinate; humans review at the gates.

Gate 1: ScopeWhat counts?
TaskAI does the repetition
Gate 2: PilotRead a small run first
TaskAI at volume
Gate 3: ReleaseWhat do we claim?
  • Human gate
  • Task (copilot or agent)
  • Owned by the organization
One workflow, drawn inside the organization that owns it. People decide at three gates; AI does the repetition between them.

An example from our own work

A benchmark that looked finished.

We asked an AI-built benchmark which model our team should standardize on for spreadsheet analysis.

Scatter chart titled Experiment results: accuracy versus average cost, with lines for four models across reasoning levels. It looks like a finished result.
The AI-built chart. It looked finished.

Three setup problems were hiding underneath.

  • Missing data. Up to 27% of answers said “Unable to determine”: the agent never received the file.
  • Failed runs. 9% returned nothing, and were counted as wrong answers and as $0 cost.
  • Faulty scoring. “A) 2.2m” was marked wrong when the answer key said “a”.

Our own workflow mistakes were counted against the models.

The same request, as a workflow

ScopeHuman: metric and grading rule, written once
PilotHuman: read one page before scaling
RunAgent: 4,400 runs, halts itself on errors
GradeAgent: 8,400 answers, rules first, then an AI judge
ClaimsHuman: sign off before it ships
Scatter chart titled After the workflow fix: accuracy versus average cost. Five models across reasoning levels now score mostly between 80 and 92 percent, with 95 percent intervals.
After the workflow fix: the same grader on 200 of the questions, with five models, including Claude Haiku. Bars are 95% intervals. This chart uses 200 questions and the first used all 757, so compare the bars below, not the two charts.

The result: same 200 questions, same grader

Excel workbooks 80 questions

22% before
82% after

+60 pts

CSV statistics 30 questions

30% before
70% after

+40 pts

Text only, no file needed 90 questions · the control

92% before
93% after

+1 pt

Files mattered; the control barely moved. Averages over 17 model configurations. “Before” is the original answers to the same questions, re-graded with the same grader. This is one benchmark, and 200 of its 757 tasks: an illustration, not proof.

Next steps

Understand the idea. Examine the evidence. Apply it to work.

Research

Papers and talks on tasks, workflows, governance, and organizational structure.

Explore the research