Every workflow has a seam: the line where AI stops and the lawyer starts. This finds yours. Break the work into tasks, rate each on three checkable questions, weight by hours, and get an honest band and range for what AI can compress.
Name each task, set its share of the billed hours, and rate it on the three questions. The share is your best estimate of relative time, not a measured budget. The hour weighting is what keeps the result a fraction of time instead of a count of tasks.
| Task | Hours | Codifiable | Into the score |
|---|
Each task gets a codifiable fraction from three checkable content signals, then that fraction is weighted by the task's share of the billable hours.
retrieval× 0.40checkable× 0.30templated× 0.30What's deliberately excluded. Two things you might expect are not scored here, on purpose. The partner review gate is a quality control on whether the firm trusts the output, not evidence the underlying work is incompressible, so it belongs downstream in the calculator's refill rate. And the cost of being wrong is orthogonal to compressibility, so a filing deadline can score high even though it's malpractice-grade. Leaving both out keeps the number about the work, not about risk.
Why a range, not a number. The headline is the hour-weighted share, and the range shows what it would be if every rating was off by up to your chosen confidence bound in the same direction. That is an admission, not a hedge: this is a rubric, not a measurement. It has not been validated against a firm's actual compression outcomes. Treat the number as directional.
The band boundaries at 75 / 55 / 35 / 15 are judgment calls, not measurements. A one-point drag can relabel a workflow across a boundary. Weigh the range, not the label.
The three signals correlate. Templated work is often determinate and retrieval-heavy, so the additive sum double-counts some of what they share. The weights and the sum are a judgment, not a hidden strength. If you disagree with them, the honest response is to distrust the precise number, not to trust the label.
Scope. This scores one workflow. The calculator will happily extrapolate it to a firm's whole book. If you feed it a dollar figure, score several representative workflows first, not one.
It will tend to agree with you. If you already believe a workflow is or is not compressible, the three sliders can land almost any band. Use it to force yourself to name the tasks and the evidence, not to confirm a prior.
Feed it into the AI Profit Paradox as its addressable-work input. The calculator now shows what the seam's uncertainty does to the dollar figure.