Commercial lease abstraction

A decomposition is a model, not a measurement. Component shares are labeled by basis: every share in this report is modeled, none is observed, and the shares sum to one with nothing dropped or renormalized away. Mapping strength is recorded per component, and a public score is a floor only where pack fit is strong. Weak mappings never borrow a score: for context-heavy work a public benchmark can overestimate, so a borrowed number is not a conservative bound.

Two objects are derived and never blended. The automatable share answers how much of the work sits on strongly mapped task types; it is exposure, not success. The chain-completion bound multiplies floor-valid floors across required components and fails closed: one unmapped required component leaves it open, with no number emitted.

Sixty percent of modeled workflow minutes sit on strongly mapped task types: parsing and clause extraction. Fifteen percent sits on a moderate mapping and twenty-five percent on weak or unmapped components. The assumption load is one hundred percent: every share is a modeled prior, not an observation, and an observed shadow of the actual team usually redistributes them.

This share is what the phrase 'exposed to automation' is allowed to mean in this report. It is not a success rate, not a savings estimate, and not a headcount effect. Those need the separate economic model, with volumes, touch time, QA staffing, and error costs, none of which is in this report.

Unattended end-to-end completion is bounded by the product of component floors across required components, because sequential work fails at the weakest link. Two of five components have usable public mappings. Consistency checks, drafting, and exception judgment do not. The report therefore leaves the bound open and names the three blocking components.

If a pilot supplied measured floors for the blocked components, the product across all five would still be a modeled lower bound, at best exploratory, until an authorized run in the operator's environment exists. Actual systems also recover: retry, review, and escalation loops raise completion above a raw chain bound, and measuring that recovery is precisely what the pilot is for.

The automatable share feeds a labor-exposure model only. A savings, capacity, or headcount claim needs addressable volume, baseline touch time, QA and exception staffing, error and rework costs by severity, and an adoption assumption with a source. All of those inputs are unknown-to-collect. None of them is here, so none of them is claimed, and no vendor is named as the winner.

Per-page parser pricing among published benchmarks bounds the cost of exactly one component. Extrapolating a parser's cents-per-page to a workflow cost would price the easy fraction and ignore the correction labor that dominates real bills. All-attempt accounting applies from the first pilot case.

Exception judgment, client relationships, and liability sign-off have no public floor and no pilot evidence yet. The named exception classes for this workflow are CAM reconciliations, percentage rent, SNDA, and go-dark. Frequency and severity in a given portfolio are unknown-to-collect. The escalation loop is where critical errors are caught, so the build implication is to design automation around escalation, not to remove it.