This sequence is labeled judgment. It is not a pilot result and not an ROI forecast.

First, deepen Grow intake write-back into Clio Manage. The public parse floor is the strongest of the set; the remaining risk is field mapping and exceptions on the firm's own mix.

Second, deepen Manage AI / Duo-era clause, date, and deadline write-back onto the matter. Quote the CUAD 2021 and ContractNLI 2021 cells only as paper floors, never as Clio. Gate on critical misses and on whether the write lands on the correct fields.

Third, treat research-memo drafting without graph write-back as a rent-or-wait surface. Harvey Tenet is the current instance of that surface: a vendor research preview, not a reason to pause graph write-back. First-party legal assistants, Microsoft or Google bundles, a DMS-native tool, and Thomson Reuters already ship the same shape.

Fourth, do not build billing-narrative or court-form completion until a pilot with deterministic checks exists. There is no honest public floor. Do not attach a savings or adoption number to any of these steps.

An advisory record freezes the feature list and the task-type mapping before quoting any floor. Public packs are feasibility evidence for a task type, not an in-product result: they say a competent generic system can do this class of work, and they say nothing about Clio's data, users, templates, or exception mix. A pack floor that survives a weak mapping is not conservative; for context-heavy features a public score can overestimate, so weak mappings carry no floor here.

The same published benchmarks would apply to any practice-management product: packs attach to workflow categories, never to a vendor. What differs for Clio is the register: Manage as the system of record, Grow as intake, Manage AI as the write-back surface, and the distribution Clio already holds. Feature-level pilots, not benchmarks, decide the build.

List token prices bound the model component of each feature. They do not bound the number that decides feature economics: correction. Reviewer minutes per accepted output, exception rates, and deterministic-check failures are the real bill, and none of that exists before a pilot. All-attempt denominators apply from the first pilot case: a failed generation is cost even when no output ships.

For that reason this report models no ROI, no adoption, and no pricing effect. An adoption number would be an assumption wearing a benchmark's clothes.

The durable asset in Clio is the record layer: matters, time, billing, deadlines, permissions. Features that read from and write back into that layer inherit its switching costs and are hard for a generic assistant to replicate without the same write authority. A drafting surface without the graph competes head-on with first-party legal assistants from the model labs and with Microsoft or Google suite bundles, on their distribution.

Customer insourcing is the ceiling on feature pricing: a firm can wire a current model to its own templates. A Harvey-class assistant attached to the firm's DMS, and a Thomson Reuters research-and-drafting seat, are the other structural hypotheses — named as hypotheses, not as observed displacements. The advisory question is which Clio features are cheaper to deepen than to rent once correction labor is counted. That is a pilot-and-price question, not a benchmark question.

The same floors apply to any practice-management product, because published benchmarks attach to task types, not vendors. A committee comparing two products in this category should expect the same public evidence and different company-specific findings: graph depth, write authority, distribution, and data terms. A public floor is never the tiebreaker between two products doing the same job.