Applied Compute halves cost boosts quality

This shows that Applied Compute is not just making legal AI cheaper, it is moving value from the model vendor into the workflow owner. Harvey’s Review Table can fan out into hundreds of thousands of model calls across large document sets, so cutting per call cost while improving answer and citation quality changes the unit economics of a product lawyers may use on live due diligence and review matters. AC2 matters because it lets teams keep the same API surface while swapping in a custom trained model behind it.1 2 3 4

• Review Table is a brutal workload for generic models. A lawyer can upload up to 10,000 files, ask many questions, and get a cited grid where each cell is its own inference call. In that setting, even small gains in per call quality and cost compound into a much better product margin and faster turnaround.1 2
• The quality gain is especially important because legal buyers care about whether an answer is tied back to the source text, not just whether the prose sounds plausible. Harvey had already rebuilt Review Table to produce more granular sentence level citations, so a model that lifts both answer score and citation score improves the part of the workflow lawyers actually check.1 5 6
• This also fits how large firms buy legal AI. Deployments often start with small seat counts, heavy security review, and close scrutiny of usage and ROI. Lower inference cost gives Harvey more room to price flexibly, absorb usage spikes, or offer demanding table based workflows without making every large matter economically painful.6 7

The next step is more workflow specific model training and routing. As legal AI shifts from chat into high volume tasks like review, drafting, and research, the winners will be the companies that own the evaluation loop, ship custom models behind stable endpoints, and keep improving quality without letting inference costs explode.3 4 6