Replacing Harvey AI: an AmLaw 200 case file
$686K reclaimed in 11 weeks. Same workflow at <5% of the cost on direct AWS / Anthropic billing.
In Q2 2025, an AmLaw 200 firm engaged us to evaluate Harvey AI for its 50-attorney litigation practice. The licence quote was $1.2K/seat/mo, $720K/yr fully loaded. The procurement committee had been negotiating for six weeks; the technology committee had been piloting for ten. By the time we got the call, the conversation was no longer “should we buy Harvey?” — it was “should we buy Harvey or build the same thing ourselves?”
We delivered the rebuild in 11 weeks. Y1 reclaim: $686K. Same workflow, <5% of the cost, on direct AWS / Anthropic billing. This is the case file.
What Harvey AI was being used for
The pilot covered four attorney workflows:
- Document review. Triaging discovery batches against case-specific criteria, producing a structured first-pass summary per document.
- Drafting assistance. Generating first drafts of motions, briefs, and contract redlines from a structured prompt + reference templates.
- Citation checking. Verifying that case citations in drafts were accurate, current, and on point.
- Research summarisation. Compressing case-law research outputs into attorney-readable executive summaries.
The four workflows together represented ~30% of associate-level work hours. The firm’s hypothesis: AI assistance would compress that 30% by half, freeing associate capacity for higher-value matters.
The pilot validated the hypothesis. The economics did not.
The architecture conversation
Our first technical review surfaced a structural observation: Harvey’s product is, architecturally, a thin orchestration layer over the same foundation models (predominantly Anthropic) that the firm could call directly. The differentiation was in three places:
- Prompt engineering — proprietary prompt templates tuned for legal workflows.
- UI — a polished interface calibrated to attorney workflows.
- Vendor relationship — single-vendor procurement, single-vendor support, single-vendor compliance posture.
Items (1) and (2) are reproducible by a competent team in a fixed-price engagement. Item (3) is the genuine value Harvey was charging for. The question for the firm was whether $720K/yr was a fair price for items (1)+(2) when (3) had alternatives.
What we built
The rebuild consisted of:
- Backend. A FastAPI service running on the firm’s existing AWS account, calling the Anthropic API directly. Document storage in the firm’s existing S3 bucket. Vector embeddings in pgvector (PostgreSQL extension), avoiding a separate vector DB. Audit logging into the firm’s existing Splunk instance.
- Frontend. A React SPA matching the firm’s brand, deployed to CloudFront. Authentication through the firm’s existing Okta tenant.
- Workflows. Four agent-style workflows (document review, drafting, citation checking, research summarisation) each with a defined prompt template, evaluation harness, and rollback procedure.
- Compliance. All data stayed in the firm’s AWS account. No data left the BAA boundary except via direct Anthropic API calls under the firm’s existing Anthropic enterprise agreement.
Total infrastructure cost: ~$2,800/mo at production load. Total Anthropic API cost: $4,200/mo at observed usage. Total run cost: ~$84K/yr, vs $720K/yr for Harvey. Y1 reclaim including build cost: $686K.
The 11-week schedule
| Weeks | Milestone |
|---|---|
| 1–2 | Architecture review, prompt-template extraction from pilot, evaluation harness scaffold |
| 3–5 | Document review workflow + UI + acceptance testing |
| 6–7 | Drafting workflow + template library |
| 8–9 | Citation checking + research summarisation |
| 10 | Compliance review, audit logging, security hardening |
| 11 | Production cutover, attorney training, parallel-run with Harvey for sign-off |
We ran the rebuild in parallel with the Harvey pilot for the final two weeks. Attorneys used both side-by-side. The acceptance criteria: per-workflow output quality at least matching Harvey on a blinded review by three senior partners. All four workflows passed.
What the firm got
- Full source code in their GitHub from week one.
- Infrastructure running on their AWS account, billed to their card.
- Prompt templates and workflow definitions documented in their repo.
- Operational runbooks, evaluation harnesses, and rollback procedures.
- A 90-day defect warranty (free fixes during the warranty window).
The firm’s IT team has owned the system since week 12. They’ve shipped two new workflows on top of it (deposition prep, expert-witness research) using the same architecture.
What this case isn’t
It isn’t an indictment of Harvey AI. Harvey is a well-engineered product, and for firms that don’t have the engineering capacity to own the rebuild, the per-seat licence may be the right answer. The case is an indictment of the per-seat pricing model at the price point Harvey was charging this specific firm — a 50-attorney deployment at $720K/yr, where the firm did have access to the engineering capacity (via us) and the architectural sophistication to evaluate the alternative.
For firms in similar positions — high per-seat enterprise AI spend, some appetite for owning the rebuild — the math is increasingly compelling.
How to know if your firm fits
Three diagnostics:
- Are you paying >$300K/yr for any single AI tool?
- Are <50% of your seats active monthly?
- Do you have, or can you get access to, an engineering team that ships?
If yes to all three, the rebuild conversation is worth having. We do a free 30-minute scoping call.
Read more: /upstream/harvey-alternative · /upstream/ · /case-studies/
Run the matching free calculator
Each one runs in 3 minutes and emails you an 8-page memo.