Co-Build vs Buy Legal AI Procurement Models 2026

When Freshfields announced its co-build deal with Anthropic on April 23, 2026, it named a procurement model that didn't exist publicly in BigLaw before. Co-build sits next to two existing models: buy (enterprise vendor relationship: Harvey, Spellbook, CoCounsel) and build (internal engineering on raw API access). Each model has a structural fit, a cost profile, and a set of firms it serves well. Most procurement teams default to buy because that's what their RFP processes are built for. The 2026 reality is that the right model depends on firm size, practice mix, and engineering depth, not on what your procurement function happens to know how to negotiate. Here's the operator read on which model fits which firm.

The three procurement models in one sentence each

Three models, structurally distinct:

- Buy. Enterprise vendor delivers finished software at a per-seat price; firm deploys, trains, uses. Examples: Harvey AI, Spellbook, Thomson Reuters CoCounsel. Procurement velocity is fast (weeks to months); deployment is shaped by vendor's roadmap. - Build. Firm contracts directly for foundation model API access (Anthropic at $5/M input + $25/M output for Opus 4.7, OpenAI similar) and writes internal tooling. Procurement is a finance question (usage-based); deployment is shaped by internal engineering capacity. - Co-build. Firm collaborates with foundation model provider on workflow design and feature prioritization, in exchange for early access and influence. Procurement velocity is slow (6-12 months negotiation); deployment is shaped jointly by firm and provider. Examples: Freshfields × Anthropic.

The second-order point: most firms operate in a mix. A firm running CoCounsel for research, Spellbook for contracts, and direct Claude API for ad-hoc work is running buy + build simultaneously. Co-build is the one model that's structurally exclusive; you can't co-build with multiple foundation model providers without conflicts of interest.

Buy: the procurement default and where it fits

Enterprise vendor procurement is what most BigLaw and mid-market law firm procurement functions are structurally good at. RFP, security review, MSA negotiation, per-seat pricing, deployment, training. The vendor delivers a finished product; the firm pays per seat per month and uses what's shipped.

When buy fits:

- Firms 100-500 lawyers without dedicated legal-tech engineering depth. Internal build requires sustained engineering investment that small and mid-market firms can't underwrite. - Workflows where the vendor's product roadmap matches firm needs. Spellbook's contract review, Harvey's transactional support, CoCounsel's Westlaw-grounded research are mature products. Building equivalents internally costs more than buying for most firms. - Procurement timelines under 6 months. Buy can deploy in 30-90 days. Build takes 6-18 months. Co-build takes 9-18 months. If the partner board wants AI in workflow this fiscal year, buy is the only option.

When buy fails:

- Workflows where the vendor's roadmap doesn't match. Vertical legal AI vendors prioritize broad-market features. Firms with niche practice areas or specialty workflows may pay enterprise prices for products that don't fit their use case. - Long-term cost compounding. Per-seat enterprise pricing compounds over time. A 500-lawyer firm paying $400/seat/month for Harvey is $2.4M/year recurring. Over five years, that's $12M with no equity build-up: pure operating expense. - Vendor lock-in on procurement contracts. Multi-year buy contracts with switching costs reduce procurement flexibility when better tools ship.

Industry observers report Harvey AI tier prices for AmLaw 100 firms in the $1,500-$2,000+/seat/month range and mid-market in the $1,200-$1,500/seat/month range per Artificial Lawyer 2025 reporting, not vendor-confirmed (Harvey doesn't publish pricing). Spellbook pricing is quote-only with industry estimates around $180-$300/seat/month per secondary sources; Spellbook hasn't confirmed those publicly.

Build: when foundation model API access plus internal engineering wins

Build means contracting directly for foundation model API access and writing internal tooling on top. Anthropic Opus 4.7 at $5 per million input tokens / $25 per million output tokens. OpenAI GPT-5.5 at $5/M input / $30/M output. Google Gemini at competitive rates. The firm pays for tokens consumed, not for finished software.

When build fits:

- Firms with at least one full-time legal-tech engineer. Build requires sustained engineering capacity. A solo legal-tech consultant or a part-time engineering function can't maintain build infrastructure. - 3-5 high-priority repeatable workflows. Build amortizes engineering investment across workflows that run repeatedly. One-off use cases don't recover the build cost. - Firms with practice areas underserved by vendors. Boutique IP litigation, specialty transactional practices (energy, sovereign debt), specific regulatory work where vertical vendors don't have mature products. - Firms with strict data-residency or audit requirements. Building on direct API access gives the firm full control over data flow, logging, and storage. Buying inherits the vendor's data handling.

The cost economics. A 200-lawyer firm running build typically lands at $30-80/lawyer/month in compute costs (varies wildly by use case intensity) plus $200-500K/year in legal-tech engineering salaries. Total: $300K-$1.5M/year for the firm. Compare to buying enterprise legal AI at $400/seat/month × 200 seats = $960K/year, comparable cost, but build retains internal IP and capability.

The second-order tradeoff. Build accumulates internal capability over time. Year 1 build is mostly cost; year 3 build is capability the firm owns. Year 5 build is a competitive moat. Buy stays expense-only across all years.

Co-build: what Freshfields just demonstrated

Co-build is the third procurement model, structurally distinct from buy and build. The firm collaborates with the foundation model provider on workflow design, feature prioritization, and (sometimes) model behavior. The provider absorbs some engineering burden in exchange for shaping the model's behavior on legal work.

When co-build fits:

- Firms with $2B+ revenue. Co-build deal value typically lands $10-30M+/year based on the Freshfields scope. Firms that can't underwrite that without restructuring partner economics aren't candidates. - Global footprint of 20+ offices. Co-build deployment requires centralized AI governance across geographies. Concentrated firms can't operate the deployment surface. - Industrial-scale training data quality. The firm needs to offer the foundation model provider something the provider wants, usually high-quality legal work product feedback at industrial scale. Magic Circle and AmLaw 50 firms have it; mid-market firms generally don't. - Sophisticated procurement function. Co-build deals require IP, data handling, and exclusivity terms outside the vendor's standard MSA. Procurement functions without dedicated legal-tech specialists struggle to negotiate the substantive terms.

When co-build fails:

- Time pressure. Co-build deals take 6-12 months to negotiate and 12-24 months to deploy at scale. If the partner board wants AI in workflow this year, co-build doesn't deliver. - Smaller firms. Most BigLaw firms (including most AmLaw 100 firms) don't qualify on revenue, footprint, or training-data-quality criteria. - Risk-averse partner cultures. Co-build commits the firm to a multi-year exclusive-ish relationship with one foundation model provider. Partner boards uncomfortable with that level of vendor concentration won't authorize it.

The upside that compounds. Early access to future models is the structural advantage that compounds hardest over time. Read the early access models competitive edge spoke for the talent-attraction angle.

The decision framework: which model for which firm

Five questions to answer before choosing procurement model:

1. What's the firm's annual revenue? Under $500M: buy or consumer-tier. $500M-$2B: buy with selective build for high-priority workflows. $2B+: co-build is on the table. 2. What's the legal-tech engineering capacity today? Zero engineers: buy or consumer-tier. One full-time engineer: buy with selective build experiments. Three to five engineers: build is viable. Ten-plus engineers: build or co-build. 3. How many offices need uniform deployment? Under 5 offices: buy is structurally simpler. 5-20: buy or build with central governance. 20+: build or co-build with centralized AI function. 4. What's the procurement timeline? Under 6 months: buy. 6-12 months: buy or selective build. 12-24+ months: build or co-build. 5. What's the practice mix? Disputes-heavy: CoCounsel-grounded research is hard to replicate, lean buy. Transactional-heavy: foundation model access compounds (multi-session memory, long-document handling), lean build or co-build. Mixed: mix procurement.

The second-order pattern: most firms answer these questions and end up with a hybrid. Buy for research workflows where vertical vendors are mature; build for high-priority specialty workflows; co-build only at the very top. The pure single-model procurement is rare outside the smallest and largest firms.

For the mid-market replication detail, the structural answer is: 70% of co-build benefit at 5% of cost via prompt-template infrastructure plus consumer-tier Claude Pro plus written AI use policy.

The Bottom Line: The verdict on fit: Buy fits 100-500 lawyer firms with mature workflow needs and short procurement timelines. Build fits 50-500 lawyer firms with sustained engineering capacity and 3-5 high-priority repeatable workflows. Co-build fits Magic Circle and AmLaw 50 firms with $2B+ revenue, global footprint, and industrial-scale training-data-quality offers. Most firms operate hybrid procurement: buy for research where vertical vendors are mature, build for specialty workflows, co-build only at the very top. The wrong model for a firm wastes 12-24 months of procurement effort; the right model compounds capability over 3-5 years.