Freshfields Anthropic vs Allen Overy Harvey Deal

Two of the largest publicly named BigLaw AI procurement deals share the same headline scale and almost nothing else underneath. Allen Overy Shearman announced its firm-wide Harvey AI rollout in February 2023, the original BigLaw AI flagship deal. Freshfields announced its multi-year Anthropic partnership on April 23, 2026, three years and two months later. Both reported firm-wide deployment. Both Magic Circle peers. Different vendors, different model architectures, different procurement postures. Here's the structural comparison: what each firm actually bet on, and what the next five years are likely to show. I covered the parallel capital-structure thesis (Blackstone's $50M into Norm Ai, which launched the Norm Law firm) in LawFuel, April 28, 2026, same structural dynamic, different layer of the stack.

The two deals at a glance

Allen Overy Shearman × Harvey AI (February 2023)

• Vendor: Harvey AI (vertical legal AI provider)
• Scale at announcement: 3,500+ lawyers, 43 offices
• Model architecture: Harvey runs on a mix of foundation models (OpenAI and Anthropic) under the hood; A&O sees Harvey's product layer
• Procurement structure: enterprise vendor relationship (buy)
• Public deal terms: not disclosed; Harvey does not publish a price
• Iteration history: A&O has deployed four generations of Harvey since 2023

Freshfields × Anthropic (April 2026)

• Vendor: Anthropic (foundation model provider)
• Scale at announcement: 5,700 employees, 33 offices
• Model architecture: direct foundation model access (Claude Opus 4.7 and successors); Freshfields builds workflow surface on top
• Procurement structure: co-development partnership (co-build)
• Public deal terms: not disclosed; multi-year, includes Cowork expansion plans and early access to future models
• Adoption: +500% in first six weeks per Freshfields press release

The headline similarity (firm-wide rollout at Magic Circle scale) hides the structural difference (vertical vendor vs foundation model).

Vertical vendor vs foundation model: the structural bet

A&O bet on vertical legal AI. Harvey is a company built specifically for legal workflows: contract drafting templates, transactional support flows, jurisdiction-specific knowledge, partner-facing UX tuned for legal review. The vendor absorbs the engineering burden of building legal-specific software; A&O receives finished product.

Details of the Freshfields arrangement are explained in the Freshfields and Anthropic deal terms.

Freshfields bet on foundation model access. Anthropic is the company that builds the model underneath several legal vendors (Harvey itself uses Anthropic models for many tasks). Freshfields receives raw model access plus co-design influence over how the model handles legal work generally, but pays in engineering and lawyer time to build the workflow surface.

The structural tradeoff:

• Vertical vendor partnerships ship faster. Harvey delivered legal-specific UX in 2023 that A&O could deploy immediately. Building equivalent UX on raw foundation model access takes 12-18 months of engineering work.
• Vertical vendor partnerships cap at the vendor's roadmap. Harvey's product team prioritizes features for the broad market. A&O can request features but doesn't control the prioritization.
• Foundation model partnerships ship slower but inherit improvements at the model layer indefinitely. When Opus 4.8 ships in (probably) Q3 2026, Freshfields' workflows benefit from the improvement automatically. Harvey's workflows benefit only when Harvey's product team integrates the new model into their stack.
• Foundation model partnerships compound advantage over time. Year 1 the vertical vendor wins on velocity. Year 3-5 the foundation model wins on capability accumulation.

The second-order point: this comparison was unambiguous in 2023 (Harvey was clearly ahead because foundation models couldn't handle legal work directly without scaffolding). It's becoming ambiguous in 2026 (foundation models can handle more legal work directly, narrowing the vendor-scaffolding advantage). It will likely flip in 2027-2028 (foundation models will handle legal work natively at quality that doesn't require vertical scaffolding for most tasks).

Procurement velocity vs structural advantage

A&O's deal in 2023 deployed in months. Harvey delivered ready-to-use software, A&O ran security review and MSA negotiation, lawyers started using the product within 60-90 days of contract signing. That procurement velocity was the structural advantage of the buy model.

Freshfields' deal will likely take 12-24 months to deploy at full scope. The +500% adoption curve in six weeks is impressive but reflects early use of available Claude features, not full co-build deployment. The Cowork expansion is future-tense in the press release; the agentic workflow co-development is underway, not finished.

The procurement velocity tradeoff:

• Buy delivers operational AI in the current fiscal year. Partner boards that need to demonstrate AI deployment to clients within 12 months structurally need buy.
• Co-build delivers structural advantage over 3-5 years. Partner boards willing to invest 12-24 months in deployment for structural model-layer advantage choose co-build.
• Buy + selective build hybrid splits the difference. Many BigLaw firms running CoCounsel or Spellbook for immediate deployment plus internal API tooling for specialty workflows operate this hybrid.

The second-order pattern: A&O's procurement choice in 2023 was structurally correct for 2023 conditions (foundation models weren't ready for legal work without vertical scaffolding). Freshfields' procurement choice in 2026 is structurally correct for 2026-2030 conditions (foundation models are increasingly capable on legal tasks directly). Both firms made the right bet for their respective timing.

Vendor lock-in dynamics: where each firm has flexibility

Vendor lock-in cuts both ways. Buy locks the firm into the vendor's product roadmap, pricing, and stability. Co-build locks the firm into the foundation model provider's model layer.

A&O's Harvey lock-in: Three years of Harvey deployment plus integrated training plus partner political capital create switching costs that make a vendor change expensive. A&O can renegotiate Harvey contract terms but switching to a different vertical vendor (Spellbook, CoCounsel) would unwind significant operational investment. The flexibility A&O has: Harvey itself runs on Anthropic models for many tasks, so A&O is indirectly exposed to Anthropic's model improvements through the Harvey relationship.

Freshfields' Anthropic lock-in: Multi-year direct relationship with Anthropic creates dependency on Anthropic's continued model improvement velocity. If Anthropic falls behind GPT-6 or Gemini 4 in late 2027, Freshfields is structurally locked into a relationship that's no longer ahead. The flexibility Freshfields has: foundation model relationships are easier to layer additional vendors on top of (run direct Anthropic plus run Spellbook plus run CoCounsel) than to switch out wholesale. Freshfields is also an early adopter of CoCounsel rebuild per Law.com coverage.

The second-order tradeoff: vertical vendor lock-in is transparent (you know what you're locked into). Foundation model lock-in is structural (you're locked into a model provider's continued performance). Both have switching costs; the pattern of those switching costs differs.

The third-order tradeoff: foundation model providers are subject to capability competition (multiple providers leapfrog each other on benchmarks every 6-12 months). Vertical vendors are subject to product-team velocity competition. Capability competition compounds faster than product-team velocity competition. Over five years, the foundation model lock-in is structurally less risky than vertical vendor lock-in, but only if the firm has the engineering capacity to switch foundation model providers when better ones ship.

Five-year outlook: which deal ages better

Predicting which deal compounds harder over five years requires assumptions about model improvement velocity, vertical vendor adaptation speed, and BigLaw partner board patience.

The baseline scenario (foundation models continue improving at 2024-2026 velocity, vertical vendors integrate but lag by 6-12 months):

• A&O on Harvey: continues to benefit from Harvey's vendor adaptation as Anthropic and OpenAI ship better models. Harvey's product team integrates new models into workflows; A&O sees better Harvey performance every 6-12 months. Per-seat pricing remains the dominant cost.
• Freshfields on Anthropic: sees model improvements at the moment Anthropic ships them, plus has co-design influence on what features ship. Internal tooling capability accumulates year over year. Engineering investment compounds into firm-owned IP.

The accelerated scenario (foundation models improve faster than vertical vendors can integrate):

• A&O on Harvey: falls increasingly behind Freshfields-style direct-access firms. Harvey's vendor scaffolding becomes a drag on capability rather than an enabler. A&O faces a procurement decision in 2027-2028: stay on Harvey at increasing capability lag, or bear the switching cost to move toward direct Anthropic.
• Freshfields on Anthropic: compounds advantage as model capability grows faster than vertical vendor product cycles can integrate.

The slower scenario (foundation model improvement velocity slows; vertical vendors catch up):

• A&O on Harvey: wins. Harvey's product depth and vertical-specific UX become more valuable than direct foundation model access. A&O looks prescient.
• Freshfields on Anthropic: has paid co-build coordination cost without proportional capability advantage.

The second-order honesty: nobody knows which scenario plays out. Both firms made structurally defensible bets given their respective timing and risk preferences. The five-year outlook is a probability distribution, not a forecast. My structural read leans toward the accelerated scenario, but with material uncertainty. Read the Anthropic Legal Ecosystem 90-day map for the broader pattern this deal sits inside.

The bottom line. The verdict on fit: A&O's Harvey deal in 2023 was structurally correct for 2023 conditions; foundation models weren't ready for legal work without vertical scaffolding. Freshfields' Anthropic deal in 2026 is structurally correct for 2026-2030 conditions; foundation models are increasingly capable on legal tasks directly. Both firms made the right bet for their timing. The five-year question is whether foundation model improvement velocity continues at 2024-2026 pace; if yes, Freshfields' bet compounds harder. If foundation model progress slows, A&O's vertical vendor bet ages better. Most BigLaw firms today don't have to choose between these two paths exclusively. Hybrid procurement is structurally available.