AI Agents for Lawyers

Comparisons

Spellbook vs Harvey: Word Add-In vs Enterprise Platform

Spellbook vs Harvey for transactional work: about $180 a seat against a reported $1,000-plus with a 20-seat minimum, scope, grounding and a ten-lawyer decision.

By Dr. Niklas SchmidtUpdated 13 November 20269 min read
On this page
  1. The arithmetic for ten lawyers
  2. What Spellbook is
  3. What Harvey is
  4. Grounding: a Word pane on GPT-5 against a multi-model stack with its own model
  5. Where they overlap: the Word pane
  6. Who buys each
  7. A decision for a ten-lawyer transactional practice
  8. Where this leaves you

For a ten-lawyer transactional practice, Spellbook's mid tier costs roughly $21,600 a year. Harvey, at reported prices, cannot be bought for ten seats at all, and twenty seats would run to about $240,000–288,000 a year before consumption charges. That arithmetic settles most cases. Above twenty seats and across several practice groups the question changes to what you need beyond contracts, and there Harvey starts to earn its price.

The arithmetic for ten lawyers

OptionPrice basis (Sep 2026)Ten seats, per yearMinimum
SpellbookTiers of roughly $99–$350 per user per month; mid tier about $180About $12,000–$42,000; about $21,600 at the mid tierNone reported
HarveyNot published; secondary sources estimate $1,000–1,200 per seat per month plus consumptionNot available at ten seats; about $240,000–288,000 for twenty20 seats (reported)
Microsoft Word Legal AgentInside Microsoft 365 Copilot at $30 per user per month (annual)About $3,600 on top of M365None
Claude Team (Claude for Word included)$25 standard or $125 premium per seat per monthAbout $3,000–$15,0005 seats
Clio Work$199 per user per month, standalone since 21 Apr 2026About $23,900None reported
LegoraReported from about $3,000 per user per yearAbout $30,00010 seats (reported)

The Spellbook and Harvey figures come from the pricing guide on this site and the secondary sources it cites, including a May 2026 price survey that anchors Harvey at $1,000–1,200 a seat and Spellbook at about $180; Harvey has never confirmed a list price, and a Reddit thread in September 2026 reported, without confirmation, a shift toward token-based billing. Treat every Harvey number as an estimate and every Spellbook number as a tier that your sales conversation will land somewhere inside.

What Spellbook is

Spellbook, from Toronto and led by Scott Stevenson, is a Word add-in that drafts, reviews and redlines contracts in the task pane, with an agent, Spellbook Associate, that carries multi-step tasks. It reports about 4,000 customers in 80-plus countries, with in-house teams now more than half of revenue and Nestlé among the named customers. It raised a $50m Series B at a $350m valuation in October 2025 and took a $40m debt facility from RBCx on 4 March 2026 to fund acquisitions in what its own release calls a consolidating market.

Two design choices define it. First, it grounds frontier models rather than training its own: Stevenson has argued in interviews that fine-tuning "encourages hallucinations", so the product feeds a GPT-5-class model your document, your playbook and your precedents and constrains the output. Second, since 30 June 2026 it has offered Autonomous Contract Management in early access, in which agents review and redline documents that arrive through DMS and email integrations, using multiple language models. That is the first move from "assistant in the pane" to "agent in the inbox", and it is the feature that most narrows the gap with Harvey for contract-heavy teams. The Spellbook review covers the product in detail.

What Harvey is

Harvey is a $15.5bn company after a $550m round on 9 September 2026, with annual recurring revenue above $400m, 3,000-plus customers, 80% of the top 100 law firms and 20% of the Fortune 500 by its own count. The product is a platform: Assistant, Vault for matter documents, Knowledge for firm precedents, 500-plus pre-built agents, an Agent Builder that had produced 25,000-plus customer-built agents by March 2026, Word and Outlook add-ins, Contract Intelligence (20 million-plus terms extracted), and since Harvey II on 18 August 2026, Memory, matter-aware Spaces and agents that inherit context. Harvey reported 700,000-plus agent tasks a day in June 2026.

The customer stories are platform stories, not contract stories. CMS rolled it out to 7,000-plus lawyers and staff across 21 member firms; GSK Stockmann reports 15–20% time savings on standard diligence and up to 75% on unstructured data rooms; PwC's Harvey partnership has produced 10,000-plus diligence red-flag reports; A&O Shearman sells four Harvey-built agents to its own clients. In the ILTA 2026 survey Harvey is used by 43% of firms overall and 84% of firms with 700-plus lawyers. The Harvey review and the small-firm question go further; the relevant point here is that the 20-seat minimum is not an accident. Harvey is priced for organisations that will use Vault, Knowledge and agents across practice groups, and its CEO has reportedly said in a Reddit AMA that the company is moving downmarket toward small firms and non-profits, which is worth watching but has not yet produced a published small-firm price.

Grounding: a Word pane on GPT-5 against a multi-model stack with its own model

Both products put a frontier model behind a legal interface; they choose the model differently.

Spellbook picks one frontier family and grounds it. Harvey runs OpenAI, Anthropic, Google and Mistral models side by side, turned on Claude Opus 5 on 24 July 2026 and Fable 5.1 on 1 September 2026, and has released Harvey Tenet, a post-trained open-weight legal model, in research preview. Multi-model routing matters for redlining specifically, because the models are moving fast on that task: Anthropic says Opus 5's first-turn NDA redlines scored "nearly double Opus 4.8", and GC AI's RedlineBench rose from 47.9 to 57.0 when it switched to Fable 5.1 in September 2026. A single-family product inherits its family's ceiling; a multi-model platform can move.

The counter-evidence is that model choice inside a vendor is not the same as model quality. In Vals' February 2025 report, lawyers beat every tool on redlining (79.7% against Harvey's 65.0%) even as Harvey won document question-answering at 94.8%. And a 30-lawyer firm's head of technology described a bake-off in August 2026 in which intake notes contained a disqualifying heirship issue and "every system missed it"; only Claude Opus on maximum effort caught it without prompting. His inference, that subscription vendors "are not incentivized to provide the best model due to their cost", is one to test rather than accept, but it means your pilot must ask each vendor which model and which reasoning effort it is actually running on your documents.

Where they overlap: the Word pane

The overlap is the task pane, and the pane is becoming a commodity. Microsoft's Word Legal Agent, released on 30 April 2026 in the Frontier preview and built by the team Microsoft hired from Robin AI, reviews a contract clause by clause against a playbook and writes tracked changes with explanatory comments; it requires Anthropic models to be enabled in the tenant. Claude for Word does tracked changes and is included in paid Claude plans. Litera's answer to the agent boom is a deterministic redline engine. A July 2026 survey of Word add-ins lists Spellbook, DraftWise, Definely, LegalOn, Harvey, Gavel Exec, Claude for Word, Copilot and Lexis Create in the same category. The Word Legal Agent page compares the Microsoft option in detail.

For Spellbook the pane is the product, and it competes on depth: playbooks, clause libraries, negotiation history, the Associate agent and now autonomous review. For Harvey the pane is one entrance to a platform that also holds the data room, the precedent bank and the agents that run overnight. If your lawyers live in Word and your matters are contracts, the pane is what you are buying, and the contract review and contract drafting pages show what a good pane does. If your matters are deals, with diligence, disclosure schedules and closing checklists, the pane is a small part of the work.

Who buys each

Spellbook's buyers are in-house teams (more than half of revenue), small and mid-sized firms, and anyone whose contract volume is high and whose deal volume is low. Harvey's buyers are large firms (80% of the top 100 by its count) and large in-house functions (HSBC, HubSpot, P&G, T-Mobile, Merck and PwC among the named ones; 500-plus in-house teams and 50 asset managers).

The reasons lawyers give are consistent. In a thread on rejecting Harvey and Legora in August 2026, the pro-Harvey argument was that "there is also no token usage to manage. It is an all inclusive price", and a practising lawyer defending the $15.5bn valuation wrote that "we cant upload our client files into Claude. Harvey is encrypted and trusted... vaults, agents, playbook workflows – it's amazing." Those are platform arguments: governance, all-in pricing and matter-scale storage. Nobody in those threads argued that Harvey's redline of a single NDA was worth $1,000 a month more than a Word add-in's.

A decision for a ten-lawyer transactional practice

  1. Count the seats honestly. Ten lawyers is under Harvey's reported minimum; if two paralegals and a knowledge manager would also use the tool, you are at thirteen, still under.
  2. Name the three contract types that consume most hours (say NDAs, supply agreements and SaaS terms) and write a one-page playbook for each. Any tool without a playbook produces generic redlines.
  3. Price the whole stack: the add-in, the Microsoft 365 or Claude seats you already pay for, and the research tool you use for anything beyond contracts.
  4. Run the three-document test below on Spellbook, on the Word Legal Agent if you have Copilot, and on Claude for Word as a control.
  5. Get the data terms in writing for the winner: no training on your content, retention period, and where the documents are processed. Module 2 of the AI Lab for Lawyers course redlines the same NDA in Claude, ChatGPT, Harvey and Legora side by side, which is the cheapest way to see what the platform premium actually buys.
  6. Buy the add-in, and set a date to revisit Harvey: when you pass twenty seats, when diligence volume would justify a Vault, or when a published small-firm tier appears.

The three-document test

Use closed files with names changed, and score each tool on the same sheet.

Document 1 – An NDA with three planted problems: a perpetual term, a
  one-way non-solicit, and a governing-law clause that contradicts the
  forum clause. Score: problems found (of 3), problems invented, quality
  of the replacement wording as a tracked change.
Document 2 – A supply agreement with a limitation-of-liability clause
  that caps at fees paid "in the preceding month". Score: did it flag the
  cap against your playbook's 12-month position, and did it explain why?
Document 3 – A 40-page SaaS agreement plus your playbook. Score: time to
  first full redline, number of playbook positions applied correctly,
  and whether the tool stopped to ask before changing anything it was
  unsure about.
For each tool record: model in use, reasoning effort, and whether the
  vendor could tell you either.

The playbook-automation guide turns the positions, fallbacks and escalations into a reusable file that works in any of these tools.

Where this leaves you

Spellbook and Harvey are compared because both put AI on contracts, but they are priced and built for different organisations. Under twenty seats, the add-in category (Spellbook, the Word Legal Agent, Claude for Word, Clio Work) is where the decision is, and the differences between those products are about playbooks, autonomy and data terms rather than the model. Above twenty seats and across practice groups, Harvey's Vault, Knowledge and agents are what the premium buys, and the honest comparison there is Harvey against Legora, not against a Word pane. The one rule that applies at every size: a redline you have not read is not a redline, whatever it cost.

Frequently asked questions

Is Spellbook cheaper than Harvey?

Yes, by an order of magnitude. Spellbook publishes tiers of roughly $99 to $350 per user per month, with a mid tier around $180. Harvey does not publish prices; secondary sources put it at $1,000 to $1,200 per seat per month with a 20-seat minimum plus consumption charges. For a ten-lawyer practice that is about $21,600 a year against a product that, at reported terms, cannot be bought at all.

Can Spellbook do what Harvey does?

For contract drafting and review inside Word, largely yes: playbook-based redlines, clause suggestions, and since 30 June 2026 an Autonomous Contract Management mode in early access that reviews and redlines through DMS and email integrations. It does not offer Harvey's matter Vault, firm knowledge search, 500-plus pre-built agents, Agent Builder or diligence at data-room scale. Spellbook is a contract tool; Harvey is a firm platform.

Which is better for a small law firm, Spellbook or Harvey?

Spellbook, as of September 2026, because Harvey's reported 20-seat minimum rules most small firms out and its price only pays off across several practice groups. Harvey's CEO has reportedly said the company is moving toward small firms and non-profits, so check for a published small-firm tier before deciding. Compare Spellbook against Microsoft's Word Legal Agent, Claude for Word and Clio Work rather than against Harvey.

Does Harvey have a Word add-in like Spellbook?

Yes. Harvey offers Word and Outlook add-ins and a Contract Intelligence capability that has extracted more than 20 million terms, but the add-in is a door into the platform rather than the product itself. Spellbook is Word-native by design and its whole interface lives in the task pane. Microsoft's own Word Legal Agent, released on 30 April 2026, now redlines against a playbook in tracked changes as well.

Which AI model does Spellbook use compared with Harvey?

Spellbook grounds frontier models of the GPT-5 class on your document and playbook rather than fine-tuning; its CEO has said fine-tuning encourages hallucinations. Harvey runs a multi-model stack across OpenAI, Anthropic, Google and Mistral, turned on Claude Opus 5 on 24 July 2026 and Fable 5.1 on 1 September 2026, and released its own post-trained open-weight model, Tenet, in research preview on 20 August 2026.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion