Spellbook

Spellbook provides a structured first pass on contracts in Microsoft Word. In Comprehensive Review, the user sets the represented party, jurisdiction, deal context, review settings, and review depth, then inspects the identified risks, reasoning, redlines, and comments. The product also supports playbooks, precedent-aware drafting, contract questions, and market comparisons. [S1][S2]

Spellbook's terms say it sends document data and user inputs to third-party LLMs, with stated zero-retention and no-model-training commitments for those providers. The terms also allow limited quality-assurance retention and derived anonymous statistics under stated conditions. Customers must not rely on the output alone as legal advice and should use a duly licensed lawyer when appropriate. Buyers need to verify the data handling and approval terms in their own contract. [S5]

Best for: in-house legal teams and transactional law firms with recurring contracts, a named lawyer accountable for final advice and approvals, an established Word-based process, and a focused pilot boundary.

Not for: buyers seeking unsupervised legal advice or automatic contract approval; teams without approved playbooks, permissions, and escalation paths; or work where the source document, jurisdiction, or risk profile cannot be reviewed by a qualified lawyer.

Run a first-pass contract review

• Trigger: A lawyer or authorized legal operator opens a draft agreement that needs a defined first review.
• Inputs: The contract text, represented party, jurisdiction, deal context, selected review depth and settings, plus any buyer-approved playbook or precedent context.
• Output: General Risks and Proofread results, reasoning, and suggested revisions or comments.
• Human checkpoint: The responsible lawyer inspects the flagged issues and evidence, edits suggestions as needed, and chooses whether to apply any change or route it for commercial approval.
• Success measures: time to first pass, material-issue recall on a reviewed sample, false-positive rate, suggestion acceptance or rejection, escalation rate, and time to final approval.

Spellbook's Comprehensive Review guide describes the party, jurisdiction, deal-context, setting, depth, reasoning, and apply/edit steps. It does not make the output legal advice. [S2][S5]

Review against a team playbook and precedents

• Trigger: A recurring agreement arrives that should follow the team's negotiated positions and templates.
• Inputs: The agreement, the buyer's approved playbook rules and fallbacks, relevant precedent language, counterpart and business context, and a named lawyer's review instructions.
• Output: Flagged risks, proposed redlines, comments, draft language, and an exception queue for review.
• Human checkpoint: A lawyer compares changes against the team's policy, matter facts, precedent, and approval thresholds before accepting a redline or giving advice.
• Success measures: playbook adherence, missed mandatory clauses, exception coverage, reviewer agreement, edit rework, negotiation cycle time, and unauthorized-change rate.

Spellbook's product page says users can build detailed playbooks with fallbacks and draft from preferences, precedents, and legal context. Those are vendor capability statements; the buyer should test the actual playbook, jurisdiction, and version behavior with its own contracts. [S1]

Use selected iManage material in a review request

• Trigger: A reviewer needs permitted matter documents to compare agreements, extract clauses, summarize risks, or ground a drafting request.
• Inputs: One or more documents manually selected from iManage, the user's existing iManage permissions, and a specific review or comparison question.
• Output: A focused answer, comparison, extraction, summary, or drafting context for lawyer review.
• Human checkpoint: The reviewer confirms that access, ethical walls, selected source documents, and the resulting analysis are appropriate for the matter before relying on it.
• Success measures: correct-document selection, permission errors, ethical-wall exceptions, answer traceability, useful citation or source context, manual rework, and review time.

Spellbook documents iManage selection for Ask and Associate, says existing iManage permissions and ethical walls are respected, and says it does not create a persistent library sync in the current integration. Verify the exact deployed scope and tenant configuration with both vendors. [S3]

The operating model is lawyer or authorized operator → select the agreement and matter context → choose the party, jurisdiction, review settings, and playbook → review returned issues and proposed edits → lawyer edits, accepts, rejects, or escalates → normal approval and execution process. Spellbook's documentation makes the user-driven settings and apply/edit steps visible; it does not document a substitute for legal judgment. [S2][S5]

For iManage, the documented path is administrator enablement where required → user authorization → manual selection of individual documents for a focused request. Spellbook states that it respects iManage permissions and ethical walls, but a buyer should test every role, workspace, error path, and matter boundary using non-production samples before enabling the integration widely. [S3]

Verified facts from primary sources

• Spellbook's current product page describes risk spotting, proposed redlines, detailed playbooks, drafting from preferences and precedents, deal-history or legal-source questions, and contract comparisons; it says the product is integrated with Microsoft Word. [S1]
• Comprehensive Review requires user selections for represented party, jurisdiction, deal context, settings, and depth, and lets users edit or apply revisions and comments after viewing results and reasoning. [S2]
• Spellbook documents manual selection of iManage documents for Ask or Associate and says it does not create a persistent iManage library or folder sync. [S3]
• The public legal terms define order-based subscription fees. No public starting price was established from the reviewed sources; the Help Centre documents a seven-day self-service website trial and paid-subscription cancellation process. [S5][S6]
• The terms say document data and prompts are transmitted to third-party LLMs to generate output and set out stated third-party LLM zero-retention and no-model-training commitments, quality-assurance retention up to 90 days, and customer ownership of customer data and outputs subject to the terms. [S5]

Vendor statements and contractual commitments to validate

• Spellbook says it supports contract review and drafting for in-house teams and law firms, including risk detection, redlines, playbooks, drafting, questions, and market comparisons. Test capability and reliability against the buyer's agreement types, jurisdictions, and approval rules. [S1]
• Spellbook says its iManage integration respects existing permissions and ethical walls. Confirm the implementation, user lifecycle, logs, exports, error behavior, and scope with the buyer's iManage administrator. [S3]
• The terms state that Spellbook will maintain agreements with third-party LLMs that prohibit retaining customer data longer than necessary to generate output and prohibit use for model training. Obtain the current provider list, DPA, subprocessor list, service configuration, exceptions, and contract incorporation before processing sensitive material. [S5]
• The privacy policy says optional Google Drive and Google Docs integrations have specified access and deletion boundaries. Validate the enabled feature set, retention, access tokens, residency, and contract terms for the buyer's tenant. [S4]

B2Bagents assessment

Spellbook is best assessed as a legal copilot with meaningful execution inside a review workflow, not as an autonomous counsel. That assessment follows from the documented user configuration and apply/edit steps, paired with the terms' explicit limitation that output must not be solely relied on as formal legal advice and should be lawyer-reviewed or vetted when appropriate. Its value case should be built on defensible review throughput and quality measures, not a generic time-saved claim. [S2][S5]

Material unknowns

• Edition-specific pricing, unit, minimum commitment, implementation, service levels, usage treatment, overages, renewal, and termination assistance.
• Exact Microsoft 365, Word, Google, document-management, API, SSO, role, audit-log, export, and admin-control behavior for the buyer's edition and tenant.
• Current model providers, subprocessors, data-residency options, DPA, regional terms, assurance reports, penetration-test scope, incident commitments, and any configuration that changes quality-assurance, retention, or derived-data treatment.
• Accuracy, issue recall, false positives, legal-source coverage, jurisdiction coverage, source citation quality, playbook consistency, and performance on the buyer's contracts and counterparty positions.
• How privilege, confidentiality, records retention, legal holds, document deletion, and post-termination access apply to every enabled workflow and jurisdiction.

Start with one recurring, lower-risk agreement type that already has a settled playbook, a named legal owner, and non-production or otherwise approved sample documents. Define the represented party, jurisdiction, review depth, allowed source material, change-application rules, escalation threshold, and the legal owner who may approve final advice. Do not start with novel, high-stakes, or privilege-sensitive matters until the data and control model is approved.

Establish a baseline for first-pass time, time to final approval, reviewer hours, material issue detection on independently reviewed samples, false positives, missed clauses, redline acceptance, edit rework, exception rate, playbook adherence, negotiation cycle, access incidents, and approval-log completeness. Test difficult cases: conflicting precedent, unfamiliar jurisdiction, ambiguous deal context, missing schedules, unusual indemnity or limitation language, multi-party agreements, misleading source text, tracked-change conflicts, sensitive matter documents, revoked iManage access, and failed integration authentication.

Agree success thresholds in advance: a meaningful time reduction without a material drop in independently reviewed issue coverage; correct playbook behavior; traceable lawyer review for every final redline; no unexpected access across ethical walls; defined data handling; and an acceptable rate of rework and escalations. Stop the pilot for material missed issues, unexplained or unsupported suggestions, loss of approval traceability, unexpected document exposure, access-control failure, unclear data terms, or any use that displaces required lawyer judgment.

• Harvey: existing directory entry oriented to legal transactional work; compare a broader legal AI environment with Spellbook's Word-centered contract review and playbook workflow.
• Crosby: existing directory entry for transactional legal work; compare workflow boundaries, lawyer review controls, integrations, and commercial model.
• Lawhive: existing directory entry with an AI-native legal-service model; compare a buyer-operated software workflow with a service-delivery boundary.

1. For our purchased edition, which inputs, documents, prompts, outputs, and metadata are processed or retained by Spellbook, third-party LLMs, and every subprocessor—and for how long?
2. Which Word, Microsoft 365, iManage, Google, SSO, role, audit, export, and admin controls are available to our tenant, and how do they behave in a role change, offboarding, or integration failure?
3. How does the product identify its legal source, jurisdiction, party position, and playbook basis for each suggestion, and how can a lawyer correct or challenge it?
4. What is the observed issue-recall and false-positive behavior for our agreement types, counterparty positions, and jurisdictions when results are compared with independent lawyer review?
5. What contractual protections cover privilege, confidentiality, data residency, retention, QA access, derived data, model providers, security incidents, deletion, export, and termination?
6. What exactly do implementation, training, subscription, usage, overages, support, renewal, cancellation, and exit cost for the first pilot and scaled deployment?