Spellbook is an AI redlining add-in that runs inside Microsoft Word. For SMB and mid-market legal teams, it sits on the contract you are already editing, suggests clause language against your playbook, flags deviations, and proposes redlines you can accept or reject without leaving the document. The redlines that survive review are the ones with recorded rationale. The ones that disappear silently cost you the next negotiation.
The workflow
1. Load the playbook into Spellbook. Upload fallback positions, preferred clauses, and dealbreakers. Spellbook compares incoming drafts against these. If you have a clause library already built, point Spellbook at it.
2. Open the counterparty draft in Microsoft Word. Spellbook activates in the side pane, scores risk, and highlights clauses that diverge from your playbook. For inbound NDAs, see the NDA review workflow; the redlining pattern here applies once you move past intake.
3. Review suggestions with attorney judgment. Spellbook proposes a redline. The attorney accepts, rejects, or modifies. For complex ambiguity, drop the clause into ChatGPT or Claude for a second reading. Hebbia is useful when the question requires pulling precedent from a corpus of prior executed contracts.
4. Record the rationale at the moment of decision. Why did you accept the liability cap at 2x fees instead of 1x? Why reject the mutual indemnification? Spellbook records the redline. It does not record the reasoning. Capture that somewhere the next negotiator can find it.
5. Send to CLM for signature and storage. Ironclad, DocuSign CLM, LinkSquares, or Evisort take the executed artifact. The signed PDF becomes the system of record.
Worked example: vendor MSA
A vendor sends an MSA. Spellbook flags three deviations: unlimited liability, 90-day payment terms, and an auto-renewal with no notice window. The attorney accepts Spellbook's redline on liability (caps at 12 months of fees), rejects the redline on payment terms (the business agreed to 90 days as a price concession), and modifies the auto-renewal redline to require 60 days notice.
Six months later, the same vendor sends a renewal. A different attorney opens it. Spellbook flags the same three clauses. The new attorney has no idea why payment terms were left at 90 days. They redline it back to 30, the vendor pushes back, the deal stalls.
Where the workflow leaks
Spellbook holds suggestions and redlines. Ironclad holds the executed artifact. Neither holds the why. The accept/reject rationale, the GC-approved deviation, the playbook exception that became a precedent: all of it lives in Slack threads, lawyer memory, or nowhere.
One way to solve this is a workspace like Dock that holds the redline rationale, the playbook deviation, and the approver decision as structured rows, with pointers (ironclad_workflow_id, spellbook_session_id) back to the CLM record. The CLM stays the system of record. Dock holds what the agent and the attorney interpreted around it. Irreversible actions, sending the redline to counterparty or executing the signature request, still pass through two-key handshakes.
Why this matters
According to Thomson Reuters, 77% of legal professionals use AI for document review and 58% for contract drafting (Thomson Reuters, 2024). Teams that compound learning across negotiations beat teams that treat each contract as a fresh problem. Rationale capture is the difference.
Start with the pillar: how to do legal review with AI.
FAQ
Is Spellbook safe for confidential contracts? Spellbook publishes SOC 2 Type II compliance and zero data retention agreements with its model providers (Spellbook, 2025). Confirm the configuration matches your firm's data policy before uploading.
How does Spellbook compare to general AI redlining tools? Spellbook is narrower than general AI redlining approaches because it lives inside Word and is tuned for transactional legal. ChatGPT and Claude are broader but require copy-paste. Hebbia is stronger when the question spans many prior contracts.
Does Spellbook replace a CLM like Ironclad or DocuSign? No. Spellbook handles redlining. Ironclad, DocuSign CLM, LinkSquares, and Evisort handle workflow routing, signature, and storage. They are complementary.
Who is accountable when an AI redline is accepted? The attorney who accepted it. Tools assist; they do not sign. Recording the attorney identity alongside the redline matters, which is why agent identity is foundational when AI is in the loop.