The complete guide to AI contract review

This guide is for the people who end up holding the contract: in-house counsel, deal desk and sales operations leads, procurement managers, and founders at companies with no lawyer on staff. If a vendor has pitched you an AI tool that reviews NDAs in seconds, or someone on your team is already pasting MSAs into a chatbot, this is the background you need before deciding what to buy and what to forbid.

After reading it you should be able to tell AI contract review apart from contract lifecycle management, judge vendor accuracy claims against independent research, ask the right data security questions, compare the three main categories of tools and their pricing models, and run a two-to-four week pilot that tells you whether a tool works on your paper.

This guide compares and explains. It is not legal advice, and it is not a substitute for a qualified lawyer reviewing your contracts or your professional obligations.

Summary

• AI contract review reads a contract, finds the clauses that matter, compares them with your positions, and proposes redlines. CLM manages the contract’s whole life; review is one step in it.
• The playbook is the product. A tool with no playbook behind it is applying someone else’s idea of “standard.”
• Independent tests show AI tools match or beat lawyers at extracting information but trail them at redlining.
• Hallucination is measured, not hypothetical: even purpose-built legal research tools produced wrong answers on more than one in six queries in a Stanford study.
• Business tiers of the major AI vendors do not train on your data by default, but retention, sub-processors, and consent duties still need checking.
• Pricing runs from about $18 per user per month for a general assistant to custom, volume-based quotes for CLM suites, as of October 2026.
• Pilot on 30 to 50 of your own past contracts with known answers, and measure missed issues before you measure time saved.

1. What AI contract review is, and how it differs from CLM

AI contract review software takes a contract, usually a Word document or PDF, and does four jobs.

Clause extraction. It finds and labels the provisions that matter: limitation of liability, indemnification, governing law, auto-renewal, assignment, termination for convenience. This is the oldest and most studied task. The CUAD dataset, built by The Atticus Project, labeled 41 clause categories across 510 commercial contracts, and found that the relevant clauses make up only about 10% of a typical contract, which is why the task is described as finding needles in a haystack.¹1 Source: Hendrycks, Burns, Chen and Ball, “CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review,” arXiv

.06268, accessed Oct 2026.

Playbook-based redlining. A playbook is your organization’s written negotiating position for each clause: the preferred language, the fallback you will accept, and the point at which a human must approve. The tool compares the counterparty’s draft to the playbook and proposes tracked changes. Juro, for example, describes its AI Review feature as auto-redlining contracts against custom playbooks.²2 Source: Juro, “Pricing”, accessed Oct 2026. Feature names and pricing basis as described by the company. Ironclad says its review agent in Word lets lawyers apply preferred positions and switch to fallbacks, with a rationale attached to each suggested edit.³3 Source: Ironclad, “Introducing Contract Review Agent for Microsoft Word”, June 16, 2026, accessed Oct 2026. Capabilities as described by the company.

Risk flagging. It marks provisions that deviate from the playbook or from market norms. Ironclad’s examples include missing indemnification caps, aggressive liability language and non-standard termination terms.³

Obligation tracking. It pulls out dates and duties (renewal notice windows, payment terms, reporting requirements) so someone can act on them after signature. Ironclad’s Word agent generates a summary covering obligations and key dates, and Juro’s AI Extract tags incoming contracts.³²

Review versus CLM

Contract lifecycle management (CLM) software covers the whole life of an agreement: intake, templates, approvals, negotiation, e-signature, a searchable repository, and renewals. AI review is one stage of that life, the negotiation stage, plus some post-signature extraction.

A CLM is a system of record that takes months to implement and touches sales, procurement and finance. A review tool can be a Word add-in one lawyer installs this afternoon. The boundary is blurring as each adds the other’s features, so ask which problem you have. If contracts get lost, renewals get missed and nobody knows what was signed, you have a CLM problem. World Commerce & Contracting estimates the average business loses almost 9% of contract value each year through poor contract management, with best performers at about 3%.⁴4 Source: World Commerce & Contracting, “Contract Management Whitepaper, August 2025”, accessed Oct 2026. If contracts are tracked fine but sit in the legal queue for a week, you have a review problem.

2. Who uses it, and who it is not for

In-house legal teams are the core buyers. Their use case is volume: NDAs, vendor agreements, data processing addenda, and customer paper that arrives faster than the team can read it. The tool does the first pass, and the lawyer spends time on the clauses it flags.

Sales and deal desk teams use review to keep deals moving. A deal desk (the group that approves non-standard pricing and terms) can run the customer’s markup against the playbook and send only true exceptions to legal. This works when the playbook clearly says what sales may concede without approval.

Procurement reviews supplier paper, which is the mirror image of sales. The risks differ (supplier liability caps, data security terms, audit rights), so procurement usually needs its own playbook.

Small companies without counsel are the most tempted and the most exposed. A founder reviewing a reseller agreement with a general-purpose chatbot gets fluent output with no one qualified to check it. The tool can help a non-lawyer understand what a clause says and prepare better questions for outside counsel. It should not be the last reader of a contract that carries real liability.

It is not for: bespoke, high-value transactions where every clause is negotiated from scratch (M&A, complex financing), contracts in jurisdictions or languages the vendor has not tested, and teams that will not maintain a playbook.

Where contract review sits in the wider sales process is covered in the complete guide to building a B2B revenue stack.

3. How it works: models, playbooks, and Word

Almost every current product sits on a large language model (LLM), licensed from a frontier lab or fine-tuned by the vendor. The vendor’s work is what surrounds the model: parsing the document so clause boundaries survive, retrieving the right playbook rule for each clause, constraining the model’s output to tracked changes, and showing why each change was proposed.

The playbook layer

A playbook converts judgment into rules the model can apply. For each clause type it typically holds:

• the preferred position and sample language,
• one or more acceptable fallbacks,
• red lines that are never acceptable,
• escalation rules (who approves what).

Vendors differ in how much of this they supply. LegalOn advertises more than 135 pre-built playbooks in its base plan, with custom playbooks in higher tiers.⁵5 Source: LegalOn, “Pricing”, accessed Oct 2026. Security and training statements are the company’s own. Ironclad’s support documentation says its Professional Services team is still required to create and tune playbooks before customers use playbook redlining, and that its tool for generating playbooks from precedents was in early access.⁶6 Source: Ironclad Support, “Ironclad Jurist for Microsoft Word”, accessed Oct 2026. The accuracy figure is from Ironclad’s own testers. Read that as a signal about effort: a playbook is a project, not a setting.

The Word add-in

Lawyers negotiate in Microsoft Word with tracked changes, so most serious tools live there. Spellbook describes its core product as a Word add-in covering review, drafting, benchmarks and playbooks.⁷7 Source: Spellbook, “Pricing”, accessed Oct 2026. Security and training statements are the company’s own. LegalOn includes a Word add-in alongside a web platform.⁵ Ironclad’s review agent works on any document opened in Word, with more context when the contract is also managed in Ironclad.³ Ironclad renamed its Jurist assistant to Contract Review Agent on October 8, 2026, so older documentation still uses the old name.⁶

The practical test is formatting. A counterparty’s draft with three rounds of tracked changes, comments and renumbered sections is hard to parse. The Vals benchmark found tools did better when clauses arrived as clearly labeled plain text than when given a formatted, redlined document, which they struggled to read.⁸8 Source: Vals AI, “Vals Legal AI Report,” February 27, 2025, accessed Oct 2026. The report’s narrative and results table give different redlining scores for the best tool; the table figure is used here.

Contract review is one kind of AI agent work; the complete guide to AI agents in enterprise operations covers the broader category.

4. Accuracy and hallucination: what the studies show

Vendors publish accuracy figures from their own tests. Ironclad, for example, cites an internal tester figure for how often its output is grounded in the provided playbook.⁶ Those numbers are worth asking about but not worth relying on. Independent evidence is thinner and more sobering.

Extraction is close to solved; redlining is not

The most useful independent study so far is the Vals Legal AI Report, published in February 2025. It compared legal AI tools with a baseline of lawyers working without AI. On data extraction, the best tool (Harvey Assistant) scored 75.1% against a lawyer baseline of 71.1%. On redlining, the lawyers scored 79.7%, and the best tool scored 65.0%.⁸ The report also noted that on complex changes to a single contract, the tools often inserted standard text instead of making the nuanced edit, and that vendors were tested on open-ended assistant features rather than dedicated redlining workflows, which may understate them.⁸

So extraction and obligation tracking are where AI first passes are already competitive. Redlining, the job most vendors lead with, is where human review adds the most.

Hallucination

A hallucination is output that is fluent and wrong: a clause the contract does not contain, a citation that does not exist, a confident statement of what “market standard” is. Stanford RegLab researchers found that general-purpose chatbots answering verifiable questions about federal court cases hallucinated between 69% of the time (GPT-3.5) and 88% (Llama 2).⁹9 Source: Stanford Law School, Dahl, Magesh, Suzgun and Ho, “Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models,” 16 J. Legal Analysis 64 (2024), accessed Oct 2026. Those were 2023-era models, and current models are better. The more relevant finding is the follow-up: Stanford RegLab and HAI tested purpose-built legal research tools that use retrieval to ground their answers, and found Lexis+ AI and Ask Practical Law AI gave incorrect responses more than 17% of the time, and Westlaw AI-Assisted Research more than 34%.¹⁰10 Source: Stanford HAI, “AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries”, May 2024, accessed Oct 2026. Over 200 pre-registered legal queries.

Those studies tested legal research, not contract review, where the source document is right there. The risk changes shape rather than vanishing: a summary that misstates a notice period, a flagged “missing” clause that is present under another heading, or a fallback that quietly shifts risk back to you.

The professional consequence is well documented. In Mata v. Avianca, a federal judge in New York fined lawyers and their firm $5,000 in June 2023 after they filed a brief citing cases ChatGPT had invented. The court said there was nothing inherently improper about using a reliable AI tool, but that lawyers have a gatekeeping duty to check what they file.¹¹11 Source: ABA Journal, “Lawyers who ‘doubled down’ and defended ChatGPT’s fake cases must pay $5K, judge says”, accessed Oct 2026. The same logic applies to a signed contract.

5. Data security, confidentiality, and professional duties

A contract under negotiation contains pricing, customer names, security commitments and sometimes personal data. Treat a review tool like any vendor that will hold your most sensitive documents.

What to ask the vendor

SOC 2. SOC 2 is an attestation report produced by an independent CPA under AICPA standards, covering a service organization’s controls for security, availability, processing integrity, confidentiality and privacy.¹²12 Source: AICPA & CIMA, “SOC 2”, accessed Oct 2026. Ask for the Type II report, which tests whether controls operated over a period, and read the exceptions section. Spellbook and LegalOn both state they hold SOC 2 Type II.⁷⁵

Model training. Ask, in writing, whether your documents or prompts are used to train any model, the vendor’s or the underlying lab’s. LegalOn says its private AI does not use your data to train third-party foundation models.⁵ Spellbook says it has zero data retention agreements that prevent customer data from being used for training.⁷ Anthropic says that by default it does not use inputs or outputs from commercial products such as Claude for Work and its API for training, with exceptions if users submit feedback or opt in.¹³13 Source: Anthropic Privacy Center, “Is my data used for model training?”, accessed Oct 2026. OpenAI says that by default it does not train on content from ChatGPT Business, Enterprise and Edu workspaces.¹⁴14 Source: OpenAI Help Center, “Privacy Center in ChatGPT”, accessed Oct 2026. Confirm current terms with OpenAI before procurement.

Retention. “Not used for training” is different from “not stored.” OpenAI says ChatGPT Enterprise owners can set a custom retention period with a minimum of 90 days.¹⁴ Ask where documents are stored, for how long, in which region, and which sub-processors (including the model provider) see them.

Consumer accounts. The defaults above apply to business tiers. An employee using a personal chatbot account is outside them. Decide what is allowed and write it down.

What bars have said

For lawyers, data handling is also an ethics question. In July 2024 the American Bar Association issued Formal Opinion 512, its first formal opinion on generative AI. It ties use of these tools to existing duties under the Model Rules: competence (understanding the benefits and risks of the technology), confidentiality, communication with clients, and reasonable fees. On fees, the opinion says a lawyer may bill for time spent entering information and reviewing output, but generally may not charge a client for learning to use the tool.¹⁵15 Source: The Florida Bar News, “ABA issues first ethics guidance on a lawyer’s use of AI tools”, July 31, 2024, accessed Oct 2026. Read the full opinion before relying on it.

State guidance goes further on specifics. Florida Bar Ethics Opinion 24-1 (January 2024) recommends obtaining a client’s informed consent before using a third-party generative AI program if doing so would disclose confidential information, and warns that “self-learning” tools may store client information and reveal it in response to later queries.¹⁶16 Source: The Florida Bar, “Ethics Opinion 24-1”, January 19, 2024, accessed Oct 2026. The State Bar of California’s practical guidance says lawyers should not enter confidential client information into a tool with material confidentiality or security risks without informed client consent, and that reasonable diligence means more than relying on a vendor’s marketing assurances.¹⁷17 Source: State Bar of California, “Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law”, accessed Oct 2026. The State Bar revised this guidance in 2026; check the current version. These opinions are advisory and vary by state, so in-house teams should check the rules where their lawyers are licensed.

The vendor security questions here overlap with what a penetration test covers; see the complete guide to penetration testing for SaaS companies.

6. The tool landscape and what it costs

The market splits into three categories. Named tools below are examples, listed neutrally; none is a recommendation.

	Word-native reviewers	CLM suites with AI	General-purpose LLMs
Examples	Spellbook, Gavel Exec, LegalOn	Ironclad, Juro	Microsoft 365 Copilot, ChatGPT Enterprise, Claude for Work
Where it runs	Word add-in, often plus web app	CLM platform, with Word add-in	Chat interface, Office apps
Playbooks	Built-in and custom	Custom, often with vendor help	Whatever you write into the prompt
Repository and obligation tracking	Limited to growing	Core feature	None by default
Typical pricing model	Per seat	Custom, by contract volume or platform	Per seat, low list price
Setup effort	Days	Months	Hours, but discipline is on you
Best for	Legal teams negotiating in Word	Companies that also need a system of record	Reading and summarizing, low-stakes first drafts

Sources: vendor pricing and product pages.⁷⁵¹⁸18 Source: Gavel, “Gavel Exec Pricing”, accessed Oct 2026. List prices. ²³¹⁹19 Source: Microsoft, “Microsoft 365 Copilot Plans and Pricing”, accessed Oct 2026. List prices; promotional terms per Microsoft.

Pricing examples, as of October 2026

Most dedicated vendors do not publish prices. These are the ones that do, or that state their model.

Tool	Pricing model	Published list price (as of October 2026)
Gavel Exec	Per user, monthly or annual	$160 per user per month, or $1,740 per user per year billed annually; team pricing for 10+ seats18
Spellbook	Per seat, custom quote	Not published; 7-day free trial7
LegalOn	Team size and capabilities, custom quote	Not published; all plans include unlimited usage5
Juro	Contract volume and integrations, not per user	Not published; unlimited users on all plans2
Microsoft 365 Copilot Business	Per user add-on to an eligible Microsoft 365 plan	$21 per user per month on annual billing, promotionally $18 for the first year through December 31, 2026; $25.20 monthly billing19
Microsoft 365 Copilot (enterprise)	Per user add-on	$30 per user per month, paid yearly2020 Source: Microsoft, “Microsoft Copilot Plans and Pricing, Enterprise”, accessed Oct 2026. Requires a qualifying Microsoft 365 license.

A few things the tables hide. Gavel’s paid plan includes 1,000 completions per month, so heavy users should check the cap.¹⁸ Microsoft notes that some advanced AI features carry usage-based charges beyond the license.¹⁹ CLM pricing is usually a platform fee plus implementation, and Juro charges extra for deeper integrations such as Salesforce and HubSpot.² For a fair comparison, price the total cost per reviewed contract over a year, including setup and playbook work, not the seat price.

7. How to run a pilot on your own contracts

Build a test set

1. Pick 30 to 50 past contracts of the two or three types you see most (for many companies: NDAs, customer MSAs, vendor agreements). Include some messy ones: scanned PDFs, heavily marked-up drafts, counterparty paper.
2. Create an answer key. For each contract, have a lawyer record the issues that should be flagged and the changes your team actually made. This is the slow part, and it is what makes the pilot worth anything.
3. Write a pilot playbook covering the 10 to 15 clauses that drive most of your negotiation. Use the same playbook for every tool you test.
4. Check the data terms first. Do not upload real contracts until the vendor’s security review and, for lawyers, the consent questions in Chapter 5 are resolved. Redacted or already-public contracts are an option for an early round.

Measure the right things

Score each tool’s output against the answer key, clause by clause.

• Missed issues (recall). Of the problems your lawyer found, how many did the tool flag? This is the number that matters most, because a miss is invisible.
• False alarms (precision). Of the issues the tool flagged, how many were real? Too many false alarms and reviewers start ignoring flags.
• Redline usability. What share of suggested edits could be accepted as written, accepted with light edits, or had to be rejected?
• Fabrication. Any summary or flag that states something the contract does not say. Count these separately; one is a finding, several is a disqualifier.
• Time. Measure reviewer time with and without the tool on comparable contracts, including the time spent checking the tool’s work.

Precision and recall can be combined into an F1 score, which some vendors quote. Ask vendors which contract types and which clause set their figure covers, then compute your own on your test set.

Run the pilot for two to four weeks with the reviewers who will use the tool daily. Their willingness to keep using it is a result too.

8. Rollout, playbook building, and common mistakes

Rollout

Start with one contract type and one team. NDAs are the usual first choice: high volume, low variance, and low stakes if something is missed. Expand to the next contract type only after reviewers have used the tool on the first for a month and the playbook has been revised at least once based on what they saw.

Set a review rule from day one. A common pattern: AI output on low-risk paper may go out after a single human read; anything touching liability, indemnity, data protection or IP gets a lawyer’s sign-off regardless of what the tool says. Write the rule down and make it visible in the tool if the tool supports approval routing.

Building the playbook

Start from what you already do, not from a template. Pull the last 20 negotiated versions of your most common agreement and record where you held firm and where you gave ground. That history is your real playbook. Then:

• write each position in plain language with a sample clause,
• state the fallback and who can approve it,
• name the red lines,
• assign an owner who updates the playbook each quarter.

If your organization cannot agree on its positions, a tool will not resolve that. It will apply whichever version someone typed in.

Mistakes to avoid

Buying before the playbook exists. Without your positions encoded, the tool applies a generic standard, which may not be yours.

Piloting on the vendor’s contracts. Clean sample NDAs flatter every tool. Your counterparties’ paper is the test.

Measuring time saved and nothing else. A tool that halves review time and misses one indemnity cap in twenty has not saved anything.

Letting non-lawyers treat output as approval. For deal desk and procurement users, the tool should route exceptions, not grant them.

Ignoring the consumer-account problem. If the approved tool is slow or restricted, people will paste contracts into personal chatbots. Give them an approved path.

Forgetting post-signature value. Extraction of renewal dates and obligations is where AI is most reliable, per the benchmark evidence in Chapter 4, and it is often left unused.

9. The verdict

AI contract review is good at reading and getting better at writing. Independent evidence says it matches or beats unaided lawyers at pulling information out of documents and still trails them at redlining. Hallucination rates in purpose-built legal tools are lower than in general chatbots, but not zero. The tools that work best are the ones applying a playbook your team wrote, inside Word, with a lawyer reading the output before it goes out. Without a playbook, you are paying for speed on positions nobody chose.

Frequently asked questions

Is AI contract review the same as CLM?

No. CLM manages the whole contract lifecycle, from intake and approval to signature, storage and renewal. AI contract review covers the negotiation step and some extraction. Many CLM suites now include AI review, so you may get both from one vendor, but you can also buy review on its own.

Can a company without a lawyer rely on AI contract review?

It can use AI to understand a contract and prepare questions, which is useful. It should not treat the output as legal advice on contracts that carry meaningful liability. Stanford’s research found even purpose-built legal AI tools wrong on more than one in six benchmark queries, and no one on a non-legal team is positioned to catch those errors.

Will the vendor train its AI on my contracts?

It depends on the vendor and the plan. LegalOn and Spellbook say customer data is not used for training, and Anthropic and OpenAI say their business tiers do not train on customer content by default. Get the commitment in the contract or data processing addendum, and check retention separately.

Do lawyers need client consent to use these tools?

It depends on the tool and the jurisdiction. ABA Formal Opinion 512 grounds the question in the confidentiality duty, and Florida’s Opinion 24-1 recommends informed consent when a third-party generative AI tool would disclose confidential information. Check your state bar’s guidance and your engagement terms.

How accurate is AI redlining?

In the Vals benchmark, the best AI tool scored 65.0% on redlining against 79.7% for lawyers working without AI. Accuracy on your contracts depends on your playbook, document formatting, and contract types, which is why a pilot on your own paper matters more than any published figure.

Can I just use ChatGPT or Copilot instead?

For summaries, explanations and comparisons, a business-tier general assistant is often enough and costs less. For consistent playbook redlining across a team, dedicated tools are built for it: the playbook is shared, edits arrive as tracked changes, and each edit is tied to a rule.

Sources and notes

1. Source: Hendrycks, Burns, Chen and Ball, “CUAD: An Expert-Annotated NLP Dataset for Legal Contract Review,” arXiv

   .06268, accessed Oct 2026. ↩

2. Source: Juro, “Pricing”, accessed Oct 2026. Feature names and pricing basis as described by the company. ↩ ↩² ↩³ ↩⁴ ↩⁵

3. Source: Ironclad, “Introducing Contract Review Agent for Microsoft Word”, June 16, 2026, accessed Oct 2026. Capabilities as described by the company. ↩ ↩² ↩³ ↩⁴ ↩⁵

4. Source: World Commerce & Contracting, “Contract Management Whitepaper, August 2025”, accessed Oct 2026. ↩

5. Source: LegalOn, “Pricing”, accessed Oct 2026. Security and training statements are the company’s own. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

6. Source: Ironclad Support, “Ironclad Jurist for Microsoft Word”, accessed Oct 2026. The accuracy figure is from Ironclad’s own testers. ↩ ↩² ↩³

7. Source: Spellbook, “Pricing”, accessed Oct 2026. Security and training statements are the company’s own. ↩ ↩² ↩³ ↩⁴ ↩⁵

8. Source: Vals AI, “Vals Legal AI Report,” February 27, 2025, accessed Oct 2026. The report’s narrative and results table give different redlining scores for the best tool; the table figure is used here. ↩ ↩² ↩³ ↩⁴

9. Source: Stanford Law School, Dahl, Magesh, Suzgun and Ho, “Large Legal Fictions: Profiling Legal Hallucinations in Large Language Models,” 16 J. Legal Analysis 64 (2024), accessed Oct 2026. ↩

10. Source: Stanford HAI, “AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries”, May 2024, accessed Oct 2026. Over 200 pre-registered legal queries. ↩ ↩²

11. Source: ABA Journal, “Lawyers who ‘doubled down’ and defended ChatGPT’s fake cases must pay $5K, judge says”, accessed Oct 2026. ↩

12. Source: AICPA & CIMA, “SOC 2”, accessed Oct 2026. ↩

13. Source: Anthropic Privacy Center, “Is my data used for model training?”, accessed Oct 2026. ↩

14. Source: OpenAI Help Center, “Privacy Center in ChatGPT”, accessed Oct 2026. Confirm current terms with OpenAI before procurement. ↩ ↩²

15. Source: The Florida Bar News, “ABA issues first ethics guidance on a lawyer’s use of AI tools”, July 31, 2024, accessed Oct 2026. Read the full opinion before relying on it. ↩ ↩²

16. Source: The Florida Bar, “Ethics Opinion 24-1”, January 19, 2024, accessed Oct 2026. ↩

17. Source: State Bar of California, “Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law”, accessed Oct 2026. The State Bar revised this guidance in 2026; check the current version. ↩

18. Source: Gavel, “Gavel Exec Pricing”, accessed Oct 2026. List prices. ↩ ↩² ↩³

19. Source: Microsoft, “Microsoft 365 Copilot Plans and Pricing”, accessed Oct 2026. List prices; promotional terms per Microsoft. ↩ ↩² ↩³

20. Source: Microsoft, “Microsoft Copilot Plans and Pricing, Enterprise”, accessed Oct 2026. Requires a qualifying Microsoft 365 license. ↩