CLM buying
CLM Vendor Security Questionnaire: The Questions to Send and the Answers to Accept
A contract management system ends up holding the most concentrated set of commercial secrets your company has: prices, liability caps, customer names, side letters, termination rights and every concession you ever made. Most security questionnaires that procurement teams send were written for HR tools or file storage. They ask about firewalls and password length, then miss the questions that matter for AI-driven contract software, such as whether your contracts feed a model and who else reads the text when the tool drafts a clause. This page gives you a CLM vendor security questionnaire organised into ten groups. For each group you get the questions to send, why the group matters for contract data in particular, and what a satisfactory answer looks like in practice. Near the end, one vendor's real answers are worked through as an example, including the places where the honest answer is "not yet". ## Classify what the system will hold before you write a single question The right depth of review depends on what goes into the tool. A team that will upload signed NDAs and simple order forms has a different risk profile from one that will store M&A disclosure schedules or health-data processing terms. Spend thirty minutes listing the contract families that will live in the system and tag each one: | Contract family | Typical sensitive content | Review depth | |---|---|---| | NDAs, simple vendor forms | Counterparty names, scope of confidential information | Standard | | Customer MSAs and order forms | Pricing, discounts, liability caps, service credits | Elevated | | Employment and contractor paper | Compensation, personal data, restrictive covenants | Elevated | | Deal documents, financing, disputes | Valuation, undisclosed liabilities, settlement terms | Highest | If any row lands in "Highest", the exit, deletion and audit-log groups below stop being nice-to-haves. Write the classification into the questionnaire cover note so the vendor understands why you are asking. ## Group 1: Does our contract text train anyone's model? Questions to send 1. Do you use customer documents, prompts or outputs to train, fine-tune or evaluate any model, yours or a third party's? 2. Is that position in the contract, or only in a marketing page or FAQ? 3. If you use aggregated or de-identified data for product analytics, describe exactly what is collected. 4. Can an administrator switch off any optional data sharing, and is the default off? Why it matters. A clause your team negotiated for weeks becomes a pattern another customer can benefit from if it trains a shared model. For contract data, the training question is the first question, not a footnote. What a good answer looks like. A plain "no" for customer contract content, written into the order form, MSA or a data processing addendum, with a narrow and specific description of any telemetry. A weak answer says "we take privacy seriously" and points to a settings toggle that is on by default. ## Group 2: Which model provider sees the text, and on what terms Most AI contract tools call a large language model hosted by someone else. That is not a problem in itself; it simply moves part of your risk to a second company. Questions to send - Which model provider or providers process our documents, and for which features?
- What terms govern that provider's use of our data, and can we read them?
- Is any prompt or output retained by the provider, for how long, and for what purpose (for example abuse monitoring)?
- If you switch model providers, will you notify us in advance? Good answer. A named provider, a link to the provider's commercial terms, and a clear statement on retention. For example, Anthropic's commercial terms state that Anthropic may not train models on customer content from its services. A vendor building on such terms should be able to point you to the clause rather than paraphrase it. ## Group 3: Where the data lives and who can move it Questions to send - In which country and cloud region are documents stored at rest, and where are they processed?
- Is single-region residency available, on which plan, and does it cover backups and AI processing?
- Do support staff in other countries have access to customer documents? Under what approval? US buyers often skip residency because their data is already in the US. That is a mistake when a vendor's engineering or support team sits in another country, because remote access is a transfer for many purposes. Ask where people who can open your documents actually work. ## Group 4: Encryption that covers the copies people forget Questions to send - Is data encrypted in transit and at rest, and with what standards?
- Are search indexes, extracted clause data, embeddings and backups encrypted too?
- Who holds the keys? Is customer-managed key support offered, and on which plan? Contract tools do not just store the PDF. They create extracted text, clause-level records, search indexes and, in AI products, vector representations of your text. A good answer confirms that all of these derived copies are covered, not only the original file. ## Group 5: Access, SSO and offboarding Questions to send - Do you support SSO through SAML or OIDC? On which plan?
- Do you support role-based access control, and can access be restricted to specific matters, folders or contract types?
- How fast does access end when a user is removed from our identity provider?
- How is your own staff's access to production data controlled and logged? The commercial wrinkle here is that SSO is often reserved for top plans. Price it in from the start. A team that buys a lower tier and adds SSO later can face a larger jump than expected. ## Group 6: An audit trail you can export Questions to send - What events are logged: logins, views, downloads, edits, shares, permission changes, AI requests?
- How long are logs kept, and can we export them to our own logging system?
- Are signature events recorded with time, IP and signer identity? Logs matter most on the day something goes wrong: a leaked draft, a disputed edit, an employee who downloaded the clause library before resigning. If the vendor can only show logs on a screen and not export them, your incident response depends on their support queue. ## Group 7: The sub-processor list, read line by line Ask for the full list, not a summary. For each entry you want the company name, the service it provides, the data it touches and its location. Then check three things: 1. The model provider from Group 2 appears on the list.
- Email, analytics and support-ticket tools that might receive document snippets are listed.
- The vendor commits to give notice of new sub-processors and a way to object. An incomplete list is the most common finding in CLM reviews, and it usually reflects a fast-growing vendor rather than bad intent. It still needs fixing before signature. ## Group 8: Retention, deletion and the backup question Questions to send - Can we set retention periods by contract type or workspace?
- When we delete a document, what happens to extracted data, indexes and AI history?
- How long until deleted data leaves backups?
- Will you certify deletion in writing at the end of the contract? A satisfactory answer gives a number of days for backup expiry and confirms that derived data is deleted with the source. ## Group 9: Incidents, notice periods and who calls whom Questions to send - What is your contractual notice period for a security incident affecting our data?
- Who is our named security contact, and what is the escalation path out of hours?
- Have you had a reportable incident in the last three years? What changed afterwards? The notice period belongs in the contract, not in a policy page the vendor can edit. Pair it with your own state breach-notification obligations so the vendor's clock is shorter than yours. ## Group 10: Attestations, and how to read them honestly Questions to send - Do you hold a SOC 2 report? Type 1 or Type 2, what period, which trust services criteria, and may we read it under NDA?
- Do you hold an ISO 27001 certificate? What is the scope statement?
- Have you published a CSA CAIQ or STAR entry?
- If you hold none of these, what is your plan and date? SOC 2 reports are attestation reports issued by a CPA firm under the AICPA's SOC framework; a Type 2 report tests whether controls operated over a period, while a Type 1 looks at design at a point in time. The Cloud Security Alliance STAR registry lets cloud providers publish a self-assessment questionnaire (the CAIQ) that you can read without asking. Read the scope section first: a report that covers the hosting provider but not the vendor's own application tells you less than its logo suggests. ## Exit and export: the question most buyers ask too late Ask this before signing, not at renewal: - In what formats can we export documents, metadata, clause data and audit logs?
- Is export self-serve, or does it require a support ticket and fee?
- Will exported documents keep numbering, formatting and version history?
- How long do we have after termination to export before deletion? A good answer is self-serve export to common formats (PDF and Word for documents, CSV or JSON for metadata) with a stated post-termination window. ## Turning answers into a decision Grade each group as Met, Partly met or Not met, then apply two rules. First, any "Not met" in Group 1 (training) or Group 7 (sub-processors) blocks the purchase until fixed in the contract. Second, gaps in Groups 5, 6 and 10 can be accepted for lower-sensitivity contract families if you record the decision, the owner and a date to revisit. For more on how these answers feed a formal evaluation, see our CLM RFP checklist for US buyers, and for the contract language that should carry the vendor's promises, see SOC 2 requirements in vendor contracts. ## Worked example: Adira's answers, including the gaps To show what this looks like with a real vendor, here is how Adira answers the groups above based on its published position. Adira is an AI contract lifecycle tool built on Anthropic's Claude. | Group | Adira's answer | Read it as | |---|---|---| | Training | Does not train on the customer contract corpus | Met; ask for it in the contract | | Model provider | Built on Anthropic's Claude | Ask for the sub-processor entry and terms | | Residency | Data residency controls on Enterprise | Met on Enterprise only | | Encryption | Encryption in transit and at rest | Ask about derived data and backups | | Access | Role-based access on all plans; SSO/SAML on Enterprise | Price Enterprise if SSO is mandatory | | Audit logs | Audit logs on Enterprise | Confirm export format | | Retention | Retention controls on Enterprise | Confirm backup expiry in days | | Attestations | Its security page sets out its certification position | Ask Adira for current status, scope and dates | | Export | One-click export to PDF and Word | Ask about metadata export | The honest reading: Adira's answers on training and core encryption are clear, while the heavier governance controls sit on the Enterprise plan, and a buyer who requires an audited report should ask Adira for its current attestation status in writing before shortlisting. A smaller team with mostly standard and elevated contract families can run its own check during the 7-day free trial using the questions above, then put the answers that matter into the order form. ## A one-page version to paste into your cover email 1. Do you train on our documents, prompts or outputs? Where is that in the contract?
- Which model provider sees our text, under which terms, with what retention?
- Where is data stored and processed, and who can access it from where?
- Are derived copies (indexes, clause data, embeddings, backups) encrypted?
- SSO, RBAC and offboarding speed: which plan?
- Which events are logged, for how long, and can we export them?
- Full sub-processor list with locations and notice of changes.
- Retention settings, deletion of derived data, backup expiry in days.
- Incident notice period and named contact.
- SOC 2 or ISO 27001 status with scope and dates, or the plan.
- Export formats and the post-termination window. Send the eleven questions with your contract classification table and ask for answers in writing. The quality of the reply, more than its speed, tells you how the vendor will behave when something goes wrong.
Frequently asked questions
- Is a generic SaaS security questionnaire enough for a CLM purchase?
- Usually not. Generic questionnaires cover encryption and access well but rarely ask whether customer contracts train an AI model, which model provider sees the text, or how clause data is deleted at exit. Add the AI and sub-processor questions on this page to whatever standard form your security team already uses.
- What should I do if a CLM vendor has no SOC 2 report yet?
- Do not treat it as an automatic fail for a small team, but do not accept a vague promise either. Ask for the controls in writing, a named security contact, the sub-processor list, and a dated plan for any audit. Then decide whether the sensitivity of your contracts allows a vendor at that stage, and record the decision.
- Who should own the questionnaire inside a US company?
- Legal owns the business case and the data classification, security owns the technical review, and procurement owns the paper. The common failure is each group assuming another one asked the AI training question. Name one person who signs off on the AI and data answers specifically.
- Should the AI model provider be listed as a sub-processor?
- Yes. If a vendor sends contract text to a third-party model provider to generate drafts or reviews, that provider handles your data and belongs on the sub-processor list with its location and the terms that govern it. A vendor that leaves it off is either running its own model or has an incomplete list, and you should find out which.
- How long should a CLM vendor take to return a security questionnaire?
- A vendor with its answers already written can return a focused questionnaire within days. Long delays on basic questions such as encryption, sub-processors and deletion usually mean the answers do not exist yet, which is useful information in itself.
Sources
See how Adira drafts in your voice and reads contracts from your side.
Explore the showroomWorking through a contract like this? Weave is Adira’s free tool to read, mark up, and connect any contract in your browser — no account needed.
Try Weave — free