CSCounsel Sync HQ
Guides › Legal AI

A Legal AI RFP That Separates Vendors: Questions, Scoring Weights and Red Lines

Most legal AI proposals look identical until you ask questions that force a yes or no. Here's the question set, a 100-point scoring model, and the contract terms to hold firm on.

Counsel Sync HQ editorial team · November 2026 · 8 min read

Why Vendor Answers Usually Can't Be Compared

You've shortlisted three or four legal AI tools, and each one claims to be "enterprise-grade," "secure by design," and "doesn't train on your data." The proposals come back and still can't be compared, because every vendor answered a slightly different question. A good legal AI RFP fixes that. Each question has a constrained answer format, each answer maps to a score, and a short list of contract terms decides who gets cut before scoring starts.

Use the questions below as written or adapt them, require answers in the format shown, and score every vendor the same way. The method suits a 15-lawyer firm buying a drafting assistant and a global legal department replacing a contract review platform. Only the weights change.

Advertisement

Define Use Cases and Data Classes Before Writing Questions

Vendors can only give comparable answers if you describe the job the same way to all of them. Before you issue anything, write down two things.

  1. Three to five named use cases, each with a volume estimate. For example: first-pass NDA review (say, 60 a month), summarizing deposition or hearing transcripts, research memos in two jurisdictions, and clause extraction across a 2,000-contract legacy portfolio. These figures are illustrative.
  2. The data classes the tool will touch. Privileged client communications, personal data, special-category or sensitive personal data, export-controlled material, and anything a client has placed under contractual handling restrictions, such as outside counsel guidelines.

Data classes drive residency and retention requirements. If a client's guidelines bar processing outside the EU, or a public-sector client in the Gulf requires in-country hosting, treat that as a knockout criterion, not a scored preference.

Training Data and Retention: Questions That Force a Yes or No

"We don't train on customer data" is the most common answer in legal AI vendor due diligence, and it's often incomplete. Break it into parts and require a yes or no plus the contract clause reference for each:

  • Is customer input, output or feedback (including thumbs-up/down ratings) used to train, fine-tune or evaluate any model, whether yours or a third party's?
  • Do the underlying model providers retain prompts for abuse monitoring? If so, for how many days, and has a zero-retention arrangement been agreed?
  • How long are prompts, outputs, uploaded documents and embeddings retained by default? Can you set retention to zero, and does deletion reach backups and vector indexes?
  • Can vendor staff access customer content? Under what approval process, and is that access logged and visible to you?
  • Is customer data logically separated per tenant, and are retrieval indexes ever shared across customers?

A vendor that answers "no training" but stays vague on feedback data or embeddings should score lower than one that commits to all five in writing.

Residency, Sub-Processors and the Security Questionnaire

Your legal AI security questionnaire should ask where data is stored and where it's processed. They often differ, since inference may run in a different region from storage. Ask for a region-by-region table covering storage, inference, backups and support access.

Requirements vary widely. The EU and UK GDPR regimes require a lawful transfer mechanism for data leaving the region. India's DPDP Act lets the government restrict transfers to specified countries. Saudi Arabia and the UAE attach their own conditions to cross-border transfers. Brazil's LGPD, South Africa's POPIA and Singapore's PDPA each set separate transfer rules. Have local data-protection counsel confirm what applies to your client base, and don't rely on the vendor's summary.

On sub-processors, require:

  • A complete current list, including model providers, hosting, logging and support tools, with location and function for each
  • Advance notice of changes (30 days is a common ask), plus a right to object and terminate without penalty
  • Current independent attestations (ISO/IEC 27001, SOC 2 Type II, and ISO/IEC 42001 where available), with report date and scope stated
  • Breach notification within a fixed window, ideally short enough for you to meet your own regulatory deadlines, which differ by jurisdiction

Accuracy Evidence: Test on Your Documents, Not Their Demo

Accuracy claims in proposals are close to useless because each vendor measures them differently. Replace them with a structured pilot. Give every shortlisted vendor the same set of 20 to 50 representative documents, anonymized or consented, plus a short answer key prepared by your own lawyers.

Ask each vendor to describe its evaluation method in writing: which benchmark it uses, how it handles citations, and what the tool does when it can't find support for an answer. Then score the pilot on:

  • Correctness against your answer key
  • Citation fidelity: does every assertion point to the right passage or authority?
  • Refusal behavior: does it say "not found" rather than invent an answer?
  • Consistency: run the same prompt three times and compare

Hypothetical example: say a 40-lawyer firm pilots two NDA review tools. Tool A flags 92% of the issues in the answer key but cites the wrong clause about 1 time in 10. Tool B flags 85% with accurate citations every time. For most firms, Tool B is the safer choice, because lawyers can verify it quickly. With Tool A, every output needs a full re-check.

Model Changes and Version Control

Legal AI products swap underlying models often, and a change can quietly shift output quality on work you've already validated. Ask:

  • Which foundation models power each feature today, and from which providers?
  • How much notice will you get before a model change? Is there release documentation, and can you pin a version for a period?
  • Does the vendor re-run regression tests on legal tasks before switching, and will it share the results?
  • Are outputs logged with the model version, so a disputed work product can be traced?

If you have EU operations or clients, also ask how the vendor is preparing for its obligations under the EU AI Act and how it will support yours as a deployer. A vendor that can't name its own models won't be able to support an audit.

Pricing Questions That Expose the Three-Year Cost

Ask every vendor to price the same scenario: your actual user count, your stated volumes, and three years. Then have them break the quote into the components below. Figures are indicative 2026 market ranges that vary considerably by contract, region and volume. Confirm with vendors.

Cost componentTypical 2026 range (USD)What to ask
AI add-on to an existing practice or document platform$40–$150 per user/monthWhich features are included, and which are metered?
Specialist legal AI platform$150–$500+ per user/monthMinimum seat count and annual commitment
Usage or credit overagesHighly variablePrice per unit, alerts and hard caps
Implementation and integration$5,000–$75,000 one-timeFixed fee or time and materials?
Renewal upliftCommonly single-digit percentages a year if uncappedWill you cap it in the contract?

Ask whether pricing differs by region and currency, too. Some vendors charge more for in-region hosting in the EU, India or the Gulf.

A Weighted Scoring Model You Can Adapt

Run knockout gates first: required residency, no training on customer data, and acceptable sub-processors. Score only the vendors that pass. Each criterion gets a 0–4 score (0 = no answer or refused, 2 = meets the requirement with caveats, 4 = meets it in writing with evidence), multiplied by its weight to give a total out of 100.

CriterionLaw firm weightIn-house weight
Accuracy on your pilot set25%20%
Data handling and confidentiality25%20%
Security attestations and residency fit15%15%
Integration (DMS, CLM, Microsoft 365)10%20%
Model change governance10%10%
Three-year commercial terms10%10%
Vendor viability and support5%5%

In-house teams weight integration higher because adoption depends on the tool sitting inside contract and matter workflows. Firms weight confidentiality higher because privilege obligations run to many clients at once. Have at least two scorers, a lawyer and someone from IT or security, score independently before comparing results.

Contract Red Lines: Terms to Refuse to Sign Without

For most buyers, these legal AI contract terms are non-negotiable. If a vendor won't agree to them, it shouldn't win, whatever its score.

  1. No training, fine-tuning or evaluation on customer content or feedback, binding on sub-processors too.
  2. Customer ownership of inputs and outputs, with no broad licence back to the vendor beyond what's needed to deliver the service.
  3. Configurable retention and certified deletion, covering backups and embeddings, within a fixed period after termination.
  4. Sub-processor change notice with a right to terminate and a pro-rata refund.
  5. Confidentiality obligations that recognize privilege, plus a commitment to notify you of any legal demand for your data unless the law prohibits it.
  6. Liability caps that carve out data breaches and confidentiality breaches, or set a separate, higher cap for them.
  7. Renewal price caps and an exit-assistance clause that returns your data in a usable format.

Have your procurement or technology counsel review the final drafts against your regulatory and client obligations. Your next step: send the question set to your shortlist with a 15-business-day deadline and a required answer template, and drop any vendor that skips a knockout question.

Frequently asked questions

How many vendors should receive a legal AI RFP?

Three to five is the practical range. With fewer than three, you lose negotiating leverage. With more than five, structured pilots on your own documents become too time-consuming for the lawyers grading them.

How long should the evaluation take from RFP to signature?

For most firms and legal departments, 8 to 12 weeks is realistic: about three weeks for responses, three to four weeks of piloting, and two to four weeks of contract negotiation. Cross-border residency reviews or client consent requirements can add time.

Can we use our standard IT security questionnaire instead of an AI-specific one?

Use it as a base, but it won't cover training data use, model provider retention, embeddings, version changes or output accuracy. Add those as a separate AI module so every vendor answers them in the same format.

Should we pilot with real client documents?

Only if client consent, your engagement terms and the vendor's pilot contract allow it, and only after the data-handling terms are signed. Many buyers use anonymized documents or internal precedents for the first round. Get advice from your own counsel on what your client obligations permit in each jurisdiction.

Written by the Counsel Sync HQ editorial team. Prices, features and rules change often - confirm details with the vendor or a licensed professional before you decide.