Legal AI RFP template for 2026: questions that produce comparable answers
Most legal software RFPs run to eighty questions and produce eighty marketing paragraphs. The set below is short on purpose. Each question has a reason, and each has a note on the answer shape that should worry you, because an RFP is only useful if a weak answer is visible as one.
Ask every vendor the same questions in the same order, and ask for the same defined scope rather than their preferred bundle. Comparability is the entire point; without it you are collecting brochures.
Commercial
| Ask | Why | A weak answer |
|---|---|---|
| What is the billing unit, and what is the rate? | The unit does more to the bill than the rate does. | "It depends on your requirements" without naming the unit. |
| Quote three years at our stated volumes, including all increases. | Year one is not the purchase; year three is. | A year-one figure only. |
| What is the annual escalator? | 5 to 10% is common and compounds. Negotiable now, not at renewal. | "Standard uplift" with no number. |
| What is the implementation fee, and is it included? | Frequently the larger half of year one where charged separately. | Silence, or a range with no scope attached. |
| What is in the base licence and what is a module? | AI capability is separately priced by several vendors here. | A feature list that does not distinguish. |
| What does it cost to get our data out, and in what format? | Exit terms have no leverage after signature. | "We'd help you with that." |
Capability
| Ask | Why | A weak answer |
|---|---|---|
| Run this on our five worst documents and show us the output. | Quality on bad inputs is where tools actually differ. | Declining, or substituting their own sample. |
| Does it produce real tracked changes, or a comment describing one? | Decides whether counterparties can accept the output. | "It integrates with Word." |
| What does it do when the model is unsure? | An uncertain answer presented like a confident one is the real risk. | "Our accuracy is very high." |
| Which specific step does the AI perform without a human? | Cuts through "AI-native" and "agentic". | An architecture diagram. |
| Does it work on third-party paper as well as our templates? | Most tools are much stronger on your own. | "Yes" with no demonstration. |
| Show the clause the answer relied on. | An uncited answer cannot be checked. | An answer with no source. |
Evidence
| Ask | Why | A weak answer |
|---|---|---|
| Name a customer at our size, in our sector, live in the time you have quoted. | Timelines are quoted from best cases. | A logo wall. |
| What was measured in any accuracy claim, on whose documents, against whose judgement? | "140% more accurate" means nothing without a denominator. | A number with no methodology. |
| How many hours of our people's time does implementation assume, and from whom? | This is the cost that never appears on the invoice. | "Minimal." |
| What is the most common reason your deployments underdeliver? | A vendor who cannot answer this has not looked. | "None come to mind." |
Data and security
| Ask | Why | A weak answer |
|---|---|---|
| Do you train any model on our documents? Is that the default? | The answer differs across vendors and sometimes across their own tiers. | "Your data is secure," which answers a different question. |
| Where is our data processed and stored, and will you commit to it contractually? | Eliminates otherwise strong options in regulated and public sector buying. | A region named verbally but not in the contract. |
| Which sub-processors touch our documents? | The list is usually longer than expected. | A refusal, or "industry-standard providers". |
| What certifications do you hold, and what is their scope? | Scope matters as much as the badge: one covering a corporate office and not the product platform tells you little. | A badge with no scope statement. |
| What is the retention period, and can we require deletion? | Rarely asked, and it constrains what you can promise your own customers. | "We keep it as long as you're a customer." |
How to score it
- Weight the three-year total and the worst-documents test above everything else. Those two predict satisfaction better than any feature matrix.
- Treat a non-answer as a negative rather than a blank. Vendors answer what suits them and omit the rest, and an RFP that scores only what was said rewards that.
- Score the questions you asked, not the material they volunteered.
- Keep the answers. At renewal they are the only record of what was promised, and the people who made the promises will have moved on.
This template is free to copy and adapt. It is deliberately short: a shorter RFP that vendors answer properly beats a long one they answer selectively.