What 17 legal-aid RFPs in 2026 asked vendors about AI
By MV2 of Munim, Inc. This study and this post were made with AI agents working in my name. They describe public procurement documents and are not legal advice. Legal-aid organisations hire outside help for many jobs: i
By MV2 of Munim, Inc. This study and this post were made with AI agents working in my name. They describe public procurement documents and are not legal advice.
Legal-aid organisations hire outside help for many jobs: intake systems, websites, evaluations and strategic plans. Their requests for proposals (RFPs) are public. I wanted to know what those documents say about AI in 2026. Do they tell the vendor how it may use AI on the job? And how often do they ask for AI as part of the work?
I pre-registered four predictions before reading any text. I published the coding rules, then froze the coding itself and published it as well. Only after that did the grader run.
What was counted
- Frame: RFPs and RFIs posted from 1 January to 4 November 2026 on LSNTAP's jobs-and-RFPs board (live pages, archive captures and files the board still hosts), plus a few postings that web search found on the same site.
- 17 postings could be coded. Nine more survive only as one-line teasers, so they could not be coded.
- Checked mechanically. The grader downloaded all 18 source files again, matched every SHA-256, and found all 34 quotations word for word in the PDF text.
What came out
| Prediction | Result |
|---|---|
| At most 30% have a clause on the vendor's own AI use | held: 3 of 17 (17.6%) |
| That share is higher in July to November than in January to June | failed: 2 of 9, then 1 of 8 |
| None bans vendor AI use | held |
| At least 25% of build-the-software postings include AI in the work | held: 5 of 8 (62.5%) |
The second prediction needs a caveat up front. The gap is one posting. If one borderline call is read the other way, it would have held. With only search-found postings removed, it can't be tested at all. Read it as "no sign of a rise", not as a decline.
What the clauses say
- Disclosure plus human review. Two of the three vendor-AI postings come from one programme, and both carry the same standard clause: "Any use of AI in processing data under this contract must be disclosed, and any deliverables assisted by AI must be reviewed by a human being before finalization." Both also ask for the policies that will govern the vendor's use of AI.
- Data rules on the product, not the vendor. Several build RFPs put data limits on the delivered system instead. One says the system "must exclude client personally identifiable information from training data". Another says "all proposed solutions" must follow "prohibitions on using customer data to train AI models". I coded those as rules about the product. I list them as borderline calls because a reader could reasonably see them as rules on the vendor.
- More AI in the work than rules about it. 9 of the 17 postings want AI in the work itself: an AI research assistant, AI intake and referral tools, AI-generated video tutorials, a conceptual design for an AI knowledge base, and training staff to build Copilot agents.
Limits
The sample is small, and one coder made every call. That coder was AI agents working in my name, with no second human coder. Every call is published with its quotation and with the reading that would flip it, so you can check each one and disagree. Legal-aid RFPs posted only on an organisation's own website are not in the frame. Munim, Inc. may respond to one of the open postings in this set. That is part of why the coding was frozen and published before the grader ran.
Data, code and full write-up (CC BY 4.0 / Apache-2.0): https://github.com/munimtechnologies/bar-ai-guidance/blob/main/analysis/legal-aid-rfp-ai.md
MV2, Munim, Inc., [email protected]
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes โ full credit and traffic to the original publisher.