Skip to content

Example Everything on this site is sample data — one of the configurations this template ships with, built from the same repository.

Example: AI use case catalog
Civic AI Community of Practice AI Use Case Catalog

Pilot · State agency — Procurement division

Vendor product

Clause summaries for procurement review

Result in one line: Cut first-pass contract review from five business days to one

A licensed product that extracts clauses from scanned vendor agreements, summarizes them in plain language, and flags where they differ from our standard terms or are missing entirely.

Published June 2, 2026 Verified July 30, 2026

Documentation or write-up
Skills needed to set it up
Contractor or vendor
Readiness
Needs a paid license Needs a data agreement Human review built in
Data it touches
Internal, non-public data
Who sees the output
Internal staff
License
Not open source — description only
Cost to stand up
$25k–$100k
Portable to other platforms
No — tied to its platform
Use case category
Administrative & task automation
Reviews it went through
Legal or contracts review Security review or authority to operate Records retention review
What do these mean?
Contractor or vendor
Outside help is required to stand this up.
Needs a paid license
Depends on a paid product, API or subscription.
Needs a data agreement
A data-sharing or BAA-type agreement is required.
Human review built in
A person checks the AI's output before it is used.
Internal, non-public data
Non-public operational data without personal identifiers.
Internal staff
Used only inside the organization.
Not open source — description only
The write-up is what is shared, not the artifact itself.
No — tied to its platform
Depends on one vendor's services end to end.
Administrative & task automation
Forms, intake, scheduling, records — the routine work that eats staff time.
Security review or authority to operate
A security assessment, ATO or equivalent sign-off.
Records retention review
Confirmed how long the outputs must be kept and where.

Problem

Our team of three reviews around 200 vendor agreements a year. Many arrive as scanned PDFs from vendors who printed our template, wrote on it and scanned it back. The first pass — reading the whole document, finding the clauses that differ from our standard terms, and noticing the clauses that are not there at all — took a specialist most of a week per contract.

The missing clauses were the expensive ones. A clause that differs is visible; a clause that was quietly deleted is invisible unless someone is reading with a checklist in hand, and those were the gaps that used to surface after signature.

What we built

We licensed a commercial product instead of building anything, and we are writing it up because the evaluation is the reusable part.

The tool reads a scanned agreement, extracts clauses, writes a plain-language summary of each, and compares them against a library of our own standard terms that we uploaded during setup. The output is a review worksheet: one row per clause, with the page it came from, a summary, and a flag saying whether it matches our template, differs from it, or is absent.

How it works

Scanned pages go through the vendor’s OCR first, which is why output quality tracks scan quality closely. The comparison is against our own clause library, so the tool is only as good as the standard terms you give it — normalising ours took longer than the vendor’s implementation did.

Nothing is decided by the tool. A contracts specialist opens every flagged clause in the source document before it goes anywhere. What changed is what gets read first, not what gets approved.

Results

Across 38 contracts in the pilot, first-pass review dropped from a median of five business days to one. The tool found an average of two missing standard clauses per contract — usually records retention or right to audit — which is the number that persuaded our leadership.

Accuracy on badly scanned documents was noticeably worse: roughly one in six contracts needed a page rescanned before the extraction was usable. Two contracts in the pilot produced summaries that misstated a liability cap, both caught in review, both from pages where the OCR had merged two columns.

Lessons learned

Budget for the standard terms library. Ours existed as a Word document carrying twelve years of tracked comments, and normalising it was three weeks of work we would have had to do eventually anyway. That library is now used in places the tool is not.

Ask the vendor where documents are processed, what is retained, and whether anything is used for training — in writing, before the pilot, not during contract negotiation. Our question set is linked above; it is nine questions and it took one meeting.

Do not put live negotiations in. A draft under active negotiation is the one document where a leak has an immediate cost, and there was no version of the pilot that needed them.

How to reuse

This is a paid product on someone else’s infrastructure, so the path is procurement rather than deployment. We bought off another agency’s cooperative contract, which removed about four months from the timeline. The evaluation memo is the document we used to make the case internally and is probably more reusable than anything else here.

About

Area of work
  • Finance, procurement & contracts
  • Legal & compliance
Review status
Reviewed & approved

How it's built

How AI is involved
AI is part of the solution
Types of AI
  • Document Q&A (RAG)
  • Classification & NLP
  • Computer vision
AI tools & models
  • Vendor-hosted clause extraction service
  • Optical character recognition (vendor-hosted)
Where it runs
  • Vendor / SaaS hosted
Vendor or partner
Commercial contract-analytics vendor

Sharing & licensing

Access terms
This is a commercial product; what we are sharing is the write-up. The evaluation memo, the question set we put to the vendor and the clause comparison template are all linked above and are free to reuse. Contact us if you want to talk through the pricing or the contract terms we negotiated.
Portability notes
The product is the vendor's hosted service end to end. What transfers is the evaluation approach and the clause taxonomy, not software. If the vendor disappeared tomorrow we would keep the taxonomy and lose the tool.

What it took

Cost to keep running
Over $50k/yr
How it was bought
  • Cooperative or piggyback contract
Who it affects
Internal to procurement, with no output that reaches the public. The thing we watch is whether the tool flags unfamiliar contract language more often for small and minority-owned vendors, whose paperwork looks least like the large vendors whose agreements the product has seen most of. A flag is only ever a prompt for a specialist to read the clause; it is never a score, and nothing about a bidder is derived from it.

Data & access

No PII/PHI in the shared material
Yes
Data sources
  • Executed and draft vendor agreements
  • Standard terms and conditions library
Data-governance caveats
Agreements are uploaded to the vendor's tenant under a data processing addendum that fixes the processing location, sets a 90-day retention limit and prohibits training on our documents. Draft agreements under active negotiation are excluded from the pilot entirely. No agreement text appears in this entry or its screenshot.

Printed from AI Use Case Catalog
https://crypticpy.github.io/phct/examples/ai-use-cases/catalog/procurement-clause-review/