How to Write an RFP for AI Tools in a Sustainability or Resilience Program

Missouri reports that AI-assisted review of 3,367 agricultural grant applications in early 2026 took 70 staff hours and $3,458, including $720 of computing, against an estimated 3,367 hours and $122,693 by hand. Subject-matter experts kept the final scores (state award submission). Results like that are why document-heavy programs are buying AI, and those same programs are where an unchecked error ends up in a grant decision, a permit or a public report.

An AI tool for a sustainability or resilience program reads, extracts or drafts from documents such as plans, permits, grant applications, utility bills and supplier questionnaires, while a named person approves each output before anyone relies on it. You need one when a review queue outgrows the staff who own it. The RFP succeeds when it fixes the task, the test and the approver before anyone sees a demo.

The rules moved in 2025 and 2026. OMB issued new federal AI memos in April 2025 and a third in December (OMB), and NIST says its AI Risk Management Framework is being revised under a White House plan that tells it to drop references to climate change (NIST). New York, California, Texas and Maryland now attach terms to government AI use and purchases, and the language models inside these tools retire on schedules their makers set.

What this work should deliver

The outcome is a defined task done with a measured error rate, a named reviewer who approves every output, records you can produce when asked, and a way to switch the tool off and finish by hand.

Who this is for: sustainability, resilience and grants staff at local governments, state agencies, ports, universities and nonprofits with a queue of documents to read, such as hazard mitigation plans to check, grant applications to screen, utility bills to extract for an emissions inventory or a sustainability report to draft. It is the wrong purchase for general writing help, which seat licenses cover (see the budget section). And if the real problem is two spreadsheets that disagree, settle data ownership first, because AI will repeat the disagreement faster.

Settle these before you write tasks:

• The task and its volume. Lakewood, Washington defined its 2024 tool by the work: legal and code research behind about 2,000 permits a year (RFP). Name the document types, the fields or findings you need and the monthly volume.
• The approver. Lakewood required city staff to review outputs before any official use. Name the role, and the record that shows who approved which output.
• What the tool may touch. List the data it may read, such as public documents, personal information or confidential bids, and the actions it may never take, such as sending, publishing or changing a record.
• Decision weight. Federal agencies treat AI as ‘high-impact’ when its output is a principal basis for decisions with legal, material, binding or significant effects, and require pre-deployment testing, an impact assessment and human oversight (M-25-21). If your tool’s output will drive a decision about an applicant, a permit or a person, borrow that standard.
• Records. Decide where prompts and outputs live and for how long, before the vendor’s default retention decides for you.

Our one-workflow build brief is a template for writing the first of these down, and our sustainability data and technology hub collects related guides.

Rules and standards to cite

Most AI rules bind federal or state agencies, not counties, nonprofits or companies. Cite them anyway, by number and date, because they hold the most tested contract language available.

• Procurement route. With federal funds, 2 CFR 200.320 uses the FAR thresholds in force since October 1, 2025: up to $15,000 without quotes, up to $350,000 with quotes from an adequate number of qualified sources, and a formal, advertised solicitation above that (FAR 2.101). Documented cooperative arrangements meet its competition requirements (2 CFR 200.318); Williamson County, Texas bought ChatGPT Enterprise seats in 2025 through a Texas Department of Information Resources cooperative contract. State and local rules can be stricter, so check yours first.
• NIST. The AI Risk Management Framework 1.0, released January 26, 2023, is voluntary, and its Generative AI Profile, NIST AI 600-1, followed on July 26, 2024. NIST says the framework is being revised under America’s AI Action Plan, which directs it to remove references to misinformation, diversity, equity and inclusion, and climate change (AI Action Plan). Cite version 1.0 by date, so a revision can’t change what your contract requires.
• OMB memos. M-25-21 and M-25-22 (April 3, 2025) and M-26-04 (December 11, 2025) appear without rescission on OMB’s memoranda page as of October 2026, and they bind federal agencies only. M-25-22 says agency processes should ensure contracts permanently prohibit using non-public agency inputs and outputs to train commercial AI without explicit consent, and asks solicitations to guard against vendor lock-in through knowledge transfer, data and model portability, clear licensing and pricing transparency. M-26-04 tells agencies buying large language models to request the vendor’s acceptable use policy and its model, system or data cards.
• New York and California. New York’s Appendix C-AI (April 2025) requires 90 days’ written notice before a vendor brings AI into a product, keeps AI that isn’t integral to the system off by default, and makes prompts, uploaded content and outputs state property that may not be used to train or tune any model. California requires state entities to complete a generative AI risk assessment before awarding a purchase, consult its Department of Technology on moderate- and high-risk uses, and include a clause in IT contracts requiring vendors to notify the state in writing if they will deliver generative AI or use it in ways that materially affect the work (SAM 4986.9).
• Texas and Maryland. Texas requires state agencies and local governments to adopt the state’s AI code of ethics and minimum standards, including guidelines for contractually requiring vendors to use risk management frameworks on heightened-scrutiny systems (SB 1964, effective September 1, 2025). Since July 1, 2025, a Maryland state unit may not buy or deploy a new AI system unless it meets the Department of Information Technology’s policies (§ 3.5-805). The law reaches public universities except for research and academic uses (§ 3.5-802), and the department’s Responsible AI Policy of May 23, 2025 extends to contractors and calls for technical documentation such as the specific model used.
• Federal preemption. A December 2025 executive order set up a Justice Department task force to challenge state AI laws, but told the drafters of its legislative proposal not to propose preempting state laws on state government procurement and use of AI (EO 14365).
• Records. Washington State Archives says generative AI interactions about public business are public records, including those made on personal devices or accounts (guidance). NARA’s August 21, 2026 guidance lets federal agencies dispose of AI-related records only under an approved schedule (NARA).
• Accessibility and naming. A public-facing assistant is web content under the ADA Title II rule, with compliance from April 26, 2027, or April 26, 2028 for entities under 50,000 people and special districts (28 CFR 35.200). And under a September 29, 2026 executive order, new federal agency documents use ‘Super Intelligence’ where they used AI (EO 14434), so search for both terms.

The term templates miss is model change. Microsoft retires generally available models in its Foundry service 18 months after launch, closes them to new customers at 12 months and won’t extend retirement dates (Microsoft). OpenAI commits to at least six months’ notice and will shut down several GPT-4o and GPT-4.1 models on January 14, 2027 (OpenAI). A tool you accept in the spring may run on a different model by the next reporting cycle, so the contract has to require notice, a fresh acceptance test and a way back.

Sample scope of work

Adapt the bracketed items. Write Tasks 1 and 3 before you talk to vendors.

1. Task definition and data boundary. Everything else is tested against this memo.

The Contractor shall document the task, the document types and monthly volume in scope, the data classes the tool may process, the actions it may not take (including sending, publishing or changing any record) and the role that approves each output. Configuration shall not begin until the [Agency] approves this memo.

2. Proof of concept on your documents. California paid each of five vendors $1 to test generative AI in a six-month sandbox in 2024 (California), and M-25-22 tells federal agencies to keep their evaluation data out of the vendor’s reach.

Before full deployment, the Contractor shall run the tool on [200] of the [Agency’s] documents, including superseded versions, missing years and figures that conflict between sources. The [Agency] shall keep a separate test set that the Contractor shall not see before acceptance testing.

3. Accuracy and acceptance. California’s tax agency listed ‘result accuracy of at least 95 percent’ among the acceptance criteria for its proof of concept (agreement). Denver required bidders for its AI plan-review contract to define their expected accuracy threshold and explain how they handle false positives and false negatives (contract).

The [Agency] shall accept the tool only if, on the held-back test set, it [extracts the required fields with at least 95% accuracy, cites the source page for every finding and flags every seeded conflict]. The Contractor shall report false positives and false negatives separately and repeat the test after any change of model, prompt or data source.

4. Human review and audit trail. Lakewood’s RFP required staff review before official use, citations to sources in prompts and outputs, and retention under the state Public Records Act.

No output shall be used in an official capacity until a named [Agency] reviewer approves it. The tool shall record the reviewer, the decision, the output version and the sources cited, and shall export that record on request.

5. Data protection and model training. Vendor defaults differ: OpenAI says API data hasn’t trained its models since March 1, 2023, but keeps abuse-monitoring logs up to 30 days unless it approves zero retention (OpenAI). Denver bars using city data to develop AI models without written consent, then consents only for models in the city’s own instance, a workable alternative to a flat ban.

[Agency] data, prompts, uploaded content and outputs remain the [Agency’s] property. The Contractor and every subprocessor, including any model provider, shall not use them to train, retrain, tune or improve any model, shall process and store them in [the United States], shall keep them no longer than [30] days after processing unless the [Agency] directs otherwise, and shall disclose every subprocessor before award.

6. Model disclosure, changes and cost. New York requires 90 days’ written notice before a vendor introduces AI, and M-25-22 encourages federal agencies to require that a new model version meet performance standards before deployment or be rolled back.

The Contractor shall name each model and version the tool uses, give [90] days’ written notice before changing a model or adding AI features, rerun the acceptance test before any change goes live, and keep the accepted version available until the [Agency] accepts the change. The cost proposal shall separate the Contractor’s fees from model and API usage charges, state the assumed monthly volume, and cap usage charges at [amount] a month without written approval.

7. Security, hosting, integrations and records. Denver requires an annual SOC 2 Type 2 audit, quarterly vulnerability scans, a penetration test and 99.5% monthly uptime. Read the exclusions: its uptime guarantee doesn’t count outages of third-party components, so ask whether the model provider is one.

The Contractor shall maintain [99.5]% monthly availability and list every component excluded from that measure, provide a current SOC 2 Type 2 report each year, integrate with [the Agency’s single sign-on and document management system], retain prompts and outputs under [the state records retention schedule], and make any public-facing interface conform to WCAG 2.1 Level AA.

8. Handover and exit. California’s 2024 agreement gave the state a 30-day transition to extract its data, then required destruction within 30 days and a written statement confirming it.

On termination or expiration, the Contractor shall deliver within [30] days all [Agency] data, prompts, outputs, review records, configuration and test results in open, machine-readable formats, then delete them and certify the deletion in writing. The tool shall keep a switch that stops AI processing while staff finish the work by hand.

Deliverables to require

• Task and data boundary memo, approved before configuration.
• Test plan, held-back test set and proof-of-concept results.
• Acceptance report with accuracy by document type, counting false positives and false negatives separately.
• Reviewer workflow and an exportable audit log.
• Model and subprocessor list, with the vendor’s acceptable use policy and model or system cards, the documents M-26-04 asks federal buyers to request.
• Data protection terms covering location, retention and deletion.
• Security evidence: a current SOC 2 Type 2 report or GovRAMP status, and incident notice terms.
• A monthly usage and cost report: documents processed, model calls and charges.
• Records retention settings matched to your schedule, and an export procedure.
• Staff training, a manual fallback procedure and a tested off switch.
• An exit package of data, prompts, outputs, review records, configuration and test results.

How to evaluate proposals

Treat vendor accuracy claims as hypotheses. The winning vendor in Denver’s 2026 plan-review procurement claimed more than 99% accuracy for its code compliance model after a brief calibration with city reviewers, whose decisions override the system (Denver contract). Only a test on your held-back documents shows what that number means for your queue.

M-25-22 tells federal agencies to test proposed solutions before award to the greatest extent practicable, and M-25-21 lets them query a service and observe its outputs when they can’t see the model. With federal money, 2 CFR 200.320 requires published evaluation factors with their relative importance and an award on price and other factors. 2 CFR 200.319 excludes firms that drafted your requirements, which can include a vendor whose free pilot shaped them.

In interviews, ask each finalist to run part of your held-back set live and show you an output it got wrong.

Timeline

Plan the second acceptance test along with the first. Denver’s contract runs five years, from March 1, 2026 to March 1, 2031; by our arithmetic, that is more than three times the 18-month life of a generally available Foundry model. Budget staff time to repeat the test whenever the model changes.

Budget and cost drivers

Small tools are cheap to start. Lakewood capped implementation of its generative AI permit research tool at $30,000 from a state digital permitting grant, with monthly or annual fees after launch excluded (RFP). Missouri reports building its grant-review process in house for $3,458, including $720 of computing, and California’s six-month proofs of concept cost the state $1 per vendor.

Seat licenses buy a general tool, not a tested workflow. Williamson County, Texas bought 150 ChatGPT Enterprise seats in 2025 at $470.40 a seat a year against a $480 list price, $70,560 in all, under a 150-seat minimum. The quote excluded API use, which would be billed at OpenAI’s list prices, and added quarterly true-ups for new users (county agreement). Microsoft lists Microsoft 365 Copilot at $30 per user per month, paid yearly (Microsoft).

Platforms run larger. Denver’s AI plan-review contract allows up to $4,591,740 over five years; its first-year milestones total $353,060, and $120,000 of that, about a third, pays for integrations with the city’s identity, GIS and permitting systems (contract). South Carolina’s environmental agency posted a $1,777,000 sole-source notice in July 2026 to expand its permitting platform with AI-driven analysis modules (notice), and the Department of Energy’s PermitAI went from a $300,000 pilot to $10 million in new funding (PNNL).

For a custom build, one studio’s published range for an AI assistant or automation is $25,000 to $70,000 over six to twelve weeks (CF Labs estimator). Our budget brief for custom sustainability software lists the lines that come after launch, and our build-or-buy comparison covers when a subscription wins.

What moves the price:

• Volume. Usage charges scale with documents and model calls, and every new document type needs its own test cases.
• Integrations. A third of Denver’s first-year milestones went to connecting three city systems.
• The accuracy bar. A higher pass mark means more test documents, more tuning and more reviewer time.
• Data location and retention. OpenAI charges 10% more on its data residency endpoints for models released from March 5, 2026, and zero retention needs its approval (OpenAI).
• Model choice. Ask bidders to price the model they propose and one alternative, with the volume assumption shown.
• Repeat testing. Every model change needs a new acceptance run, and logging prompts, outputs and approvals for your full retention period adds storage and export work.

Red flags in proposals

• A demo run only on the vendor’s own documents.
• One accuracy number with no test set, no error types and no statement of how it was measured.
• ‘We never train on your data’ in the marketing but not in the contract, or silence on the model provider’s retention.
• No named model or version, and no notice before either changes.
• Usage billed ‘at cost’ with no volume assumption, cap or monthly report.
• A tool that can send, publish or change records without a person’s approval.
• An uptime guarantee that excludes the components the tool depends on.
• No off switch, and no manual procedure for finishing the work without the tool.

How we’d approach it

We would start with one task, one named reviewer and a test set the vendor never sees, then decide whether the tool has earned a second task. CF Labs builds AI for sustainability teams on that rule: tools that read, extract and draft, with a person accepting or rejecting every output, and client data never used to train models without consent. We hold our own product, Trovance, to it, and nothing it drafts publishes until a person approves it. Building these tools makes us stricter about acceptance tests, not looser; our AI pilot acceptance checklist sets out the tests we would put in the contract, and our assurance guide covers checks on figures you publish.