Blog AI company guides

AI RFP template: how to write an RFP for AI development.

What an RFP for AI development must ask that a software RFP doesn’t: evaluation on your own data, a proof stage, data use, model accounts, ownership and exit. With a scoring table and a complete template to copy.

Fig. 0Bids scored on cases the bidder never saw

An AI RFP needs everything a software RFP asks for (scope, integrations, team, plan, support) plus questions a software RFP never asks. It must ask each bidder how they would measure quality on your data and which acceptance numbers they would propose. It should include a proof stage, where shortlisted bidders run their approach on a sample of your data and are scored on cases they never see. And it must ask where your data goes and who may train on it, which models and accounts the system depends on, who owns the prompts, evaluation sets and any fine-tuned weights, and how you would leave.

Nobody can promise results on data they haven’t seen, so a good AI RFP asks for an approach and a way to prove it, not a fixed specification. Below: when to use an RFP, what each section should ask, how to run the proof stage and score responses, and a template to copy.

Do you need an RFP at all?

An RFP works when you can describe the problem, you know what data exists, and several firms could credibly do the work. It works badly when nobody has looked at the data or the problem is still vague. Every bid then rests on guesses, and the most confident guess can win.

In that case, start with a short discovery engagement: a few weeks on your data and workflow, ending in a written plan, a first evaluation set and a go or no-go recommendation. The UK government’s Guidelines for AI procurement (8 June 2020) say “Discovery phases or proofs-of-concepts can demonstrate if the AI system is likely to meet your wider requirements”. They also tell buyers not to expect “a fully detailed and definite plan, as AI development is an iterative process”.

A middle path is a phased RFP. The World Economic Forum’s AI Procurement in a Box: AI Government Procurement Guidelines (June 2020) describes a challenge-based process that evaluates competitors’ minimum viable products in phase one, “with only the winner going on to develop the full solution”. The proof stage in the template below is a small version of that.

Your situationWhat to use
Problem clear, data assessed, several capable firmsAn RFP with a proof stage
Problem clear, data not yet assessedA short discovery first, or an RFP whose first phase is the data assessment
Problem still vagueDiscovery with one or two firms, then decide whether to go to market
One firm already proven on similar work, and no rule requiring competitionA direct discovery engagement, then a statement of work

How is an AI RFP different from a software RFP?

A software RFP assumes that what is built as described will work. With AI, the output varies, quality depends on your data, and the model underneath changes on the provider’s schedule. NIST’s Generative AI Profile (NIST AI 600-1) names one risk, confabulation: “the production of confidently stated but erroneous or false content”.

TopicA software RFP asksAn AI RFP also asks
RequirementsFeatures and specificationsThe problem, the business measure, and what a correct outcome looks like on real cases
QualityA test plan and a bug processHow the bidder would measure quality on your data, and which acceptance numbers they would propose
ProofA demo and referencesResults on a sample of your data, scored on held-back cases
DataMigration and storageThe gaps the bidder sees in your data, and who may train models on it
DependenciesTech stack and licencesWhich models it calls, whose accounts they run in, and what happens when a model changes
SecurityAccess control and penetration testingHow it resists prompt injection, and what it can do if an attack works
OwnershipSource codePrompts, evaluation sets, scoring scripts and any fine-tuned weights as well
SupportUptime and bug fixesQuality monitoring after launch, and a retest on every model change

What should each section of an AI RFP ask?

Describe what you need in detail and leave the method open. The UK guidelines put it this way: “Use output-based requirements, which allow the supplier to propose how they will respond to your requirement.”

Background, problem and business measure

Describe the work today and the problem, with real numbers: volumes, task times, where errors happen. The WEF guidelines say: “Focus on developing a clear problem statement, rather than on detailing the specifications of a solution.” Then name the number that should move, today’s baseline and the target. “Resolve 60% of order-status questions without a person” gives bidders something to design for. “Implement an AI agent” doesn’t.

Scope and out of scope

List the workflows, users, channels, languages and actions in scope, then what isn’t in scope yet, so every bidder answers the same question.

Data available and constraints

Describe each data source, its format and size, and what you know is wrong with it. The WEF guidelines suggest you “Highlight known limitations (e.g. quality) of the data in the RFP and require tenderers to describe their strategies on how to address these shortcomings.” Add the constraints: personal data, where it may be processed, and what bidders may see before award. If you share a sample, the UK guidelines say to provide the same data to all suppliers at the same time, and suggest non-disclosure agreements and techniques such as masking or synthetic data.

Required integrations

List every system the AI must read from or write to, and how access works today. Ask which integrations carry the most risk and what the bidder needs from your team.

Evaluation and acceptance

Don’t fix the acceptance numbers alone before anyone has seen the data. Ask each bidder how they would build an evaluation set from your cases, what they would measure and which numbers they would propose, with reasons. The WEF guidelines say: “If possible, work with bidders to determine what the thresholds should be.” A bidder who answers with a measurement plan understands the work. One who names an accuracy figure for a model they haven’t tried on your data doesn’t. Evals before features shows what a good evaluation set looks like.

The proof stage

Say whether there will be one, who is invited (two or three shortlisted bidders keeps it manageable), how long it runs, whether it is paid, and how it is scored. OMB memo M-25-22, Driving Efficient Acquisition of Artificial Intelligence in Government (3 April 2025), which replaced the earlier memo M-24-18, sets out what US federal agencies should ask of AI suppliers. It says that when evaluating proposals, agencies must, “to the greatest extent practicable, test proposed solutions to understand the capabilities and limitations of any offered AI system or service.”

Security and prompt injection

Ask how the system treats outside text, which tools and permissions it gets, which actions need a person’s approval, and how attacks are tested before launch. NIST AI 600-1 describes indirect prompt injection as attackers “injecting prompts into data likely to be retrieved”, and notes researchers have shown it can steal proprietary data or run malicious code remotely. Prompt injection is an input-validation problem explains why permissions, not prompts, limit the damage.

Privacy and data use

Ask for every service that would see your data, where it is processed and stored, how long it is kept and how it is deleted. Ask directly whether the bidder or any model provider may train on your inputs or outputs. M-25-22 says agency processes should ensure contracts “permanently prohibit” using non-public agency data and outputs to further train publicly or commercially available AI, absent explicit agency consent. You can ask bidders to confirm the same.

Models and accounts

Ask for every model and service the system would depend on, and whose account each runs in. NIST AI 600-1 notes that “GAI value chains involve many third-party components such as procured datasets, pre-trained models, and software libraries.” Ask how model versions are pinned and how a provider’s model change is tested before it reaches users. Look for production keys held by you, with supplier access you can remove.

Ownership, lock-in and exit

Ask what you would own: code, prompts, tool definitions, evaluation sets, scoring scripts, cleaned data and any fine-tuned weights. Ask bidders to list the pre-existing tools they would reuse and the licence you would get. The UK guidelines tell buyers to consider who should own any new intellectual property, since a new algorithm is likely to be developed or an existing one tailored.

Then ask how you would run the system without the bidder, or move it to another firm. M-25-22 says agencies should generally include protections against lock-in in the solicitation itself, such as knowledge transfer, clear data and model portability practices and clear licensing terms. The WEF guidelines say: “Make knowledge transfer a requirement under the RFP.”

Team and production references

Ask for the named people who would do the work, their roles and how much of their week your project gets. The WEF guidelines suggest bidders “provide evidence of the skills and qualifications of the proposed project resources who will develop and deploy the AI solution”. Ask for a system live for at least six months, a reference who runs it, and an anonymised evaluation report and incident write-up.

Delivery plan and support after launch

Ask for phases tied to results, not only dates: discovery ending in a go or no-go decision, a build that ends when the held-back set passes, a pilot, then launch and handover. Ask who watches quality after launch and how model retirements are handled. The UK guidelines say “lifespan testing, not a one-time decision, is required”.

Response format and timetable

Fix a page limit, the order of answers and the dates. Ask for commercial terms in a separate document, so the panel scores the approach first.

How do you run the proof stage fairly?

A written response shows how a bidder thinks. A score on your data shows what they can do. Every bidder should face the same test, with no way to tune to it.

  • Build the test first. Collect real cases, each with the correct outcome signed off by someone who does the work today. Include awkward cases, refusals, requests for a person and a few attacks.
  • Split it. Give every shortlisted bidder the same development sample at the same time, and keep the rest back. M-25-22 says data used for independent evaluations “should not be accessible to the vendor, and should be as similar as possible to the data used when the system is deployed”.
  • Run the held-back cases yourself. If a bidder runs them through its own system, it has seen them. Ask for something you can run in your own environment. If that isn’t possible, treat the held-back set as used once the proof stage ends, and build a fresh one for acceptance.
  • Score everyone the same way. Same cases and runs, scored by a script or by reviewers who don’t know whose answer they are reading.
  • Keep it small and fixed. One time limit and scope for everyone. Say whether it is paid and who keeps what: a fair split is that bidders keep their methods, and you keep your data, cases and scores.
  • Read the failures, not only the score. A bidder whose errors are refusals and handovers is safer than one with fewer errors that are confident and wrong.
  • Close it cleanly. Bidders who aren’t chosen delete your data and confirm it in writing.

How do you score AI RFP responses?

Score every response on the same criteria, with weights agreed before you open the first one. Score each criterion from 0 to 5, multiply by the weight, and note the reason. Commercial terms are evaluated separately, once you know which bidders can do the work.

CriterionWeightWhat a strong response shows
Understanding of the problem10Restates your problem and measure, names the risks in your data, says which parts don’t need AI.
Evaluation approach and proposed acceptance20A concrete plan to build and score an evaluation set from your cases, with proposed numbers and reasons.
Proof stage results20Scores on your held-back cases by category, with failures that are safe rather than confident and wrong.
Security and agent controls10Outside text treated as untrusted, least privilege, approval for irreversible actions, attack testing.
Data handling and privacy10Every service that sees your data listed, no training on it, retention and deletion stated.
Models, ownership and exit10Your accounts, pinned models, a full ownership list and a handover plan your team could use.
Team and production references10Named builders, a system live for six months, and a reference who runs it.
Delivery and support after launch10Phases tied to results, a go or no-go point, monitoring and retesting after launch.

Set pass marks as well as weights. A 0 or 1 in evaluation, security or data handling should rule a bidder out for anything customer-facing, whatever the total. If you skip the proof stage, move its weight to evaluation approach and references. The WEF guidelines say technical, business, legal and ethical experts are needed to score the bids, so build the panel that way.

What are the red flags in an AI RFP response?

  • An accuracy promise before seeing your data.
  • No evaluation plan. “We will test it thoroughly” instead of a test set and numbers.
  • Pushback on held-back cases. They want to see every question before the test.
  • Security means the system prompt. A prompt can’t be the only control.
  • Keys stay with the bidder. Accounts move to you only “on request”.
  • Ownership covers software only. No mention of prompts, evaluation sets or weights.
  • An unnamed team, or demo-only references. No named builders, and nothing with dates or numbers still running.

The AI RFP template

Copy it, replace every [square bracket], and delete what doesn’t apply. Send it with your standard terms and a confidentiality agreement.

AI RFP template
AI REQUEST FOR PROPOSAL

Project: [project name]
Issued by: [company name] ("Buyer")
Questions to: [name, email]
Responses due: [date, time]

1. BACKGROUND AND PROBLEM
1.1 How the work is done today: [process, people, tools, volumes]
1.2 The problem: [one or two sentences]
1.3 Why AI might help: [reason]. We are open to approaches that use less AI or none.

2. BUSINESS MEASURE
Measure: [measure]. Baseline [value]. Target [value] by [date].

3. SCOPE
3.1 In scope: [workflows, users, channels, languages]
3.2 Actions the system may take: [list, with limits]
3.3 Out of scope: [e.g. a second channel, write access to the CRM]

4. DATA AVAILABLE AND CONSTRAINTS
4.1 Data sources: [source, format, size, date range]
4.2 Known issues: [gaps, errors, out-of-date content]
4.3 Personal data: [categories, where it may be processed, retention rules]
4.4 What bidders may see before award: [nothing, a masked sample, or a synthetic sample]
4.5 Describe how you would assess our data and deal with the issues in 4.2.

5. REQUIRED INTEGRATIONS
5.1 Systems to read from and write to: [system, access method, actions]
5.2 Describe the integration risks you see and what you need from our team.

6. EVALUATION AND ACCEPTANCE
6.1 Describe how you would build an evaluation set from our cases: size, coverage, and who signs off the expected outcomes.
6.2 Propose acceptance criteria as numbers, with reasons, covering:
    a. Task pass rate
    b. Critical failures, and how you define them
    c. Factual answers supported by a cited source
    d. Saying so or handing over when the answer is not in the data
    e. Required handovers to a person
    f. Response time
    g. Prompt-injection and misuse cases
    h. The lowest score allowed in any one category
6.3 Describe how cases are scored and how many runs count.
6.4 Part of the evaluation set is held back by the Buyer and never shared. Confirm you accept acceptance testing on it.
6.5 Final acceptance numbers are agreed in the statement of work after discovery.

7. PROOF STAGE
7.1 Shortlisted bidders: up to [3]. Duration: [x] working days, [dates]. Paid or unpaid: [state which].
7.2 Every shortlisted bidder receives the same development sample of [x] cases at the same time.
7.3 Scoring uses [y] held-back cases that bidders never see, run by the Buyer [in its own environment, or through an endpoint the bidder provides].
7.4 Scored on [measures from 6.2] by [a script, or reviewers who do not know which bidder produced each answer].
7.5 Bidders keep their methods. The Buyer keeps its data, the cases and the scores. Bidders not selected delete all Buyer data within [x] days and confirm in writing.

8. SECURITY AND PROMPT INJECTION
8.1 List every entry point for untrusted text, and the tools and permissions the system would have.
8.2 Which actions need a person's approval, and how is that enforced in code?
8.3 How do you test prompt injection and misuse before launch and after each change?
8.4 If an attack succeeds, what is the worst the system could do?

9. PRIVACY AND DATA USE
9.1 List every service that would see our data, where it is processed and stored, and for how long.
9.2 Confirm that neither you nor any model provider will train models on our inputs or outputs without our written consent.
9.3 Describe how our data is returned or deleted at the end, with written confirmation.

10. MODELS AND ACCOUNTS
10.1 List every model and third-party service the system would depend on, with version and purpose, and whose account each runs in.
10.2 Our preference: production accounts and keys held by the Buyer, with access granted to you.
10.3 How are model versions pinned, and how is a model change tested before it reaches users?

11. OWNERSHIP AND IP
11.1 Confirm the Buyer would own: source code, prompts, tool definitions, evaluation sets, scoring scripts, configuration, cleaned or labelled data, and fine-tuned weights or adapters.
11.2 List any pre-existing tools you would reuse, who owns them, and the licence the Buyer would get.
11.3 Name any part of the system the Buyer could not export or run without you.

12. LOCK-IN AND EXIT
12.1 Describe the handover package: documentation, runbooks, architecture decisions, and how to run the evaluation set.
12.2 Knowledge transfer: [x] hours with our team or a new supplier. Describe your plan.

13. TEAM AND PRODUCTION REFERENCES
13.1 Named people who would do the work, their roles, and how much of their week goes to this project. Any subcontractors.
13.2 At least [2] AI systems you built that have run in production for six months or more, and a reference contact for one.
13.3 With client details removed: an evaluation report across several releases, and an incident write-up.

14. DELIVERY PLAN
14.1 Phases tied to results: discovery ending in a written go or no-go recommendation; build ending when the held-back set passes; pilot; launch and handover.
14.2 Timeline for each phase, and what you need from our team in each.
14.3 What happens if scores fall short of the acceptance numbers.

15. SUPPORT AFTER LAUNCH
15.1 Who monitors quality after go-live, which measures, and how often the evaluation set reruns.
15.2 How model updates and retirements are handled.
15.3 Incident response times by severity, and the fallback when a connected system is down.

16. RESPONSE FORMAT AND TIMETABLE
16.1 Answer sections 4 to 15 in order, using the same numbering. Limit: [x] pages, plus annexes.
16.2 Commercial terms: [submitted separately, in the format in Annex A]
16.3 Questions by [date]. Answers shared with all bidders by [date]. Shortlist [date]. Proof stage [dates]. Decision [date].
16.4 Responses are scored on the criteria and weights in Annex B. Commercial terms are evaluated separately.
16.5 This RFP is not an offer to contract. Confidentiality: [terms]

Annex A: [commercial response format]
Annex B: [evaluation criteria and weights]
Annex C: [data description, and the sample if shared]

What happens after you choose a supplier?

Turn the winning response into a statement of work with the acceptance numbers agreed during discovery, the held-back set, data terms, accounts, ownership and exit. AI statement of work: what to include, with a template covers each section and follows on from this template. Before you shortlist, use questions to ask an AI development company as your interview questions.

This guide is not legal advice: a lawyer should review the RFP terms, the confidentiality agreement and the final contract, including the law that applies to you.

For how we scope AI work, including the discovery phase and the evaluation set we agree before building, see our services and AI consulting.

Questions buyers ask

What should an RFP for AI development include?

The problem and business measure, scope, data and its limits, integrations, and questions on evaluation, security, privacy, models, ownership, exit, team, delivery and support. Add a proof stage on your data and a fixed response format.

How is an AI RFP different from a software RFP?

A software RFP asks for features and trusts that what is built will work. An AI RFP asks how quality will be measured on your data, tests the shortlist on cases they never see, and covers training on your data, model dependencies, prompt injection and ownership of prompts and evaluation sets.

Should an AI RFP include a proof of concept?

Yes, for anything that matters. Keep it small and identical for every shortlisted bidder, score it on held-back cases, and say in advance whether it is paid and who keeps what.

Who should set the acceptance criteria, us or the bidders?

Both. Ask bidders to propose numbers with reasons, then agree the final numbers in the statement of work once someone has looked at the data. Keep the held-back cases and the definition of a critical failure in your hands.