How to Write a Request for Proposal for an AI Project

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Write a Request for Proposal for an AI Project.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for How to Write a Request for Proposal for an AI Project.

Describe the problem, not the solution: what happens today in numbers, what a good result looks like, your systems and data, your budget range and your deadline. Ask every supplier the same questions in the same format, give them two to three weeks and a named contact, and say in advance how you'll score the replies.

The most common RFP mistake on AI projects is naming the technology instead of the problem. Ask for "an AI chatbot" and you'll get three proposals for three different chatbots, priced on three different assumptions, which can't be compared. Ask how suppliers would cut the twelve hours a week your office spends answering the same member questions, and the proposals will argue about the thing you actually care about.

Follow me on Instagram@sagnikteaches

Is a formal RFP worth it for a job this size?

An RFP (request for proposal) is a document you send to several suppliers asking each to propose how they'd solve a problem, and at what price, in a format you set. For a small, well-defined job, it's more process than you need. A one-page brief sent to two or three consultants does the same work in a fraction of the time; where to find a good AI consultant includes one you can adapt.

Connect on LinkedInSagnik Bhattacharya

A fuller RFP, three to five pages, earns its time when any of these apply:

Subscribe on YouTube@codingliquids
  • A committee, board or trustees must approve the choice and will want to see a fair comparison.
  • The budget is large enough that a wrong choice would really hurt.
  • A funder requires competitive quotes.
  • The project will handle personal or sensitive data, so you need every supplier's answers on data in writing.
  • You're genuinely unsure which approach is right and want suppliers to propose different ones on equal terms.

If none applies, write the brief and skip the rest. If two or more apply, the RFP will save more time in arguments later than it costs to write.

Eight sections every AI project RFP needs

  1. About us. Who you are, your size, and who will work with the supplier. Two short paragraphs.
  2. The problem, in numbers. What happens today, how often, how long it takes, and what it costs or causes. This is the section suppliers read most closely.
  3. What success looks like. The measurable outcome you want and how you'll check it, written as acceptance criteria where possible. How to scope an AI project covers writing deliverables and acceptance criteria.
  4. Current systems and data. The tools you use, where the relevant data lives, its volume and format, and who holds the accounts.
  5. Constraints. Budget range, deadline, data-protection requirements, tools you must use or must avoid, accessibility needs.
  6. What we want in your response. The questions each supplier must answer, the pricing table they must fill in, and a page limit.
  7. How we'll evaluate. The criteria and their weights, stated before anyone replies.
  8. Process and timeline. The deadline for questions, the submission date, when you'll decide, and one named contact.

Section 7 is the one small organisations leave out, and it's the one that makes the rest fair. Suppliers write better proposals when they know what you'll reward, and your committee can't later be swayed by the best presentation if the weights are already on paper.

A members' club RFP, filled in

The club in this example is invented, but its RFP is one you could adapt directly. A sports and social members' club has about 650 members, a part-time secretary and a volunteer committee. Member questions arrive by email, phone and social media; court and room bookings run on a shared spreadsheet; renewals are chased by hand each spring. The committee's condensed RFP:

SectionWhat the club wrote
About usA members' club with about 650 members, one part-time secretary (20 hours a week) and a volunteer committee of eight. The secretary will be the day-to-day contact.
The problemAbout 12 hours a week of the secretary's and volunteers' time goes on routine member questions (fees, guest rules, opening times, booking changes) and on keeping the bookings spreadsheet up to date. Double bookings happen about twice a month. Renewal chasing takes about 30 hours each spring.
What success looks likeRoutine questions answered within an hour, including evenings, with at least half needing no human. Members can book and change bookings themselves. No double bookings. Renewal chasing under 10 hours. Checked over the eight weeks after launch against the same weeks last year.
Systems and dataMicrosoft 365 business plan (email, shared files), a membership spreadsheet of about 650 rows, the bookings spreadsheet, a website we can edit, and a WhatsApp group for announcements. All accounts are in the club's name.
ConstraintsPhase one budget $6,000 to $10,000; running costs under $200 a month. Live before the spring renewal season. Member data must stay in services with a data processing agreement. Anything members use must work on a phone.
Your responseAnswer the questions in the attached list, fill in the pricing table, and keep the proposal to eight pages.
EvaluationUnderstanding of our problem 25%, approach and phasing 20%, evidence of similar work 15%, data handling 15%, handover and support 10%, two-year cost 15%.
ProcessQuestions by email to the secretary by day 7; all answers shared with every supplier by day 10; proposals due by day 21; shortlisted suppliers meet the committee in week 4; decision at the committee meeting in week 5.

Notice what the RFP doesn't say. It never asks for a chatbot, an AI agent or a particular platform. One supplier might propose a booking system with a member FAQ assistant; another might propose an AI agent in the club's existing Microsoft tools. The RFP lets them compete on how well they solve the club's problem, with the same numbers in front of them.

Questions an AI project RFP must ask that others don't

A standard RFP asks about experience, method and price. An AI project needs a few questions more, because AI systems fail in particular ways and depend on services that change. Add these to the response section:

  • Which AI models and providers will the solution use, and under what data terms? You want the provider names, whether your data is used for training, and how long prompts and outputs are kept.
  • How will you test for wrong answers before launch? Ask for the number of test cases and whether they'll come from your real enquiries.
  • What happens when the AI doesn't know? The hand-off to a person, and how members are told they're talking to an AI.
  • What will it cost to run, month by month, at our volumes? Licences, usage fees and support, for 24 months.
  • What's your plan if a model or product you rely on is withdrawn? Real examples make this concrete: OpenAI's custom GPTs stop running on 11 December 2026, and the Sora API closed on 24 September 2026. Ask how quickly the supplier could move the solution to another model.
  • How will we know if it stops working? Automations fail quietly: Zapier pauses a Zap when 95% of its runs error over seven days, and Power Automate turns a flow off after 14 days of continuous failure. Ask what alerts you'll get.
  • What will we own at the end? Workflows, prompts, code, accounts and documentation. Who owns the AI workflows a consultant builds explains what to ask for.

Forcing comparable prices: the response table

Proposals written in prose hide their assumptions. A pricing table every supplier must fill in, in the same format, makes them comparable. The club's version, with one supplier's illustrative answers:

LineFixed priceDaysAssumptions
Discovery and design$9002Secretary available for two half-day sessions
Member question assistant$2,8006Club supplies current fees, rules and FAQs in writing
Self-service bookings$2,6005Uses an off-the-shelf booking tool, not a custom build
Renewal reminders$8002Membership spreadsheet cleaned by the club first
Testing with real cases$7001.560 past member questions supplied by the club
Documentation and handover$6001One handover session with the secretary and two volunteers
Running costs (per month, 24 months)$120n/aBooking tool, AI usage and email sending at 650 members
Rate for changes$450 a dayn/aAgreed in writing before work starts
Excludedn/an/aPayments integration; changes to the website design

The assumptions column matters as much as the prices. This supplier assumes the club cleans its membership spreadsheet and supplies 60 past questions. If another supplier's lower price assumes nothing, it may have left out the work, or it may plan to charge for it later as a change. Either way, you can now ask.

Drafting the RFP with an AI assistant, safely

An AI assistant can turn a committee's notes into a tidy RFP in minutes. Constrain it, or it'll fill the gaps with requirements you never had:

Turn the notes below into a request for proposal with these sections:
About us; The problem; What success looks like; Systems and data;
Constraints; Your response; Evaluation; Process and timeline.
Rules:
- Use only facts, figures and requirements in my notes.
- Describe problems and outcomes, not technologies. Don't name any
  AI product, model or platform as a requirement.
- If a section needs information my notes don't contain, write
  [TO CONFIRM: what's needed].
- Plain English, under 1,200 words.
Notes: [paste committee notes, figures and constraints]

Here's a passage from a first draft (illustrative):

Constraints: The solution must be available 24/7 with a guaranteed 99.9% uptime and must be built on GPT-5 or an equivalent leading model. The club's budget is approximately $15,000. [TO CONFIRM: live date]

Three problems in two sentences. The club never asked for 99.9% uptime, which would push suppliers towards costly setups. "Built on GPT-5" names a technology, the exact thing the prompt ruled out, and a model name that may be outdated by the time the work starts. And the $15,000 budget was invented; the notes said $6,000 to $10,000. The [TO CONFIRM] flag is the one part that worked as asked. Read every line of the draft against the notes and delete anything you can't trace back to them; the assistant saves typing, not thinking.

How many suppliers, how long, and who answers questions

Send the RFP to three to five suppliers. Fewer gives you too little to compare; more creates reading your committee won't finish, and wastes suppliers' time on a contest most of them can't win.

Allow two to three weeks for responses. A good proposal for a project this size takes a supplier a day or two to write, and the best consultants are usually busy. A one-week deadline tends to attract the least busy suppliers rather than the best.

Name one contact and set one deadline for questions. Answer every question in writing, and share the answers with all suppliers, without saying who asked, so everyone proposes on the same information. Ask committee members not to discuss the project with suppliers individually while the RFP is open. It's courteous to the suppliers, and it keeps the decision where it belongs.

For the final step, meet the top two. Better still, pay each for a short, fixed-price discovery session, which shows you how they work with your real information; how to run a paid discovery phase sets out how.

Scoring the proposals before you read the prices

Price has a way of colouring everything else once it's seen. A simple safeguard is to score quality first. Ask suppliers to put the pricing table in a separate file, have two committee members score the other five criteria independently, agree the quality scores, and only then open the prices.

Score price on the two-year cost (build plus 24 months of running costs), not the build alone. Give the lowest two-year cost the full 15 points, and each other supplier a share in proportion. The club's results (illustrative):

SupplierQuality score (out of 85)Build priceRunning cost a monthTwo-year costPrice score (out of 15)Total
A72$8,400$120$11,28015.087.0
B52$6,200$260$12,44013.665.6
C71$9,800$80$11,72014.485.4

Supplier B had the cheapest build and would have looked the obvious choice on a quote-by-quote comparison. Its running costs made it the most expensive over two years, and its proposal scored lowest on understanding the club's problem. A and C were close, which is when the paid discovery sessions earn their fee: the committee chose A after seeing how each handled the club's real member questions. Before signing, the committee went through the winning proposal with the checks in how to evaluate an AI implementation proposal.

Mistakes that turn an RFP into three incomparable pitches

Here's how it goes wrong without the structure above. A club committee emails five suppliers: "We're looking for an AI chatbot for our website. Please send a proposal and price." Replies range from $900 to $14,000. The cheapest is a template chatbot with no connection to bookings; the dearest is a custom app with a members' portal. Each answers a different question, and the committee spends two meetings arguing about which question it meant to ask. The same mistakes recur:

  • Naming a technology instead of a problem. Fix: describe the work and the outcome, as in the club's RFP.
  • No volumes. "Lots of emails" can't be priced. Fix: counts per week or month, from a two-week sample.
  • No budget range. Fix: give a range, and say whether it covers running costs.
  • No acceptance criteria. Fix: say how you'll check the result, so suppliers price the testing it needs.
  • Asking for free design work. Fix: ask for an approach, not a finished design, and pay the shortlist for discovery.
  • Forgetting running costs. Fix: the 24-month row in the pricing table.
  • Leaving out data and ownership questions. Fix: the AI-specific questions above.

Each fix costs a few minutes of writing. Together, they're the difference between choosing a supplier and choosing between guesses.

RFP questions from first-time buyers

Should we put our budget in the RFP?

Yes, as a range. Without one, suppliers guess, and you get proposals sized for budgets you don't have, which can't be compared. A range such as $6,000 to $10,000 for phase one lets each supplier propose what they'd do within it. It doesn't weaken your negotiating position much, because the pricing table still makes each supplier justify every line.

Can we ask suppliers for a small test before choosing?

Yes, but pay for it. Asking several suppliers to design your solution for free in their proposals is unfair and gets rushed work. Shortlist two from the written replies, then pay each for a short, fixed-price discovery session or a small test on your own data. What you learn about how they work is worth far more than the fee.

Do we have to choose the cheapest proposal?

Not unless a funder or your own rules require it, so check any grant conditions first. Most small organisations can choose on value, as long as the scoring criteria and weights were set out in the RFP and applied the same way to every proposal. Keep the score sheets; they explain the choice to your committee and to the suppliers who lost.

How should we tell unsuccessful suppliers?

Promptly, politely and with a reason. A short email thanking them, naming the criteria where their proposal scored lower, and offering a ten-minute call if they'd like more detail costs you little. Suppliers put real hours into a good proposal, and the ones treated well are more likely to respond next time you send an RFP.

Further reads

Sources: OpenAI help articles on custom GPT retirement and the Sora discontinuation; Zapier and Power Automate documentation on failed automations; general procurement practice for small organisations.

Want your AI project brief sharpened before it goes out?

On a 1:1 call we'll turn your notes into a problem statement with real numbers, set the scoring weights, and check the RFP asks the AI-specific questions that matter.

Book a 1:1 call with me