AI Agent RFP: How to Write a Brief Vendors Can Quote

An AI agent RFP (request for proposal) is a short document that tells vendors exactly which job the agent will do, which systems it touches, how you will judge it and what you need to own at the end, so their quotes can be compared line by line. Most businesses skip it or write one page that says "we want an AI agent for customer service". The result is five quotes between $4,000 and $90,000 that describe five different projects. A good brief is four to six pages, takes a day to write, and is the cheapest risk control you will buy in the whole project.

This post is part of our guide to rolling out agents, which starts at wizeb.com/blog/how-to-implement-ai-agents-in-business. Once the brief is out, wizeb.com/blog/how-to-choose-an-ai-agent-development-company covers how to judge the replies.

The Problem: Vague Briefs Produce Quotes You Cannot Compare

When the brief is vague, every vendor fills the gaps with their own assumptions. One quotes a prompt on top of a no-code tool. One quotes a custom build with integrations to three systems. One quotes a six-month "AI transformation". None of them is wrong, and none of them can be compared. You end up choosing on price or on how much you liked the sales call, which is how projects end up in the pile Gartner warned about: it predicts over 40% of agentic AI projects will be cancelled by the end of 2027 because of rising costs, unclear business value or weak risk controls (gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027).

All three of those causes can be addressed on paper before anyone writes code. Unclear value is fixed by naming the job and the baseline. Rising cost is fixed by asking for running costs at your volume, not just the build fee. Weak risk control is fixed by stating what the agent may and may not do. That is what an RFP is for.

What a Good AI Agent RFP Contains: 10 Sections

1. 1The business problem, in numbers. What happens today, how often, and what it costs. "We answer 1,200 supplier invoice queries a month; two people spend about 60% of their week on them" beats "we want to improve efficiency".
2. 2The one job the agent does. A single workflow with a clear start and end. If you have several in mind, pick the first and list the rest as later phases. Our guide at wizeb.com/blog/first-ai-agent-for-business helps you choose.
3. 3Inputs and outputs. What arrives (emails, form submissions, calls, PDFs) and what the agent produces (a reply, a CRM update, a booked slot, a draft for approval). Attach 10–20 anonymised real examples. Vendors quote far more accurately from real samples than from descriptions.
4. 4Systems and access. Every tool the agent reads from or writes to, with the plan or edition you are on (HubSpot Professional, Xero, Zendesk, a Postgres database, a legacy portal with no API). Say who owns each one and whether a test environment exists.
5. 5Rules and limits. What the agent may decide alone, what needs a human sign-off, money or refund limits, and the conditions for handing over to a person.
6. 6Volume and speed. Requests per day and at peak, the response time you need, and the hours it must run. This drives the running cost more than anything else.
7. 7Data and compliance. Where the data may be stored and processed (US, UK, EU), whether personal or regulated data is involved, and which rules apply to you, such as GDPR. Ask how the vendor handles model providers' data retention.
8. 8The pass mark. How you will decide it works before it goes live: for example, "correct on 90% of a 200-case test set drawn from last quarter, with zero wrong refunds". Without this, "done" means whatever the vendor says it means.
9. 9Ownership and exit. You own the prompts, code, workflows, test sets and accounts. Ask for handover documentation and how a different team would take it over.
10. 10What you want back. A fixed price for the build broken into phases, a monthly running cost at your stated volume, a timeline, the team who will actually do the work, and one similar project they can walk you through.

What to Leave Out

Do not specify the model or framework ("must use GPT-5 and LangChain") unless you have a hard reason, such as an existing contract or an internal platform. You are buying an outcome; naming the tools invites vendors to match the spec instead of solving the problem, and it locks you to choices that may be out of date in six months. Ask instead how they choose and how easily the model can be swapped.

Leave out the 40-question security questionnaire at this stage, too. Send three questions now (where data is stored, who can access production, how actions are logged) and the full questionnaire to the two finalists. Long first-round forms mostly filter out small, capable teams and keep the ones with a bid department.

How to Run the Process

1. 1Send it to three to five vendors, not fifteen. More replies do not give you a better choice; they give you less time to read each one properly.
2. 2Offer a 30-minute call to each before they quote. The questions they ask tell you more than the proposal: a good team asks about your exceptions, your data quality and who will own the agent after launch.
3. 3Give two weeks for replies and a fixed format. Ask everyone to price the same phases so you can line them up in one table.
4. 4Score replies on fit, not polish. Weight understanding of the problem, the plan for testing and handover, running cost at your volume, and the people named. A slick deck with no test plan should score lower than a plain document with one.
5. 5Pay for a short proof of concept with the top one or two. A two- to four-week paid test against your pass mark is the best way to choose. Our guide at wizeb.com/blog/ai-agent-proof-of-concept covers what to test and what it should cost.

Red Flags in the Replies

• A price with no running cost. Model usage, hosting and monitoring are real monthly costs; a vendor who leaves them out is either inexperienced or hoping you will not ask.
• No test plan. If the proposal does not say how it will be tested against your examples before launch, assume it will not be.
• Accuracy promised up front. Nobody can promise 98% before seeing your data. A credible vendor proposes a pass mark and a way to measure it.
• Platform lock-in hidden in the terms. Watch for prompts, workflows or accounts that live in the vendor's own platform and leave with them.
• A team that changes after signing. Ask who will do the work, and put the names in the contract if it matters to you. Our guide at wizeb.com/blog/outsource-ai-agent-development covers the contract terms in more detail.

What the Quotes Should Look Like

With a clear brief for one job, replies should land in a narrow band. A single-task agent connected to one or two systems, with a draft-and-approve stage before it acts alone, typically quotes at about $5,000 to $15,000 and three to six weeks. Several systems, actions it takes without approval, or a phone channel push it into the $15,000 to $50,000 range. If your quotes differ by more than three times, the vendors understood different projects, and the brief needs another pass. The line-by-line breakdown is in our cost guide at wizeb.com/blog/ai-agent-cost-guide.

Example: A 60-Person Distributor

This is an illustrative example, but the pattern is common. A US building-supplies distributor with 60 staff wants an agent to handle order-status and delivery-date emails from trade customers. Its first brief is one paragraph. It gets four quotes: $6,000, $18,000, $42,000 and $85,000, plus one vendor who wants a paid "discovery phase" before quoting at all.

They rewrite the brief using the ten sections. It now says the agent reads the shared sales inbox (Microsoft 365), checks orders in their ERP (read-only, via an existing API) and carrier tracking, drafts replies for approval for the first month, and must be correct on 90% of 250 past emails. Volume is about 900 emails a month, and everything is hosted in the US. They attach 20 anonymised examples.

The second round of three quotes lands between $11,000 and $16,000, each with a monthly running cost of $300 to $600, and two of the three vendors ask the same question about how backorders are recorded in the ERP. That question becomes the deciding factor: the vendor who proposes testing it in a two-week paid proof of concept wins the work. The brief took one day to rewrite and turned a guessing game into a like-for-like comparison.

How WizeB Responds to an AI Agent RFP

WizeB builds custom AI agents, workflow automations and voice agents for growing businesses in the US, UK and Europe. When we get a brief, we reply with a fixed price per phase, a monthly running cost at your stated volume, the people who will build it, and a test plan scored against your own examples. If the brief is missing something that changes the price, we say so and tell you what to add, rather than padding the quote. You own the prompts, workflows, test sets and accounts from day one. See how we work at wizeb.com/services/ai-agents.