AI Agent RFP Template for Buyers (With Red-Flag Answers)

An AI agent RFP template should describe a workflow, a data set, a success metric, and a way out. It should not describe a technology.

Most results for this topic are about agents that answer RFPs. This one is for the buyer who has to write one.

Below is a skeleton you can paste into your own document, a weighted scoring sheet, and a list of answers that should make you nervous.

Why a Generic Software RFP Fails

A standard software RFP asks for a feature list, a price, and references. Agent projects need more, because agents act. They read your data, call your systems, and sometimes change records.

The cost of getting this wrong is real. Gartner describes “agent washing,” where vendors rebrand chatbots, RPA, and assistants as agents, and estimates only about 130 of the thousands of agentic AI vendors are real. Cost, value, and risk are the three topics an RFP has to pin down.

So the RFP has two jobs. It has to filter out products that are not really agents. And it has to force every vendor to commit to the same measurable outcome.

One more thing. Buying can work. Buying blind does not, so make every vendor show evidence before they make your shortlist.

Describe the Workflow, Not the Technology

This is the section that decides everything else. Do not write “we want an AI agent for customer support.” Write the workflow.

The World Economic Forum’s procurement guidelines for the private sector (June 2023) say to define business outcome criteria and to document as-is processes and the gaps the AI solution should close. I agree. If you cannot describe today’s process, no vendor can improve it.

Here is the scope section I’d write. Fill in every line.

AI agent RFP template: scope section

1. Workflow name and owner. One workflow. One accountable person.
2. Current process. Steps, systems touched, who does each step, and how long it takes today.
3. Volume. Cases per week, with seasonal peaks. Say whether this is a measurement or an estimate.
4. Inputs. What triggers the workflow and in what format: email, PDF, ticket, form.
5. Outputs. What “done” looks like. A drafted reply, a posted entry, a routed ticket.
6. Decisions the agent may make alone, and decisions that need a person. List both.
7. Exceptions. The ten messiest cases you know of. Ask vendors to explain what the agent does with them.
8. Out of scope. Name the adjacent workflows you are not buying yet.

If you are still choosing the workflow, start with which workflows to automate first. One workflow in an RFP is a strength, not a limit.

Data, Access, and Security Sections

Agents need access, and access is where risk lives. Ask for specifics in writing.

Data questions to include:

• Which of our systems will the agent read, and which will it write to? Give us the permission list.
• Where does our data go? Name every region, sub-processor, and model provider.
• Is our data used to train any model, yours or a third party’s? Under what contract term?
• What is stored, for how long, and how is it deleted on request?
• Can the agent run in our cloud account or on our infrastructure? What changes if it does?

The data-use question has a precedent. The U.S. Office of Management and Budget memo M-25-22, issued in April 2025 for federal AI purchases, tells agencies to ensure contracts permanently prohibit using non-public agency data and outputs to further train publicly or commercially available AI algorithms without explicit consent. You are not a federal agency, but the clause is a good one to copy.

Security and control questions:

• What is the least-privilege permission set the agent needs, and how is it enforced?
• Which actions require human approval before they run?
• How can we stop the agent immediately?
• What gets logged for each run, and can we export it?
• Which standards do you hold? ISO/IEC 42001, published in December 2023, is the AI management system standard. SOC 2 or ISO 27001 covers the wider security program. Ask for the report, not a badge.

The OWASP list of LLM risks calls out excessive agency: too much functionality, too many permissions, or too much autonomy. Its main mitigation is to require a human to approve high-impact actions. Make vendors tell you how they handle that.

For a longer set of questions, see questions to ask an AI agent vendor.

Success Metric and Pilot Gate

Every vendor must answer the same metric question. Pick one hero metric, set a baseline, and ask vendors to commit to a target.

Metric block to include:

• Hero metric. For example, handling time per case, cost per case, or percent of cases completed without rework.
• Baseline. Measured on your data before the pilot. Not the vendor’s benchmark.
• Target and floor. The result that means “go,” and the result that means “stop.”
• Quality guardrail. An error rate the agent must stay under while it hits the speed target.

Then add a pilot gate. This is a short, scoped run with pass or fail criteria agreed before it starts. AI POC vs pilot vs production explains the stages. A gate is useful because it turns a vendor claim into your data.

My suggested gate language:

1. The pilot runs on your real data, in your environment or an agreed sandbox.
2. The test set is defined by you. OMB M-25-22 makes a related point for government buyers: data used in independent evaluations should not be accessible to the vendor and should resemble deployment data. Hold back a slice the vendor never sees.
3. Pass means the hero metric reaches the target and the quality guardrail holds.
4. A fail ends the contract at the pilot price, with no commitment to the next phase.
5. A pass triggers a written production plan with named owners, not an automatic purchase.

The length and size of the pilot depend on your workflow. I’d keep it short and scoped. A pilot that runs for months has stopped being a test.

Exit and Ownership Clauses

Write the way out before you sign the way in. The same OMB memo lists protections against vendor lock-in: knowledge transfer, data and model portability, rights to the code and models produced under the contract, and transparency in licensing and pricing. It also tells agencies, once they decide not to extend a contract, to work with the vendor on ongoing rights and access to data and derived products.

Turn that into RFP asks:

• Ownership. Who owns prompts, workflows, evaluation sets, fine-tuned artifacts, and logs? Say it per item.
• Portability. Can we export our data, configurations, and run history in a standard format?
• Knowledge transfer. Will you document the system well enough for another team to run it?
• Pricing transparency. What are the usage, model, and support costs, and what can change after year one?
• Termination. What is the notice period, the transition support, and the data deletion timeline?
• Third-party dependencies. What happens if an underlying model provider changes terms or prices?

AI agent ownership and exit rights goes deeper on each clause.

Scoring Sheet

Score vendors on the same scale, in writing, before you watch demos. Demos bias you.

Here is a sheet to start from. The weights are my suggestion. Change them to fit your risk.

Category	Weight	What earns a 5	What earns a 1
Workflow fit	20%	Explains your exceptions and handles them in the pilot plan	Generic answer, no exception handling
Pilot gate and metric	20%	Accepts your metric, baseline, and fail terms	Refuses a fail clause or redefines the metric
Data and security	20%	Clear data map, no training on your data, in-account option	Vague on sub-processors or retention
Control and oversight	15%	Approval gates, stop control, full run logs	”The model decides”
Exit and ownership	15%	You own configs and data, documented handoff	Lock-in by design
Team and references	5%	Named builders, relevant references you can call	Sales-only contact, no references
Price clarity	5%	All-in cost with usage assumptions stated	Hidden usage or model fees

Scoring rules I’d use:

• Two reviewers score alone, then compare.
• Any 1 in data and security or exit and ownership is a veto, whatever the total.
• Ask each finalist the same follow-up questions in writing.
• Call at least two references, and ask what went wrong.

For the vendor-side view, see how to choose an AI agent development company.

Red-Flag Answers

Some answers tell you more than a demo. Watch for these.

You ask	Red-flag answer	Why it matters
What is your success metric for us?	”Accuracy of 95 percent” with no baseline	A number with no baseline cannot be checked
Can we run a pilot on our own data?	”Our demo environment shows this”	Demos hide your exceptions
What happens if the pilot fails?	”It won’t” or a mandatory next phase	No fail terms means no real gate
Do you train on our data?	”It’s covered in the terms” with no direct answer	Make them answer in the contract
What can the agent do without approval?	”Anything it needs to”	Excessive autonomy is a known risk
How do we stop it?	No answer on kill controls	You need a hard stop
Who owns the prompts and workflows?	”The platform”	That is lock-in
Which parts are agentic?	Cannot separate the agent from RPA or a chatbot	This may be agent washing
Can we speak to a client with a similar workflow?	Only logos, no calls	References you cannot call are marketing
What does year two cost?	”Depends on usage” with no model	Usage can swing the bill

One red flag is a reason to ask more questions. Three is a reason to walk.

Bottom Line

Write the RFP around one workflow, one metric, one pilot gate, and one exit clause.

Make every vendor answer the same questions in writing, score them before demos, and treat a missing fail clause as a fail.

My next step for you: fill in the scope section for your best workflow today. If you cannot, you are not ready to send an RFP.

FAQ

How do I write an RFP for an AI agent project? Start with one workflow and describe it in detail: current steps, volume, inputs, outputs, exceptions, and who decides what. Add data and security questions, one hero metric with a baseline, a pilot gate, and exit clauses. Send the same questions to every vendor and score the answers in writing.

What belongs in the scope section? The workflow name and owner, the current process, volume, inputs and outputs, decisions the agent may make alone, decisions that need a person, the ten messiest exceptions, and what is out of scope. If a vendor cannot respond to the exceptions, they do not understand the work.

How do I score vendors? Use a weighted sheet with categories such as workflow fit, pilot gate, data and security, control, exit and ownership, team, and price. Have two reviewers score separately. Treat a very low score in data and security, or in exit and ownership, as a veto regardless of the total.

What pilot gate should the RFP require? Require a short, scoped pilot on your real data, with a metric, baseline, target, and quality guardrail agreed in advance. Hold back part of the test data from the vendor. A fail should end the contract at the pilot price. A pass should lead to a written production plan, not an automatic purchase.

What are red-flag answers? Watch for metrics with no baseline, demos instead of pilots on your data, refusal to accept fail terms, vague answers on data training, “anything it needs” for autonomy, no stop control, platform ownership of your workflows, and references you cannot call. One is a reason to probe. Three is a reason to leave.

Three Ways to Work With TMA

Need an agent built? We deploy production AI agents in your infrastructure. Working pilot. Real data. Measurable ROI. → Schedule Demo

Want to co-build a product? We’re not a dev agency. We’re co-builders. Shared cost. Shared upside. → Partner with Us

Want to join the Guild? Ship pilots, earn bounties, share profit. Community + equity + path to exit. → Become an AI Architect