How to write an RFP for an AI voice agent platform

What to give vendors, which questions to ask, how to weight the answers, and why a pilot on your own calls should decide which AI voice agent platform you choose.

  • Nikola Mrkšić

    Co-founder & CEO, PolyAI

  • 16 Nov 2026
  • 10 min
How to write an RFP for an AI voice agent platform

Share

A good RFP for an AI voice agent platform starts with your data. Give vendors your call volumes by intent, handling times, peak patterns, languages, telephony, and the systems the agent must read and write, so they can price and design honestly.

Then ask questions whose answers you can check, such as:

  • How the vendor proves a call was resolved
  • How calls reach a person, what the price includes
  • Where data is stored, and how reliability is measured.

Finish with a pilot on your own calls, weight the scoring toward what the pilot measures, and ask every vendor to state its pricing unit, inclusions, minimums, and rounding in writing.

Before you write: what to prepare

Vendors design and price from what you tell them. The more of this you share, the more comparable the answers.

  • Call volume by intent for the last 6 to 12 months, with seasonal peaks and the busiest hour.
  • Handling time by intent, including hold and after-call work.
  • Current IVR and routing, with menu structure and current containment if you measure it.
  • Languages and regions callers use, and any regulated jurisdictions.
  • Telephony: carrier, contact-center platform (for example Amazon Connect, Five9, Genesys, NICE CXone), session border controllers, and how transfers work today.
  • Systems of record the agent must use: CRM, booking, order, billing, identity verification, payments, ticketing.
  • Policies: authentication rules, payment handling, call recording consent, vulnerable-customer handling.
  • Success measures you'll use, with resolution defined the way our call deflection guide does: the action completed in the system of record and confirmed by the customer.
  • A sample of anonymized call recordings or transcripts for the intents in scope, under NDA.

Structure of the RFP document

A clear structure makes vendor responses easy to compare and shows your own team what you will measure. Include these nine sections, in this order, so every vendor answers the same things on the same terms.

  1. Company background and the problem you want to solve.
  2. Scope: intents, channels (voice, and chat or SMS if wanted), languages, regions, volumes.
  3. The data pack above.
  4. Question lists (below), with a response format and word limits.
  5. Pricing schedule, in a format you supply so answers are comparable.
  6. Pilot terms: scope, duration, success measures, and who pays.
  7. Commercial and legal terms: data processing, liability, termination, data return.
  8. Timeline: questions deadline, response deadline, shortlist, pilot, decision.
  9. Scoring method and weights.

Question lists

1. Use cases and outcomes

  1. Which of our listed intents would you automate first, and why?
  2. For each intent, what share of calls do you expect the agent to resolve without a transfer, and what assumptions sit behind that estimate?
  3. How do you define containment, resolution, and handoff? Which calls do you exclude from the denominator?
  4. Describe two deployments similar to ours in industry, volume, and integration depth. What did they measure, and over what period?
  5. Which intents would you advise us not to automate?

2. Conversation quality

  1. How does the agent handle interruptions, corrections, and callers who give several pieces of information at once?
  2. How does it handle accents, background noise, and poor phone lines?
  3. How does it confirm names, addresses, numbers, and spellings?
  4. Which languages are supported for both recognition and speech, and are any of them in a lower tier of support?
  5. Can we choose or customize the voice, and how is the voice kept consistent across releases?
  6. How does the agent behave when it doesn't know the answer?
  7. What controls stop the agent from stating policy, prices, or commitments that aren't in its approved knowledge?

3. Telephony and contact-center integration

  1. Which carriers and contact-center platforms do you integrate with natively, and which over SIP?
  2. Do you provide phone numbers, or do we keep ours?
  3. How are calls transferred to a human: SIP REFER, SIP INVITE with the vendor bridging, or a platform API? Who stays on the call after transfer?
  4. What context reaches the human agent at transfer (identity status, intent, summary, collected data), and how (SIP headers, API, CRM record)?
  5. How do you honor a caller's request to speak to a person, and how many turns does it take?
  6. How do you handle queue overflow, after-hours calls, and callbacks?
  7. What audio codecs and sample rates do you support?

4. Back-end integration

  1. How does the agent read from and write to our systems: prebuilt connectors, REST calls, or code we or you write?
  2. Who builds and maintains each integration, and who is responsible when our API changes?
  3. How is caller identity verified before account-specific actions?
  4. How are payments handled, and does card data ever pass through your systems?
  5. What happens on a call when a back-end system is slow or unavailable?

5. Measurement and analytics

  1. Which metrics are available out of the box, and how is each defined?
  2. Can we access raw call data (recordings, transcripts, events, metrics) through an API, and in what format?
  3. How do you measure repeat contacts after a call the agent resolved?
  4. How are handoff reasons captured and reported?
  5. How is customer satisfaction measured on AI-handled calls?
  6. Will you report resolution by intent, with exclusion rules stated, every month?

6. Testing and change control

  1. How do we test the agent before go-live and before each change?
  2. Can we run automated or simulated test conversations, and can they block a release that fails?
  3. Who can edit the agent's knowledge and behavior: our team, yours, or both? How long does a typical change take from request to live?
  4. How are changes versioned, reviewed, and rolled back?
  5. How are model or platform updates from your side tested against our agent before they reach our callers?

7. Security, privacy and compliance

  1. Which certifications do you hold (for example, SOC 2 Type II, ISO 27001), and for which entity and scope? Please attach current reports.
  2. In which regions are call audio, transcripts, and data processed and stored? Can we choose?
  3. What are the default and configurable retention periods for recordings, transcripts, and logs?
  4. Is our data used to train models shared with other customers? Can we opt out, and is that in the contract?
  5. Which sub-processors handle our data, including speech recognition, language model, and voice providers?
  6. How do you redact personal and payment data from recordings and transcripts?
  7. Which regulations do you support for our industry (for example, PCI DSS, HIPAA, GDPR), and what evidence can you provide?
  8. How do you handle callers' recording-consent and data-access requests?

8. Reliability and scale

  1. What was your measured production uptime over the last 12 months, and how is it measured?
  2. What uptime do you commit to contractually, and what are the service credits? (Ask for measured uptime and the contractual SLA as two separate figures.)
  3. How many concurrent calls can our deployment handle, and how quickly can that limit be raised for a peak?
  4. What happens to live calls during an outage or failover? Where do they route?
  5. How do you notify us of incidents, and what are your response times by severity?

9. Pricing

Ask vendors to complete a pricing schedule you supply, so that every answer uses the same structure.

  1. What is the pricing unit: per connected minute, per call, per conversation, per resolution, or a platform fee?
  2. What does the unit include: telephony, speech recognition, language model, voice, analytics, maintenance, support?
  3. Are any costs passed through at the provider's price, and how do they vary?
  4. Is usage billed per second, or rounded up per call or per minute?
  5. Are there minimum commitments, platform fees, or charges for testing and non-production use?
  6. What are the one-off implementation fees, and what is fixed or variable in them?
  7. How do prices change at higher volume, and over a multi-year term?
  8. Using our data pack, what would our first-year cost be at the resolution rate you estimated? Show the calculation.

Our ROI guide sets out the formulas to check these answers.

10. Implementation and support

  1. What is the plan from contract to first live calls, and to full scope? Who is on your team?
  2. What do you need from our team, and for how many hours a week?
  3. How do you launch: a share of traffic, one intent, one site?
  4. After launch, who improves the agent's performance, and is that included in the price?
  5. What support hours and channels are included?
  6. May we speak to two current customers of similar size?

Run a pilot on your own calls

Written answers on conversation quality are hard to compare, and they can't show how an agent handles your callers, your systems, and your edge cases. Shortlist two or three vendors and run a pilot on real calls for a defined set of intents, for long enough to see a full weekly cycle, and measure:

  • Resolution by intent, with your exclusion rules and repeat contacts subtracted.
  • Requests for a human and how quickly each was honored.
  • Outcomes in the system of record: bookings written, orders found, payments taken.
  • Customer satisfaction on AI-handled calls.
  • Handling time saved on transferred calls.
  • How long your team's change requests took to go live.

Agree before the pilot who pays for it, what happens to the data afterwards, and what result would lead to a contract.

How to score responses

A weighting that keeps the decision tied to outcomes (adjust to your priorities):

Table
AreaSuggested weight
Pilot results: resolution, outcomes, satisfaction35%
Integration fit: telephony, contact-center platform, back-end systems15%
Security, privacy and compliance15%
Total cost over the contract term, from the pricing schedule15%
Measurement, testing and change control10%
Implementation plan, support and references10%

Score each answer against evidence: a document, a demonstration, or a pilot measurement. An answer of "yes" without evidence scores lower than a partial answer with evidence.

Mistakes to avoid in your AI voice agent RFP

Most AI voice agent RFPs go wrong in the same six places, and each one makes vendor responses harder to compare. Avoid these common mistakes so your RFP produces answers you can verify and score.

  • Asking only yes-or-no questions. Every vendor answers yes. Ask how, and ask for evidence.
  • Comparing prices in different units. A per-minute platform fee without telephony and a per-minute rate with everything bundled in are different things. Use your own pricing schedule.
  • Accepting a resolution or containment figure without its definition. Ask which calls were counted and excluded.
  • Treating measured uptime as a commitment. Ask for the contractual SLA separately.
  • Skipping the pilot. See the pilot section above for what it measures.
  • Leaving out the people who will run it. Include your contact-center operations, workforce management, IT security and the team that will own changes.

How PolyAI answers a few of these

As a worked example, these are answers PolyAI publishes on its own pages. Ask every vendor for the equivalent.

  • Handoff method: PolyAI documents SIP REFER (the default, where PolyAI passes a transfer destination to the customer's session border controller and leaves the call) and SIP INVITE (where PolyAI bridges the call), with context passed in custom SIP headers or through a Handoff API.
  • Integrations: PolyAI documents voice integrations including custom SIP, Twilio, Amazon Connect, Five9, Genesys, NICE CXone and Dialpad.
  • Regions: PolyAI's API reference lists US, UK and EU regions.
  • Certifications: PolyAI's pricing page states "SOC 2 and ISO 27001 certified infrastructure, with regular audits and testing."
  • Measured uptime: PolyAI's API reference states 99.994% production uptime, measured over the trailing 12 months. This is measured uptime, not an SLA; the Enterprise plan carries a committed uptime SLA.
  • Testing: Dialog Studio includes simulation tests that run "automated conversations" from written caller scenarios, with assertions that pass or fail.

To see how PolyAI answers the rest on your own calls, speak to our team.

Frequently asked questions

What should an RFP for an AI voice agent include?

A data pack about your calls, systems and policies, question lists that vendors answer with evidence, a pricing schedule in your format, pilot terms and a scoring method.

How many vendors should we send the RFP to?

Enough to see the range of approaches, usually four to six, then shortlist two or three for a pilot on your own calls.

What is the most important question in the RFP?

How the vendor proves a call was resolved, because that number drives cost and ROI. Ask which calls are counted, which are excluded, and whether repeat contacts are subtracted.

Should we ask for an SLA or for uptime history?

Both, as separate figures. Measured uptime shows what happened; the SLA shows what the vendor commits to and what credits apply.

How should we compare AI voice agent pricing?

Supply a pricing schedule and require each vendor to state the unit, what it includes, pass-through costs, rounding, minimums and implementation fees, and to calculate your first-year cost from your data pack.

Do we need a pilot if the RFP answers are strong?

Yes. Only a pilot shows how an agent handles your callers, your systems and your edge cases.

Evaluating dialog agents?

A dialog agent does the work behind the call as well as the conversation. Add 20 questions to your RFP that show which vendors can deliver that.

Hear a voice agent on your own calls

Book a 30-minute demo and we’ll run a conversation from your world through a live agent.