# Getting Started: A Pilot Plan for Your First AI Phone Agent

Run a small, measured pilot before you commit. Pick one call type, set success measures, sign the compliance documents, test on recorded scripts, go live on a share of calls with a human fallback, then review the recordings and decide.

Updated 2026-10-29. Also on the [web page](https://voiceagents.fyi/guides/ai-phone-agent-pilot-plan).

A pilot tells you if an AI phone agent works for your callers. Keep it small. Measure it. Decide with data. This plan has six phases.

## Phase 1: Choose the scope

1. Pick **one call type**. Choose a frequent, simple call with a clear end.
2. Write the goal of the call in one sentence.
3. Write when the agent must hand the call to a person.
4. Pick the hours: after hours only, overflow only, or all calls of this type.

Starting with after-hours or overflow calls limits the risk. Your staff still take the main calls.

## Phase 2: Set success measures

Write the measures before you start. Measure the same things for your current process.

| Measure | How to count it |
|---|---|
| Task completion | Calls where the caller's goal was met without a person |
| Transfer rate | Calls the agent passed to a person |
| Error rate | Calls with a wrong answer or a wrong action, from recording review |
| Caller hang-ups | Calls the caller ended before the goal was met |
| Cost per handled call | Total pilot cost divided by calls handled |

Agree on the numbers that mean "continue" before you see the results.

## Phase 3: Shortlist and sign

1. Shortlist vendors from the category that fits your calls.
2. Ask our list of vendor questions.
3. Sign the documents you need **before** real callers reach the agent:
   - A BAA if calls include US protected health information. The required terms are in 45 CFR 164.504(e).
   - A data processing agreement if GDPR applies.
4. Check the plan limits. On 2026-10-29, Bland AI's Start plan allowed 10 concurrent calls and 100 calls per day. LiveKit Cloud's Build plan allowed 5 concurrent agent sessions.

If you plan outbound calls in the US, read the FCC's February 2024 ruling (FCC 24-17). AI-generated voices are an "artificial or prerecorded voice" under the TCPA. These calls need prior express consent, absent an emergency purpose or exemption.

## Phase 4: Build and test offline

1. Write the agent's instructions. Use your real policies and wording.
2. Connect one system at a time: first the calendar or CRM, then the transfer.
3. Collect real call recordings or transcripts as test scripts.
4. Include hard cases: background noise, accents, interruptions, unclear requests and angry callers.
5. Call the agent yourself with each script. Record the result.
6. Fix the instructions and test again.

Use test numbers, not your main line. On their pages on 2026-10-29, Twilio listed a US local number at $1.15 a month. Retell AI listed its phone numbers at $2.00 a month.

## Phase 5: Go live on a share of calls

1. Tell callers that they are speaking with an AI. Tell them the call is recorded.
2. Route only the chosen call type and hours to the agent.
3. Keep a human fallback for every call.
4. Review recordings every day in the first week.
5. Fix problems in the instructions, not with one-off workarounds.

Track the cost. Use your vendor's usage page. For a rough guide, 1,000 minutes at Bland AI's Start rate of $0.14 per minute is $140, and Vapi's calculator estimates $82 to $129 a month for 1,000 minutes. Your rate depends on your model, voice and telephony.

## Phase 6: Review and decide

Compare the results with your measures from Phase 2. Then choose one:

- **Expand.** Add hours, add a call type, or raise the share of calls.
- **Adjust.** Change the instructions, the integrations or the call type, and run again.
- **Switch.** Test another vendor on the same scripts.
- **Stop.** Keep the current process. Record what you learned.

Before you expand, check the plan again. Concurrency, uptime commitments and support terms often change between self-serve and enterprise plans.

## Pilot checklist

- [ ] One call type and a written goal
- [ ] Success measures and a baseline
- [ ] BAA or data processing agreement, if needed
- [ ] Consent and AI disclosure in the greeting
- [ ] Offline tests with hard cases
- [ ] Human fallback on every call
- [ ] Daily recording review in week one
- [ ] A decision meeting with the data

Prices and plan limits change. Check each vendor page before you start.

## Sources

- [Vapi, Pricing (fetched 2026-10-29)](https://vapi.ai/pricing)
- [Retell AI, Pricing (fetched 2026-10-29)](https://www.retellai.com/pricing)
- [Bland AI, Pricing (fetched 2026-10-29)](https://www.bland.ai/pricing)
- [Deepgram, Pricing (fetched 2026-10-29)](https://deepgram.com/pricing)
- [LiveKit, Pricing (fetched 2026-10-29)](https://livekit.com/pricing)
- [Twilio, Programmable Voice pricing, United States (fetched 2026-10-29)](https://www.twilio.com/en-us/voice/pricing/us)
- [FCC, Declaratory Ruling FCC 24-17, CG Docket No. 23-362 (fetched 2026-10-29)](https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf)
- [eCFR, 45 CFR 164.504, Uses and disclosures: organizational requirements (fetched 2026-10-29)](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-164/subpart-E/section-164.504)

## Questions

### Can I run a pilot without a contract?

Often, yes. Several vendors offer self-serve plans and free credits. Vapi listed $5 in free credits, Retell AI $10 and Deepgram $200 on 2026-10-29. Sign the compliance documents before you use real caller data.

### Which call type should I start with?

A frequent, simple call with a clear end, for example booking or rescheduling an appointment. Leave complaints and complex cases to people at first.

### What if the pilot fails?

That is a useful result. Review the recordings to see why. The cause may be the call type, the integrations or the vendor.