
Inbound AI Voice Cover: What to Do With the Calls Nobody Answers
Let the team answer first, then send only unanswered calls to Voice AI after a tested timeout. Give the agent a narrow job: collect contact details, answer approved questions, book appointments, send follow-up texts, trigger defined workflows, and transfer callers under explicit conditions—never guess. At the lowest published pay-per-use rate, the mechanical crossover against the $97 Unlimited plan is 1,617 AI minutes per month, or roughly 539 three-minute calls, before token and Phone System charges.
Ring first, then let Voice AI cover the gap
For a service business, the phone problem is usually not that humans should stop answering. It is that technicians, dispatchers, and owners cannot answer while driving, working with a customer, or handling another call.
HighLevel supports this operating model directly. In AI Agents → Voice AI, open the agent, select Deploy, and choose Use as Backup. Configure the human, user, or forwarding routes under Settings → Phone System → Phone Numbers → Edit Configuration, then set the Incoming Call Timeout and select Voice AI as the backup. HighLevel gives approximately 20 seconds as the starting timeout.
This keeps the normal phone experience intact. The caller first reaches the people who know the job, schedule, and customer. Voice AI enters only after that human route fails to connect.
The timeout is not just a preference about how long callers should hear ringing. The backup agent receives the call only after the timeout expires. If a forwarded mobile phone or other outside destination sends the caller to carrier voicemail first, HighLevel can treat the call as completed and never send it to Voice AI.
Test the timeout against the actual voicemail delay on every forwarded destination. If carrier voicemail answers before the backup, shorten the Incoming Call Timeout. Under the phone number’s Advanced Settings, enable Call Connect so the receiving person has to confirm the call before HighLevel treats the forwarding leg as connected. That prevents an answering machine from consuming a call intended for the backup agent.
The result should be simple: the team gets a reasonable chance to answer, voicemail does not intercept the call, and the AI takes over without making the caller start again.
Routing is different for known and unknown callers
Do not test the setup with one personal phone number and declare it finished. HighLevel’s owner-based routing can send a known contact to the contact’s assigned user or to the owner of the called number before using the general forwarding destination or AI backup.
An unknown caller can therefore follow one route while an existing customer follows another. Test both:
- Call from a number that is not stored as a contact.
- Create or use a known contact and assign that contact to a user.
- Call again from the known contact’s number.
- Let each call ring without answering and confirm that both reach the backup agent.
- Answer each route once and confirm that Voice AI stays out of the call.
If several employees are able to take inbound work, Ring More Team Members supports up to six additional eligible users ringing simultaneously. More ringing destinations increase the opportunity for a human answer, but every destination still needs to be tested for carrier voicemail and confirmation behavior.
Document the final route in plain language: which users ring, whether they ring together, how long they ring, what confirmation is required, and what happens when nobody accepts. A routing diagram is less useful than a short test script that the business can repeat after changing a phone, user assignment, or forwarding number.
Give the backup agent a narrow job
A backup agent does not need authority over the entire customer relationship. It needs enough authority to prevent an unanswered call from becoming a dead end.
A deployed Voice AI agent supports actions for answering questions, collecting and updating contact information, booking appointments, sending SMS messages, triggering workflows, transferring calls, and handing a live conversation to another Voice AI agent. These actions are added under AI Agents → Voice AI → Agent List → Actions → + New Action. Each action needs a trigger condition and a defined destination, workflow, field, calendar, or integration setting.
| Agent is allowed to | Required boundary |
|---|---|
| Answer questions | Answer only the topics and claims included in the prompt. Do not guess when required information is unavailable. |
| Collect caller information | Specify the fields to collect and when each field should be updated. |
| Book an appointment | Connect the intended calendar and define the condition that permits booking. |
| Send an SMS | Define the trigger and the content or purpose of the message. |
| Trigger a workflow | Name the workflow and the exact conversation condition that starts it. |
| Transfer the call | Define whether the destination is a human phone number or another Voice AI agent. |
The prompt also needs an explicit prohibition list. Tell the agent which topics it does not answer, which claims it does not make, what information it must collect before taking an action, when it must escalate, and how it ends the conversation. HighLevel’s Voice AI setup guidance specifically requires instructions not to guess when information is missing.
This is the same core discipline described in Conversation AI in GoHighLevel: How It Actually Decides to Reply: the agent needs usable instructions and action conditions, not a vague request to be helpful.
A practical backup role is narrower than a receptionist’s full role. It captures who called, why they called, how they want to be reached, and the next permitted step. If the approved information does not support an answer, the agent records the question and routes or escalates it instead of constructing an answer.
Use human and AI transfers for different jobs
HighLevel has two transfer actions, and they do not go to the same type of destination.
Call Transfer dials a human representative or external phone number. It requires a destination phone number. Use it when the caller needs a live team member, a high-value lead needs sales, or a sensitive issue requires human judgment.
Agent Transfer keeps the caller on the same live call and hands the conversation to another Voice AI agent. It does not dial a phone number. This is useful when the first agent performs basic intake and a second agent has a separate, tightly defined responsibility.
A root agent supports up to three destination Voice AI agents. Each destination gets its own natural-language transfer condition. Under + New Action → Agent Transfer → Connect Agent, Speak During Execution adds a spoken handoff message. Without it, the transfer happens without a spoken transition, and the destination agent’s greeting does not play.
Keep the transfer map one-directional. HighLevel warns that circular paths can create transfer loops and cause the call to fail. The root agent’s Maximum Call Time, Working Hours, and backup settings govern the connected chain; the corresponding settings on destination agents are ignored.
Write transfer conditions as observable events rather than broad intent. For example, transfer to a human when the caller explicitly asks for a person, when the issue is marked sensitive in the prompt, or when the approved information does not contain the required answer. Transfer to another agent only when that destination has a distinct responsibility and no route back to the root agent.
Some calls should not remain with an agent at all. When an AI Agent Is the Wrong Answer is the correct standard for deciding which conversations require human judgment rather than another prompt adjustment.
Test through the real phone network
The browser-based Web Call test is useful for checking prompts, question handling, data collection, and conversational behavior. It does not support Call Transfer, so it cannot prove that the production transfer route works.
Use Phone Call testing or call the deployed number to validate caller ID, routing order, transfer destinations, and destination behavior. Phone Call testing uses the account’s normal telephony setup, and the normal telephony charges apply.
Run a fixed acceptance test before launch:
- Place one unknown-caller test and one assigned-contact test.
- Answer during the human ring window and confirm the AI does not join.
- Decline to answer and confirm the agent enters only after the configured timeout.
- Let a forwarded phone approach carrier voicemail and confirm voicemail does not intercept the call.
- Trigger every human transfer condition and verify the correct destination rings.
- Trigger every Agent Transfer condition and verify the correct destination agent takes over.
- Ask an unanswered or prohibited question and confirm the agent does not invent an answer.
- Trigger each booking, SMS, contact update, and workflow action once, then verify the resulting record.
- End the call without qualifying for an action and confirm no unintended action runs.
Repeat the phone tests after changing a forwarding destination, timeout, user assignment, transfer number, working-hours rule, or agent action. The important unit of testing is the whole route, not the prompt by itself.
What the AI minutes cost
HighLevel’s pay-per-use Voice AI pricing is not one flat rate. Text-to-speech configurations combine a $0.045-per-minute Voice Engine charge, a TTS rate, LLM token usage, and separate Phone System charges.
| Configuration | Published AI time rate | Included in that figure |
|---|---|---|
| OpenAI TTS | $0.060/minute floor | $0.045 Voice Engine plus $0.015 TTS |
| Cartesia | $0.060/minute floor | $0.045 Voice Engine plus $0.015 TTS |
| ElevenLabs V2.5 | $0.080/minute floor | $0.045 Voice Engine plus $0.035 TTS |
| ElevenLabs V3 | $0.215/minute floor | $0.045 Voice Engine plus $0.170 TTS |
| Google speech-to-speech | $0.10/minute | Bundled audio, text, cache tokens, transcripts, and infrastructure |
| OpenAI realtime speech-to-speech | $0.20/minute | Bundled audio, text, cache tokens, transcripts, and infrastructure |
The TTS floors exclude LLM token usage and Phone System charges. HighLevel’s example prices a ten-minute OpenAI TTS call at $0.60 before those additional costs. The speech-to-speech prices bundle the listed AI processing components, but Phone System charges remain separate.
HighLevel also lists AI Employee Growth at $50 per enabled location per month and Unlimited at $97. Growth includes 100 Voice AI minutes shared across inbound, outbound, and widget use, followed by pay-per-use overages. Unlimited includes unlimited Voice AI across those usage types under fair-use rules. HighLevel retains the right to throttle, limit, require an upgrade, or terminate access for excessive or abusive usage. Phone charges remain payable under every option.
The call volume where Unlimited overtakes per-minute pricing
At the lowest published TTS floor, divide the $97 Unlimited price by $0.060 per minute. The result is 1,616.7 minutes, rounded to 1,617 AI minutes per month. At a three-minute average AI-handled call, that is roughly 539 calls per month.
This is a mechanical comparison, not a billing quote. It excludes LLM tokens and Phone System charges. Token costs push the actual pay-per-use crossover below 1,617 minutes, while Phone System charges remain separate under both approaches.
| AI configuration | $97 crossover | Three-minute calls |
|---|---|---|
| OpenAI TTS or Cartesia floor | 1,617 minutes | 539 calls |
| ElevenLabs V2.5 | 1,213 minutes | 404 calls |
| ElevenLabs V3 | 451 minutes | 150 calls |
| Google speech-to-speech | 970 minutes | 323 calls |
| OpenAI realtime speech-to-speech | 485 minutes | 162 calls |
The Growth-to-Unlimited comparison is different because Growth already includes 100 minutes. At the $0.060 floor, the $47 difference between the $50 and $97 plans covers 783 overage minutes. Add Growth’s 100 included minutes, and the mechanical crossover is near 883 total monthly minutes, or roughly 294 three-minute calls.
Use AI-handled minutes, not total inbound calls, for this decision. In a ring-then-AI setup, calls answered by the team do not become Voice AI conversations. Pull a full month of unanswered-call counts, multiply by the measured average AI conversation length, and compare that total with the threshold for the selected voice configuration.
Pay per use is the clearer cost choice below the applicable threshold. Unlimited becomes the stronger mechanical comparison above it, subject to its fair-use terms. The agent’s scope still matters: shorter intake and booking calls consume fewer minutes than long conversations that attempt to answer every possible question.
What to do
- Set Incoming Call Timeout to 20 seconds initially, then place an unanswered test call against every forwarded carrier voicemail and shorten the timeout if voicemail answers first.
- Enable Call Connect under the phone number’s Advanced Settings for forwarded routes so an answering machine is not treated as a confirmed human answer.
- Test the completed route with exactly two caller types before launch: one unknown number and one known contact assigned to a user.
- Limit each Voice AI action to one written trigger condition and one named destination, workflow, field, calendar, or transfer route.
- Use Call Transfer for human or external-number destinations; use Agent Transfer only for another Voice AI agent, with no circular path back to the root agent.
- Run all transfer acceptance tests through Phone Call testing or the deployed number because Web Call does not execute Call Transfer.
- Calculate monthly AI volume as unanswered calls multiplied by measured average AI call length; compare the result with 1,617 minutes at the $0.060 TTS floor, 1,213 at $0.080, 451 at $0.215, 970 at $0.10 speech-to-speech, or 485 at $0.20 speech-to-speech.
- Compare Growth with Unlimited at 883 total monthly minutes when using the $0.060 overage floor, including Growth’s 100 included minutes.
Questions people ask
Should Voice AI answer every inbound call first?
No. Configure the agent as the backup, let the human routes ring first, and send the call to Voice AI only after the Incoming Call Timeout expires.
Why does carrier voicemail prevent the AI from answering?
If the forwarded destination’s voicemail answers first, HighLevel can treat the call as completed. Shorten the timeout and enable Call Connect so an unconfirmed answering machine does not consume the call.
Can the browser test validate transfers?
No. Web Call tests prompts and conversation behavior but does not support Call Transfer. Use Phone Call testing or call the deployed number.
What should the backup agent do when it lacks an answer?
It should not guess. Configure it to collect the caller’s question and contact details, then transfer or trigger the defined follow-up path.
How many Voice AI agents can receive an Agent Transfer?
A root agent supports up to three destination Voice AI agents, each with its own transfer condition. Avoid circular paths because loops can cause the call to fail.
We build voice and conversation AI inside GoHighLevel. Book a call and we'll tell you if it's the right fit — including when it isn't. We will map the human ring route, backup timeout, agent permissions, transfer conditions, test plan, and expected minute volume before recommending a build.