AI can run bounded outbound conversations, but it does not make a call lawful or persuasive. Start with consent, suppression, disclosure, human handoff, and audit controls.
AI can place and handle outbound calls, but it does not remove the legal, reputational, or operational risks of cold calling. In the United States, an AI-generated voice is treated as an artificial or prerecorded voice under the Telephone Consumer Protection Act (TCPA). Telemarketing rules can require prior express written consent, Do Not Call suppression, identification, opt-out mechanisms, calling-time controls, and call-abandonment controls. State and sector-specific requirements may add more.
That makes the right first question “Which calls are we permitted to automate?”, not “How many calls can we make?” Have counsel review the audience, purpose, jurisdiction, consent record, and script before launch. Then use AI only inside a narrow workflow with an immediate opt-out and a reliable path to a person.
What AI should and should not do on a cold call
A bounded voice agent can be useful when the recipient, purpose, and permitted actions are already defined. Examples include:
- confirming whether a person wants to continue a conversation they previously agreed to receive;
- asking a small set of qualification questions that do not determine credit, employment, insurance, housing, healthcare, or another consequential outcome;
- scheduling a follow-up with a sales representative;
- recording an explicit “do not call again” request and suppressing the number immediately; and
- transferring an interested recipient to an available human representative.
Do not let the agent invent offers, conceal that it is automated, pressure someone who wants to leave, or make a binding decision about eligibility, price, contract terms, or regulated products. A language model can generate an incorrect answer even when the surrounding telephony and CRM integrations work correctly.
The Federal Communications Commission's declaratory ruling on AI-generated voices explains why TCPA restrictions for artificial or prerecorded voices apply to current voice-generation technology. The Federal Trade Commission's Telemarketing Sales Rule guidance covers the National Do Not Call Registry, company-specific suppression, disclosures, calling windows, abandoned calls, and prerecorded telemarketing requirements. These are starting points, not a substitute for legal review.
Put permission and suppression ahead of the dialer
A campaign should not enter the call queue unless its permission and suppression checks have passed. Build those checks outside the conversational model so the agent cannot talk its way around them.
For each number, retain the source, collection time, disclosure shown when consent was obtained, permitted caller or seller, permitted purpose, applicable expiration, and any revocation. Check the number against all required federal, state, internal, customer, and campaign suppression lists immediately before scheduling. Treat an opt-out during a call as a deterministic event: acknowledge it, end the marketing flow, write the suppression record, and prevent later retries.
The governing federal rules are available in the Telemarketing Sales Rule at 16 CFR Part 310 and the FCC's calling restrictions at 47 CFR 64.1200. Which provisions apply depends on the call, recipient, technology, and purpose.
Design the conversation around disclosure and exit
The opening should identify the seller and the commercial purpose, state that the recipient is speaking with an automated or AI voice where required or appropriate, and offer a simple way to stop. Do not bury those facts after qualification questions.
Keep the first version short:
- Identify the caller, seller, and automated nature of the call.
- Ask whether the recipient wants to continue.
- If yes, ask only the approved questions.
- Confirm any requested appointment or transfer before taking action.
- If the recipient declines or asks not to be called, suppress and end the call.
- If the agent is uncertain, a tool fails, or the recipient asks for a person, transfer or end the call safely.
Store approved claims, offers, and eligibility rules in controlled systems. Let the model choose how to express an approved response, not whether a new promise should exist.
Connect the agent to systems with explicit failure rules
An outbound workflow usually depends on a customer relationship management (CRM) system, calendar, consent store, suppression service, and human queue. Those integrations determine whether the agent can complete the promised action.
With Dasha, teams can schedule outbound calls through the API, use tools to call approved external APIs, and configure warm, cold, or webhook-routed transfers. These capabilities do not verify consent or compliance for you. Your application must decide which records are callable, which fields the agent may read or write, and what happens when a dependency is unavailable.
Use least-privilege credentials. Validate phone numbers and time zones before scheduling. Make appointment and suppression writes idempotent so a retry does not create duplicates. If the CRM or consent service cannot be reached, do not let the model guess; stop the call or route it to a trained person.
Test the failure cases before a live campaign
A happy-path demo is not enough. Test the agent against realistic interruptions, accents, silence, voicemail, wrong numbers, hostile responses, ambiguous opt-outs, transfer failures, stale CRM data, tool timeouts, and questions outside the approved script.
Dasha supports browser-based voice and chat testing. After a test call, Call Inspector can expose the transcript, recording when enabled, model inputs and outputs, tool calls, timeline, and latency breakdown. Use those records to investigate behavior, but define access, retention, notice, and recording-consent controls before collecting production conversations.
Review a statistically meaningful sample manually. Automated labels can help find calls for review, but they should not be the only evidence that disclosures, opt-outs, and transfers worked.
Measure safety and recipient experience, not just conversion
Set stop conditions before the first live call. Useful measures include:
- calls attempted without a complete permission record;
- calls placed to a suppressed number;
- disclosure completion and recipient continuation rate;
- opt-out recognition, write success, and repeat-call rate after opt-out;
- complaints, wrong-party calls, and requests for a human;
- tool, scheduling, and transfer failures;
- unsupported or inaccurate statements found in human review;
- abandoned-call and calling-time compliance; and
- appointments confirmed by the recipient, not merely created by the system.
Do not attribute a lift in meetings or revenue to the AI unless the experiment isolates the effect and accounts for audience, list quality, offer, timing, and human follow-up. More dialing is not evidence of a better customer outcome.
A sensible pilot sequence
Start with one jurisdiction, one seller, one well-documented permission source, one offer, and one outcome such as scheduling a human follow-up. Limit daily volume and concurrency. Have an owner watching live results who can stop the campaign immediately.
The launch checklist is simple:
- Legal and compliance owners approve the audience, consent evidence, script, disclosures, calling windows, suppression logic, recording policy, and retention.
- Engineering validates deterministic pre-dial eligibility and immediate opt-out writes.
- Sales approves every permitted claim and handoff rule.
- Quality reviewers test adversarial and failure scenarios, then document launch thresholds.
- Operations runs a small pilot, reviews calls daily, and pauses on any material control failure.
If a workflow cannot meet those conditions, keep it human-run. AI cold calling is a deployment choice, not a shortcut around permission, accountability, or good sales practice.