Outbound voice agent (reminders, surveys, collections)

Organisations make large numbers of routine outbound calls: appointment reminders, delivery confirmations, renewal and satisfaction surveys, gentle payment reminders. They are expensive to staff and often unanswered. A voice agent can place these calls, hold a natural conversation, update the source system and escalate to a human. The technology is ready for low-stakes, consented calls; the hard part is legal permission to call at all, honest AI disclosure and fair treatment of debtors.

Problem & value

• Today: staff or call-centre agencies phone patients, customers or citizens from lists; results are typed back by hand. Calls happen in office hours, scale poorly and many go unanswered. Text reminders are cheap but often ignored.
• AI value: calls scale up and down on demand, run outside office hours, can rebook an appointment in the same call, and write structured results back automatically. Staff focus on complex or sensitive cases.
• Evidence: vendor and aggregator material quotes per-minute platform prices from roughly USD 0.05 to 0.31 depending on stack, plus telephony and LLM costs [4]. These are vendor figures; measure answer rates and cost per completed outcome in your own pilot rather than trusting benchmarks.
• Hard constraint (EU): the call must be legally permitted. In Germany, marketing calls to consumers need prior express consent (§ 7 UWG), documented and kept five years (§ 7a UWG) [2]. Many reminder, survey and dunning calls are not marketing, but a survey with a sales undertone can be reclassified.

Variants

• Appointment and no-show reminders (clinics, dentists, workshops, authorities): confirm, reschedule, cancel; links to CONV-003.
• Customer surveys and NPS by phone (utilities, retail, telecoms): short structured interviews, open answers coded afterwards (LANG-013).
• Service notifications: delivery windows, outage information, contract or renewal reminders.
• Payment reminders and early-stage collections: polite reminder, promise-to-pay capture, payment link by SMS; hard negotiation stays human (see AGENT-023).
• Public sector: vaccination or screening invitations, census-type follow-ups, reminders for deadlines (with a statutory basis; no marketing).
• Care check-in calls (post-discharge, medication adherence): borderline health context; see CONV-010 and CONV-016.
• Lead follow-up: outbound sales calls are the most legally exposed variant, see CONV-005 and the consent section below.

How it works

1. trigger (CRM/PMS/ERP event or campaign list)
2. eligibility gate (consent / legitimate basis, opt-out and do-not-call, time window, call cap)
3. dialler (SIP trunk, caller ID, answering-machine detection)
4. greeting with AI disclosure + purpose + recording notice
5. real-time dialogue loop: ASR
6. LLM with tools (calendar, billing, CRM)
7. TTS (or speech-to-speech model)
8. guardrails (scope, no legal threats, identity check, barge-in, silence handling)
9. outcome + summary
10. write-back to CRM/ERP; opt-out flag if requested
11. escalation: warm transfer or callback ticket for a human

• Voice loop: either a cascaded pipeline (speech recognition, LLM, speech synthesis) or a native speech-to-speech model. Sub-second turn latency, interruption handling and answering-machine detection decide perceived quality.
• Tools, not free text: the LLM calls functions (find slots, book, mark promise-to-pay). Business rules such as dunning stages, discounts and allowed wording live in code and policy, not in the prompt.
• Identity check before personal data: a caller picking up may not be the data subject. The agent discloses nothing sensitive until a lightweight verification passes, and never leaves details on voicemail.
• Human-in-the-loop: transfer on request, distress, dispute, vulnerability signals or low confidence; review of sampled transcripts; humans own any decision with financial consequence.

Data preparation

The agent needs a clean call list with name, number, call reason and context (appointment, invoice, survey target), usually from CRM, ERP, practice or booking systems, and sometimes from Excel lists. Equally important are consent and opt-out data, calling-time rules, scripts, and escalation rules. Typical gaps: wrong or duplicate phone numbers, missing consent evidence (telephone marketing requires prior consent under UWG), unclear invoice status, scripts living in people's heads, and no record of call outcomes.

Preparation steps: (1) export contact and case data from CRM or ERP and deduplicate; (2) validate phone numbers and normalise to E.164; (3) record consent status and a do-not-call list per contact; (4) write script variants, FAQ answers and objection handling; (5) define outcome codes and write-back fields; (6) prepare a test set of 20-30 call scenarios. Tools: CRM exports, n8n for syncing, spreadsheet validation. Rough share of effort: about 20 percent; the main work is process and consent clarity rather than data volume.

Useful PREP use cases: PREP-005 (data export), PREP-022 (pseudonymisation), PREP-025 (purpose and legal basis), PREP-018 (test sets). See also GUIDE-003.

Implementation paths

Which path for whom

• SMB — A: hosted outbound voice-agent services with EU hosting handle standard reminders and surveys without telephony know-how.
• Enterprise — B: the agent must work against own CRM, billing and telephony stacks with defined escalation, and call volume makes per-minute pricing relevant.
• Developer — B: realtime speech APIs plus telephony (SIP) and workflow tools give flexible control over dialogue and compliance logic.
• Public sector — C: citizens' contact data and procurement rules favour a sovereign, self-hosted or German-cloud stack with open speech models.

Path A — Buy (recommended for most SMBs)

• Voice-agent platforms with outbound campaigns: Retell AI, Vapi, ElevenLabs Agents, Bland, Synthflow; EU-focused or EU-hosted options include Parloa (DE), Cognigy (DE, NICE), Vocalcom, Fonio and Voicefy-type SMB vendors, plus Deepgram Voice Agent API as a building block [4][5].
• Contact-centre suites: Genesys, NICE, Five9, Twilio Flex with built-in outbound dialler and AI agents; Microsoft and Google contact-centre offerings.
• Healthcare and practice vertical tools often bundle reminder calls with practice software.
• Check: EU processing region, DPA and sub-processor list, spoken AI disclosure, consent and do-not-call management, opt-out handling, call recording controls, transfer to humans, exports.
• Choose A when: reminders or surveys with a few scripts, tens of thousands of calls per year.

Path B — Build on cloud APIs

• Realtime models: OpenAI Realtime API with SIP, Azure OpenAI/Foundry (EU regions), Gemini Live, or a cascade of Deepgram/Azure Speech/Mistral Voxtral (EU) for ASR, an LLM (Claude, Mistral, GPT, Gemini) and ElevenLabs/Azure/Cartesia TTS [1][3][6].
• Telephony: Twilio (including ConversationRelay), Telnyx, Sinch or a European carrier over SIP; orchestration with LiveKit Agents or Pipecat.
• Integration: function calling to calendar, CRM and billing; webhook for outcomes; consent store as a hard precondition of every dial.
• Choose B when: you need custom flows, deep integration or want to avoid per-minute platform margin.

Path C — Self-host (open source)

• Stack: LiveKit Agents or Pipecat for orchestration; Whisper/faster-whisper or Voxtral for ASR; an open LLM (Mistral, Qwen, Llama class) served with vLLM; open TTS such as Kokoro, Piper or Orpheus-type models; Asterisk or FreeSWITCH with an EU SIP trunk.
• Hardware: HW3 for several concurrent calls with a mid-size LLM and low latency; HW2 for a pilot with a small model and few parallel calls.
• Effort: high operational effort for latency tuning, speech quality (especially German and regional accents) and telephony. Worth it for strict data residency or very high call volumes.

Path D — Custom training

Rarely justified. Candidates: fine-tuned ASR for domain vocabulary and dialects, a voice clone of your own brand voice (with written consent of the speaker), or a small classifier for outcome and vulnerability detection trained on reviewed transcripts.

State of the art (2026)

• Realtime voice stacks reached production quality: native speech-to-speech and optimised cascades deliver turn latencies near human conversation, handle interruptions and call SIP trunks directly [1][3]. EU speech models (Voxtral) improved multilingual realtime ASR [7].
• Platform consolidation: specialist vendors bundle dialler, voice, LLM, analytics and compliance features; agents increasingly act through tools and MCP-style integrations.
• Regulation caught up: Article 50 AI Act transparency duties for systems interacting directly with natural persons apply from 2 August 2026, so a spoken AI disclosure at the start of the call is the standard design pattern [5][8].
• Limits and failure modes: ASR errors on names, numbers and accents; voicemail and IVR mis-detection; hallucinated promises (discounts, deadlines); prompt injection by the callee; emotionally difficult conversations with indebted or vulnerable people; low answer rates because recipients distrust unknown or automated callers. Quality is best on short, structured, low-stakes calls.

Shared resources & combinations

• Same stack as CONV-002 (inbound voice): ASR, TTS, LLM, telephony, guardrails and transcripts can be one platform serving inbound and outbound.
• CONV-003 supplies slot booking and calendar tools; CONV-007 receives escalated calls with the transcript as context; CONV-013 shares authentication and transactional rules in banking/insurance.
• Downstream: LANG-013 (sentiment and open-answer analysis of survey calls), SPEECH-001 (transcription tooling), AGENT-023 (order-to-cash, dunning strategy, payment matching).
• Hardware: HW3 inference server shared with other LLM/voice workloads.

Effort, cost & KPIs

• Path A: days to weeks per campaign; usage-based per-minute pricing plus telephony.
• Path B: 1-3 months incl. consent store, CRM integration, monitoring and legal review.
• Path C: 3-6 months; GPU server, telephony and on-call operations.
• Cost drivers: minutes per call, telephony fees, number of retries, quality assurance and legal review.
• KPIs: contact rate, completion rate, task success (confirmed appointments, valid survey answers), no-show reduction, escalation rate, opt-out and complaint rate, cost per completed outcome, call-sample compliance score, for collections: payment-plan adherence and dispute rate.

People & change

Start with management and the process owner (reception, accounting or survey lead) to fix purpose and tone. Collections calls need legal and compliance early, because wording, timing and consent rules are strict and customers must be told that they talk to AI (AI Act Art. 50). The DPO checks legal basis, call recording, AVV with the voice provider and whether a DPIA is needed. IT integrates telephony and CRM; staff who currently phone (end users) co-write scripts and take escalated calls.

Works council: depends. The agent mainly serves customers, but if it replaces or reorganises staff tasks, §90(1) no. 3 BetrVG requires timely information, and if call outcomes are linked to individual employee performance, §87(1) no. 6 applies. In the public sector the staff council is involved. Only relevant where a works council exists.

AI literacy: support the team with a short introduction to what the voice agent can and cannot do (AI Act Art. 4). Pilot with a low-risk case such as appointment reminders to a small, consenting group for 3-4 weeks before surveys or payment reminders. Typical concerns: annoyed customers, wrong statements about debts and job loss; address with clear AI disclosure, a one-key transfer to a human, no negotiation or threats by the agent and listening to sample calls weekly. Humans handle disputes and hardship cases.

Guides: GUIDE-005 (data protection), GUIDE-008 (change management), GUIDE-009 (pilot).

Risks, regulation & compliance

• EU AI Act: typically limited risk (Art. 50(1)): people must be told they are interacting with an AI system unless obvious; on a phone call this means a spoken disclosure at the start. Synthetic voice output must be machine-marked under Art. 50(2) where applicable; a voice clone of a real person used to deceive recipients is a deepfake issue and must be disclosed [5][8]. Deployers must provide AI literacy (Art. 4). Art. 5 prohibits manipulative or exploitative techniques that exploit vulnerabilities (age, disability, social or economic situation), which is directly relevant to debt collection scripts; and emotion inference in workplace settings. Credit-related use that evaluates creditworthiness is high-risk (Annex III 5(b)); a collections agent that only reminds and records a payment promise is generally not, but scoring who gets which treatment may be, so review classification.
• Telephone consent (Germany/EU): consumer marketing calls require prior express consent (§ 7(2) no. 2 UWG), documented and stored for five years (§ 7a UWG); B2B calls need at least presumed consent; fines up to EUR 300,000 are possible [2]. Automatic calling machines require express consent in any case. Other member states apply ePrivacy Directive Art. 13 national rules; check each market.
• GDPR: lawful basis per purpose (contract, legitimate interest, consent), transparency in the first seconds, data minimisation, retention limits for recordings and transcripts, right to object, DPA with platform and sub-processors, transfer safeguards for non-EU providers. Recording needs a legal basis and notice; health data (appointment reminders at special clinics) are special-category (Art. 9). Art. 22 GDPR applies if the call outcome automatically triggers legal or similarly significant effects.
• Debt collection: Germany regulates registered collection services (RDG) and consumer protection rules on dunning; fair-treatment duties; never threaten or mislead; no calls at night or excessive frequency; honour disputes and hand over.
• Mitigations: consent and do-not-call check before every dial, disclosure script, identity check, allow-listed tools, vulnerability-triggered handover, audit log, sampled human QA, kill switch.

Getting started

1. Pick one low-risk call type (appointment reminder or short service survey) and document the legal basis and consent source per contact.
2. Write the script policy: AI disclosure, purpose, identity check, escalation triggers, wording limits. Have legal and the data protection officer review it.
3. Pilot with 2 platforms (Path A) on 200-500 real calls; test voicemail, accents, interruptions.
4. Integrate write-back and opt-out handling; set call windows and retry caps.
5. Review a weekly transcript sample and track KPIs before expanding to surveys or payment reminders.

Sources

1. OpenAI — Realtime API with SIP (telephony guide) — https://developers.openai.com/api/docs/guides/realtime-sip
2. IHK Nord Westfalen — Werbung per Telefon, § 7 UWG consent — https://www.ihk.de/nordwestfalen/recht/rechtsthemen/wettbewerbsrecht/werbung-per-telefon-telefax-oder-e-mail-3614212
3. Microsoft — GPT Realtime API via SIP, Microsoft Foundry — https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/realtime-audio-sip
4. Retell AI — Vapi vs ElevenLabs 2026 (vendor comparison, pricing as reported) — https://www.retellai.com/blog/vapi-vs-elevenlabs
5. European Commission — Guidelines on transparency obligations (Art. 50) — https://digital-strategy.ec.europa.eu/en/policies/guidelines-ai-transparency-obligations
6. ElevenLabs — Agents platform — https://elevenlabs.io/api
7. Mistral — Voxtral Transcribe 2 and Voxtral Realtime — https://mistral.ai/news/voxtral-transcribe-2/
8. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
9. UWG online commentary on § 7 — https://www.uwg-onlinekommentar.de/kommentierung-7-uwg/