What actually works today

• WhatsApp Business voice calling + AI: Meta is enabling voice pipelines for large business accounts via API, making it feasible to run AI voice agents for support and sales inside a channel customers already trust. citeturn2search6
• Real deployments: 911 centers use AI voice to triage non‑emergency calls; dealerships handle scheduling and parts inquiries with agents. These are not demos—they’re in production. citeturn2search0turn2search1
• Latency is viable: Orchestrators report sub‑700 ms model latency over web calls, plus ~200 ms for phone connectivity—good enough for natural turn‑taking. citeturn2search2
• Builder tooling is maturing: OpenAI’s AgentKit introduced agent evals, connectors, and a visual builder; the Responses API and Agents SDK made multi‑step agents and eval workflows easier. citeturn0search0turn0search4

Before you build: constraints and compliance

Robocalls & disclosure (U.S.): The FCC formally clarified that AI‑voiced robocalls fall under TCPA restrictions; treat consent and disclosures seriously, especially for outbound. Use branded introduction lines and capture opt‑ins. citeturn2search4

WhatsApp policy changes: Reports indicate that starting January 15, 2026, WhatsApp will ban general‑purpose chatbots via the Business API. Business workflow automations (support, booking, order status) are still permitted—so design your agent around specific, transactional workflows rather than open‑ended assistant behavior. Plan migrations accordingly. citeturn3news14turn3news12

Reference architecture for a safe, measurable voice agent

1. Telephony & channel: SIP/phone (e.g., Twilio) and/or WhatsApp Business voice calling API. Route to a voice orchestration layer. citeturn2search6
2. Orchestration & state: A runtime that coordinates STT → LLM/tool use → TTS, with conversation state, retries, and handoff rules. Sub‑700 ms latency targets keep conversations fluid. citeturn2search2
3. Speech stack: Best‑in‑class STT/TTS paired with a modern LLM. Start deterministic for payments or account changes; open‑ended for FAQs.
4. Tools & data: Connect order APIs, CRM, ticketing, and knowledge bases. Use retrieval for long‑tail answers and function calls for transactions.
5. Evals & observability: Scenario tests, trace‑based grading, and business KPIs. OpenAI’s agent evals help automate acceptance gates; instrument the stack with OpenTelemetry. citeturn0search0
6. Safety & governance: Blocklist/allowlist, rate limits, human handoff, and transaction scopes. Log identity assertions and permission checks on every step.

Deep dives from our library:

• Agent Observability with OpenTelemetry and business KPIs
• Unit economics and cost control for agents
• Stop agent impersonation: identity, permissions, transaction controls
• Red‑team your agents before go‑live
• A2A + MCP interoperability blueprint

A 14‑day launch plan (from pilot to production)

Days 1–2: Scope, compliance, and success metrics

• Pick one high‑volume intent: where’s my order, order changes, appointment scheduling, or returns.
• Decide in‑policy channels (inbound phone, WhatsApp Business voice) and confirm consent language for any outbound follow‑ups. citeturn2search4turn2search6
• Define KPIs: answer rate, first‑call containment (no agent handoff), average handle time (AHT), CSAT, and cost per resolved interaction.

Days 3–5: Wire up the plumbing

• Provision telephony (SIP/number) and/or WhatsApp Business voice calling. Route to your orchestrator.
• Stand up the speech + LLM stack; set strict function permissions for account lookups, order edits, and refunds.
• Add a human handoff to your helpdesk for confidence dips or compliance triggers (e.g., payments). Consider Zendesk handoff patterns.

Days 6–7: Build your happy paths

• Draft short, friendly call flows: greeting → purpose → authentication → task → confirmation → post‑call summary.
• Prepare structured prompts and small talk guardrails; keep replies concise (7–12 seconds).
• Implement fail‑safes: three misrecognitions → route to human; any payment/identity ambiguity → pause + human.

Days 8–9: Evals and acceptance tests

• Use agent evals to score scripted scenarios (intent, auth, tool use, policy adherence). Gate deployment on threshold scores. citeturn0search0
• Run 50–100 synthetic calls plus 20 real calls with staff. Review traces and fix top 5 error modes.

Days 10–11: Observability, cost, and safety

• Ship OpenTelemetry spans for each step (STT, intent, tool call, TTS). Dashboard answer rate, containment, AHT, CSAT, and cost/call. See our observability blueprint.
• Apply identity & permissions playbook (role scopes, secure secrets, transaction ceilings). Reference: impersonation controls.
• Model unit economics: speech + LLM + orchestration + telephony; target ≤ $0.40/resolved inbound call for FAQs, higher for transactional flows. See cost control playbook.

Days 12–14: Pilot, iterate, and expand

• Soft‑launch during business hours to ensure fast human backups.
• Iterate on top failure reasons (auth fallbacks, noisy environments, accents). Note: production users have successfully scaled in noisy, variable environments like dealerships and public call centers. citeturn2search1turn2search0
• Expand intents once containment ≥ 60% and CSAT ≥ 4.4/5 with net savings.

KPIs that predict ROI

• Answer rate: ≥ 98% during staffed hours; ≥ 95% off‑hours.
• Containment: 40–70% (FAQ and order status at the high end; complex account changes lower).
• AHT: 2–4 minutes for FAQ; 4–7 minutes for transactional flows.
• Cost per resolved call: Driven by STT/LLM/TTS plus telephony. See strategies to trim 30–60% using caching and routing in our unit economics guide.

Guardrails you shouldn’t skip

• Disclosure & consent: Open with a branded disclosure that the call uses an AI agent and how data is used. Respect opt‑out. citeturn2search4
• Authentication tiers: Lightweight (order email/ZIP) vs. sensitive (OTP, vault‑backed tokens). No changes without verified identity.
• Transaction ceilings: Refund caps and manager approval flows; log every permission check. See transaction controls.
• Handoff to humans: Confidence dips, policy boundaries, or user request → immediate warm transfer.
• Red‑team before launch: Jailbreak attempts, prompt injection via DTMF or background audio, identity spoofing. Start with our 2025 red‑teaming playbook.

What’s next (and what to park)

Now: Inbound support and scheduling on phone/WhatsApp; outbound callbacks to consented lists; multilingual handling with careful testing. citeturn2search6turn2search4

Park for later: Open‑ended, general‑purpose WhatsApp agents. Reports of WhatsApp policy changes starting January 15, 2026, mean you should keep flows task‑specific and be ready to adjust. citeturn3news14turn3news12

Template: Go‑live gate

1. Evals pass rate ≥ 90% on scripted scenarios; zero critical policy violations. citeturn0search0
2. Observability live with OpenTelemetry; alerts for error spikes and long turns.
3. Documented disclosures and opt‑out; TCPA alignment for outbound. citeturn2search4
4. Containment ≥ 50% on pilot cohort or clear savings vs. baseline.
5. Warm handoff SLAs ≤ 30 seconds to human for escalations.