Inbound vs Outbound Voice AI Unit Economics

The economic boundary changes before the conversation starts

An inbound execution usually begins when the network offers a call. The business has already acquired or retained enough attention for the caller to act. The Voice AI system is primarily serving, routing, converting, or recovering that demand. Missed calls, queues, overflow, abandonment, and wrong transfers are central failure costs.

An outbound execution should begin when a recipient becomes eligible for a specific purpose—not only when a human answers. Audience selection, suppression, attempt scheduling, number and carrier behavior, machine detection, voicemail, retries, and opt-out handling are part of the production system. Ignoring them makes the conversation look cheaper by excluding the work required to create it.

Build two funnels before calculating ROI

Twilio's Call resource distinguishes inbound from outbound calls and exposes outbound states such as queued, ringing, answered, completed, busy, failed, and no-answer through callbacks. Amazon Connect's outbound campaign metrics separately track delivery attempts, human answers, interactions, progress, exclusions, and abandonment. These are useful operational stages; the economics layer must still join them to loaded cost and a durable business result.

The shared stack is only the middle of the cost equation

Outbound costs can exist before anyone speaks. Inbound costs can spike when many people speak at once. Reconcile provider invoices to execution events, then allocate shared platform and operations cost using a written rule instead of assuming the per-minute voice rate is the total cost.

Use channel-specific leading metrics and one shared outcome metric

Do not optimize human-answer rate independently. Aggressive pacing or weak audience rules can increase answers while degrading engagement, complaint rate, outcome quality, or customer margin. Likewise, high inbound containment can be harmful if callers reconnect or corrections move to people later.

Inbound and outbound Voice AI unit-economics formulas

Worked example: $3 inbound outcomes and $9 outbound outcomes can both be good economics

The following monthly values are illustrative—not benchmarks, customer results, provider prices, or guaranteed returns. Both channels produce the same terminal event: a durable completed appointment.

The outbound outcome costs three times as much, but that does not make it inferior. If the $52 contribution is genuinely incremental and the cohort is eligible, $43 remains after channel cost. Conversely, cheap inbound calls are not automatically valuable if appointments would have been completed through an existing low-cost path. ROI requires an appropriate baseline, not just a cheaper denominator.

Outbound economics include permission, suppression, timing, and abandonment controls

Outbound legal and policy requirements vary by jurisdiction, purpose, recipient, technology, and relationship. This guide is an economics framework, not legal advice. Treat qualified compliance guidance and your written policy as production dependencies.

The U.S. Federal Trade Commission's Telemarketing Sales Rule guidance describes calling-time restrictions, do-not-call processes, records, and call-abandonment controls for covered telemarketing. Those controls create real engineering and operating work, and failures can create costs much larger than the voice bill. Transactional reminders, customer-requested callbacks, collections, healthcare communications, and sales campaigns may have different requirements; do not copy one policy across every workflow.

Amazon Connect exposes campaign exclusions, attempts, human answers, voicemail classifications, interactions, and abandonment metrics. Twilio's answering-machine detection documentation explains that human, machine, fax, or unknown classification can be synchronous or asynchronous and that detection is imperfect. Store the result, duration, configuration version, and downstream outcome so misclassification cost remains visible.

The best program may use inbound and outbound as one lifecycle

Direction is an attribute, not a product category. An inbound caller may request a callback when the queue is long. An outbound reminder may cause the recipient to call back on a trusted published number. A failed payment call may move to SMS; a service request may move from AI to a scheduled human callback. Link these contacts under one objective and one observation window.

Without cross-session identity and objective linkage, the inbound team claims the conversion while the outbound team claims the contact. A shared execution graph prevents double counting and exposes the total cost of the customer journey.

Machine-readable telemetry needs direction-specific fields

Emit one event for every execution and join it to attempts, conversation turns, tools, human work, authoritative outcomes, revenue, and corrections. Keep contact and conversation content in the appropriate operational systems; economics telemetry can rely on stable IDs, categories, states, costs, and versions.

ganivra.track({
  "event_type": "voice_execution",
  "execution_id": "vx_01K5CHANNEL7",
  "customer_id": "customer_482",
  "direction": "outbound",
  "trigger_type": "customer_requested_callback",
  "workflow": "appointment_reactivation",
  "campaign_id": "campaign_2026_09",
  "attempt_number": 2,
  "eligibility_policy_version": "eligibility_v8",
  "suppression_check": "passed",
  "call_state": "completed",
  "answer_class": "human",
  "machine_detection_ms": 2860,
  "engagement_state": "meaningful_exchange",
  "source_system_ack": "appointment_created",
  "durable_outcome": "appointment_kept",
  "telephony_cost_usd": 0.19,
  "speech_model_tool_cost_usd": 0.31,
  "human_and_operations_cost_usd": 0.22,
  "provider_reported_cost_usd": 0.50,
  "loaded_execution_cost_usd": 0.72,
  "attributed_contribution_usd": 18.40,
  "data_classification": "no_contact_recording_transcript_or_conversation_content"
});

For inbound events, replace campaign and attempt fields with offered timestamp, number or location category, queue, coverage window, intent, containment, escalation, and repeat-contact status. Ganivra's event integration connects both directions to models, tools, customers, outcomes, revenue, and margin without requiring caller contact details, recordings, transcripts, or conversation content.

How to measure inbound versus outbound Voice AI economics

1. 01
   Declare direction and trigger

   Record inbound or outbound, the event that initiated the call, the workflow, customer, location or campaign, and the policy version.

2. 02
   Define separate funnels

   For inbound, start with offered calls and valid intents. For outbound, start with eligible recipients and attempts, then distinguish machine, human, engagement, qualification, and outcome.

3. 03
   Attribute the full cost stack

   Join telephony, speech, model, tools, retries, voicemail, handoffs, quality, compliance operations, correction, support, and allocated platform cost to one execution.

4. 04
   Verify outcomes in source systems

   Require CRM, scheduler, claims, payment, dispatch, or other authoritative-system acknowledgement and define a durability window.

5. 05
   Measure incrementality and risk

   Use a credible baseline or holdout, exclude ineligible traffic, and retain opt-outs, complaints, wrong-party contacts, corrections, and repeat calls.

6. 06
   Optimize customer margin by segment

   Compare direction, intent, campaign, daypart, geography, carrier, model, voice, prompt, tool, attempt number, and customer without blending unlike denominators.

Start with one workflow and one written terminal outcome. Reconcile a sample of provider invoice lines, system acknowledgements, repeat contacts, and human work before scaling. Continue with the Voice AI unit economics guide, Voice AI pricing models, answered-call metric guide, human-handoff economics guide, and Voice AI timing-cost guide.

Frequently asked questions

Inbound vs outbound Voice AI unit economics FAQ