Quick answer: Route a call to a scripted IVR when the query has a fixed answer you can pre-record (hours, location, payments), to an AI voice agent when callers use natural language and need resolution rather than navigation, and to a human agent when the case is high-stakes, ambiguous, or explicitly flagged by the caller. On Devotel Orbit all three live on one call path, one queue, and one bill, so the routing choice is a flow question, not a platform question. The decision table below grades each call type; the rest of the guide covers the migration, the cost model, and the failure catalog that grades the rollout.
Most organizations don't have to pick one of the three. The common mistake is treating this as a vendor-selection question when it is really a routing-design question: each inbound call has a shape, and the right handler depends on that shape. Devotel Orbit ships a visual IVR builder, AI voice agents, and a human contact center as one platform, so the routing decision is a configuration you can change per intent without repurchasing anything.
The decision table: scripted IVR vs AI voice agent vs human handoff
| Call type | Route to | Why |
|---|---|---|
| "What are your hours / where are you located" | Scripted IVR (announcement) | The answer never changes; a menu reads it back in seven seconds with no model spend |
| "I want to reschedule / check a payment status / update my address" | AI voice agent | Natural-language intent resolution; the agent calls the tool, confirms, and closes |
| Regulatory / legal / complaints / vulnerable-caller flag | Human agent (queue) | High-stakes, ambiguous, or flagged callers escalate by rule, not by classifier judgment |
| "I don't know what I want, explain my options" | AI voice agent | Menu trees shed callers when the category is unclear; a conversational agent narrows options without abandonment |
| Caller says "agent" or "representative" | Human agent (queue) | The explicit escape phrase must land on the human path every time — a tenant-set rule in Orbit's contact center |
| Payment detail that matters (a missed payment, a refund) | Human agent after agent attempts | Emotional or financial weight belongs with a person; the agent gathers context first and hands it warm |
| Information complete on the first utterance | AI voice agent | The agent resolves on the partial answer; if the first utterance decides it, the IVR announcement is also fine |
| Overflow after-hours / weekend | AI voice agent | A human queue is empty; the agent either resolves or captures a numbered voicemail for callback |
Two things fall out of this table. First, the DTMF menu still earns its keep on the information-replay row: an announcement reads a fixed answer at the lowest unit cost, so you don't pay model tokens to say "We're open until 6 PM." Second, the explicit-escape row is a routing rule, not a classifier bet: "say 'agent'" and "press 0" must both land on the human queue, regardless of what the agent thinks the caller meant. On Orbit the visual canvas draws that rule once and applies across all three handler types.
Read this table against your actual call mix
Grade the table above against your recordings, not your intuition. Pull the last 60–90 days of transcripts and group them by caller language, then map each group to one of the rows above. The distribution will almost never be uniform: most organizations find a small set of intent clusters carrying the majority of volume, and the routing decision concentrates on those top intents rather than the whole tree. The migrate from IVR guide walks the catalog phase in detail; the same discipline applies here at the decision-design stage.
If the transcripts show that your top intents are ambiguous ("I don't know what to press"), the AI voice agent row dominates. If they're pure information-replay ("hours," "location," "balance"), the IVR announcement row dominates. If they're escalation-heavy ("the bill is wrong," "I want to complain"), the human-handoff row dominates. The decision table only grades correctly when the input column is real call language.
Migration rollout pattern
The routing design lands through four gates, each producing evidence before earning more traffic. This is the same pattern the full-stack migration and the IVR-to-agent migration both use, applied here at the single-decision level.
- Catalog the menu. Read what callers actually say and where they actually end up; the menu tree and the transcripts together give you the intent inventory.
- Shadow the agent. Put the AI agent behind the menu as a new option ("describe what you need") while the existing menu keeps its baseline. Run it on real traffic long enough to measure containment per intent.
- Shift top intents first. Promote the agent to the front of the call flow for the intents it resolves at least as well as the menu. The menu keeps the long tail.
- Route humans through the contact center. Any call the table grades as human-handoff goes straight into the same skills-based queues your human agents already work. The agent hands off with a summary and verified intent attached, not just a cold transfer.
Each gate is graded on contained-call share and handoff context quality. If the agent performs worse on an intent, that intent stays on the menu or moves straight to the queue. Nothing in the pattern requires taking the menu down while you decide.
The cost model
The three handler types cost differently per call, and the decision table grades them at different unit economics:
- Scripted IVR announcement. Fractions of a cent per call; the cheapest option when the answer is fixed.
- AI voice agent. Metered per talk-minute across four components: speech-to-text, the language model, text-to-speech, and telephony. A typical support or booking configuration runs roughly $0.05–$0.30+ per talk-minute all-in; the full breakdown is in AI voice agent pricing in 2026. ROI is cost per resolved conversation, scaled by the share of calls the agent closes without escalating.
- Human agent. A fully loaded talk-minute of roughly $0.50–$1.50+, so any share the AI agent contains shifts the queue's economics materially.
The decision table orders the handlers by cost ascending: start with the IVR option, then the agent, then the human. The goal is not to maximize agent containment at any cost; it's to route each call to the cheapest handler that actually resolves it. The pricing guide works the cost-per-resolved-conversation math end to end.
Failure modes to budget for
The rollout earns its keep only if each handler fails in its own way and you can see it. The AI agent failure catalog names the six failure modes a tool-calling agent hits; four of them map directly to the routing decision table above:
- Over-wide credential (mode 1). An agent that rides an ambient session token wider than the task can reach systems it shouldn't. The agent handler must exchange tokens scoped to the task before it runs.
- Deterministic gate (mode 2). A malformed input fails identically every retry; the flow must route that call to the human queue rather than looping the agent.
- Unpinned prompt (mode 4). A routing quality regression must trigger rollback to the last known-good agent version, not a menu-only fallback.
- Runaway cost (mode 5). A looping agent burns the per-conversation ceiling; when the ceiling trips, the routing rule must escalate to human, not hang up.
The routing table survives a failure in any single handler because the fallback path is designed, not blank. The intent-classification versus code-routing post covers the fallback-edge behavior in the flow builder: when nothing matches, the run takes the edge you drew. A routing design that has no drawn fallback is a routing design you will eventually pay for.
Frequently asked questions
Do I have to pick one of the three and deprecate the others?
No. The point of the routing guide is that every call has a shape and the handlers cover the shapes together. The IVR announcement row still has a purpose, the agent row carries the bulk of resolution, and the human row handles escalation. Deprecating the IVR to go all-agent or deprecating humans to go all-agent are both anti-patterns the table flags.
How do I know a call is "high-stakes" enough for a human?
Two reliable signals: the caller uses a flagged phrase ("complaint," "regulatory," "cancel everything," "my account was emptied"), or the caller says "agent" / "representative" outright. The explicit-escape rule in Orbit's contact center treats both as routing instructions before the classifier even scores the intent. A vague "cancel" from a caller with a past complaint history is also worth routing to humans on the flag alone.
My current system only has a menu tree. Can I add an agent without replacing it?
Yes. The IVR-to-agent migration guide uses the shadow pattern above precisely because the same visual builder in Orbit hosts the DTMF menu and the AI voice agent on one call path. You add the agent as an option behind the menu, run the side-by-side measurement, and promote per intent. One flow, not two systems.
What happens when the AI agent can't resolve a call?
The caller escalates warm into the contact center queue with a call summary and the verified intent attached. The person who picks up starts from what was already confirmed, not "how can I help you." That escalation context is pinned before the agent goes live as a routing rule, and the fail-safe fallback after an agent failure is the menu's announcement path, not a hang-up.
Is the AI agent's containment measured per-intent or per-queue?
Per intent, side by side with the menu's containment for that intent. Queue-level metrics hide where the agent is underperforming; intent-level containment is the number the migration gates on. The routing table above is graded on per-intent containment share, not on total calls handled.
Where to go next
The routing decision is the front door; once the decision is made, the next stop is the surface that carries the handlers:
- Read the IVR product overview for the visual builder, DTMF and spoken-intent menus, and the agent handoff surface.
- See voice AI agents for the agent handler surface, including the sample-agent previews and the latency budget.
- Read the AI voice agents pillar page for the shipped capability surface the agent row draws on.
- Review AI voice agent pricing in 2026 for the cost-per-resolved-conversation ROI framework the routing decision grades against.
- See the contact center for the human-handoff surface: skills-based queues, SLA targets, wrap-up codes, escalation context.
The routing guide's output is a designed decision table you can put in front of your own callers, with the fallback drawn and the gates evidence-driven. On Orbit the whole path is one bill, one queue, and one flow canvas, so changing the routing is a configuration decision, not a re-purchase.
Published 3 September 2026.