
September 2023 Releases
Successful outbound AI agents can navigate unfamiliar IVRs. An IVR (Interactive Voice Response) is the automated phone menu you hit when you call many companies today: "Press 1 for billing, press 2 for support". However, these menus were built for humans, not AI agents. That's exactly why teaching your AI agent to navigate them is a less intuitive, but necessary use case.
This is common among healthcare and insurance companies, where AI agents have to get routed to the correct provider, doctor, or pharmacy. For example, an agent may have to call insurance companies on behalf of a policyholder to check claim status. It needs to get through the IVR (billing, then claims, then plan type) to reach someone who can answer.

Most people think of AI agents replacing IVRs, but until every company retires theirs, your AI agents should be capable of successfully navigating them. Without successful IVR navigation, it can be put into the wrong queue, or disconnected, wasting a dial and leaving the claim unresolved. These IVRs can be organized in unexpected ways, which is why we've built testing and prompting for AI agents who frequently run into them.
Our AI agents can:
Differentiate between IVRs, answering machines, and human pick-ups, so outbound calls are routed appropriately.
Navigate a known IVR on your terms: tell your AI Agent how to move through the menu, when to speak, and which digit gets it to the outcome you want.
Isolate that logic into a single node in a multi-state agent, so IVR handling doesn't affect other parts of your agent in conversation flows that require conditional logic.
Test against real call transcripts, generating test cases anchored in your specific use case to improve your agent over time. With this approach, we recently helped a customer get their AI agent's IVR navigation pass rate to 96%.
To prepare an AI Agent for a known phone tree, each stage becomes its own state, with its own logic and its own Press Digit action. That action tells the agent what to press, when to press it, and how long to wait after silence before acting. In Regal, you configure:
.png)
Above is an example prompt of the Press Digit action, which defines a goal for an AI agent to appropriately route to the caller they'd like to connect to.
IVR navigation prompts often sit inside use cases with complex, conditional flows: an insurance verification call, a scheduling flow, a multi-step intake. Isolating IVR navigation into a single node within a Multi-State Agent lets you handle it as a first, self-contained step.

As prompts get more complex, the node responsible for pressing digits and reading a menu is isolated from the rest of the agent's logic once the call connects. IVR navigation becomes a clean handoff into the actual conversation, not a tangle of conditions.
IVRs vary by carrier, by region, by use case, and by industry. Some providers test their AI agents against synthetic phone trees, which are generic IVR structures built in a lab to represent common patterns and quirks across industries. These simulated trees are useful for benchmarking an agent's general navigation skill and catching obvious bugs, but they're built to approximate what an IVR might look like, not to reflect the actual menu your agent will hit on a real call. That gap makes them an incomplete stand-in for improving IVR navigation over time.
Rather than relying on one-size-fits-all synthetic environments, Regal users build test simulations grounded in the real IVRs their AI Agents actually encounter, informed by transcript analysis of live test calls in Copilot.

For example, for a pharmaceutical client, we recently built test cases based on the specific IVR patterns and edge cases that the client's agent was running into in the real world.
The agent was struggling with an edge case where there were lengthy disclaimers in the IVR before the press-digit prompts, and it wasn't clear when to jump in. Copilot suggested a prompt adjustment: stay silent through the disclaimer, and press the digit only once the actionable prompt starts. Ultimately, updating the agent's prompt with suggested improvements, anchored in real-world IVRs, brought their AI agent's IVR navigation pass rate from 80% to 96%.
The same agent likely wouldn't have improved as fast against generic, one-size-fits-all synthetic IVRs. The gains came from test cases and prompt adjustments based on the IVR patterns that were specific to the companies and industries the agent operated in.
Eventually, we expect a lot of these phone trees to be replaced outright by AI agents that understand natural language instead of requiring a digit press. When that happens, navigation stops being about menus and starts being about two AI agents talking to each other.
However, it is likely that evolution will happen unevenly. Financial services, insurance, and healthcare all carry compliance guardrails that slow how fast they'll trade a known, auditable phone tree for something less predictable. Until the industry moves, it's important that your AI agent navigates IVR trees more effectively, over time.
IVR navigation is an AI agent's ability to work through an automated phone menu, the "press 1 for billing, press 2 for support" system most companies still use, and reach the right destination. These menus were built for humans to navigate by ear, not for AI agents, which is why teaching an agent to handle them is a distinct, necessary capability.
Most people assume AI agents will simply replace IVRs, but that shift will take time, and until every company retires its phone tree, outbound AI agents still have to get through the ones that remain. An agent calling an insurance company to check a claim status, for example, still has to navigate billing, then claims, then plan type before it reaches someone who can answer.
A failed IVR navigation can land the agent in the wrong queue or get the call disconnected entirely, wasting the dial and leaving whatever task it was calling about, like a claim status check, unresolved. That's why classifying what picked up the call and knowing how to move through the menu correctly matters as much as the conversation that follows.
Press Digit is a built-in action that tells a Regal AI Agent what to press, when to press it, and how long to wait after silence before acting, configured per stage of a known phone tree. Builders set a Name to reference the action, a Description explaining when it should fire, and a Pause Detection Delay in milliseconds so the agent waits for the IVR to finish speaking before it presses.
Synthetic phone trees are generic structures built in a lab to approximate common IVR patterns, useful for catching obvious bugs, but they don't reflect the actual menu an agent will hit on a real call. Regal instead builds test simulations grounded in the real IVRs an agent actually encounters, using transcript analysis of live test calls in Copilot, which is what took one pharmaceutical client's IVR navigation pass rate from 80% to 96%.
Ready to see Regal in action?
Book a personalized demo.



