---
title: "How to Keep a Voice Agent on Script"
description: "Keep a phone agent on script with workflow steps, test each draft with simulated callers, and let an AI judge flag made-up facts on each real call."
canonical: "https://speko.com/keep-voice-agents-on-script"
last-updated: "2026-11-19"
---

> ## Speko page index
> The complete index of every page on this site is at: https://speko.com/llms.txt
> Read it before exploring further. It lists the exact Markdown URL for every canonical HTML page.

# How to Keep a Voice Agent on Script

Steps instead of one long prompt, tests before go-live, checks after.

Speko is a voice agent platform that keeps a phone agent on script in three layers. The call runs as workflow steps, and you test each draft with simulated callers. After each real call, an AI judge checks it for broken rules and made-up facts, also called hallucinations.

[Get started free](https://platform.speko.ai/sign-in) [Book a demo](https://speko.com/book)

## Why Does a Workflow Keep an Agent on Script?

[Workflows](https://docs.speko.ai/concepts/workflows)

- A long prompt asks a model to remember its agenda while it improvises. A workflow puts the agenda in the structure, so the same call runs the same way twice.
- On a standard agent, each step says its line word for word, stores the answer in a typed slot and branches on what the caller said.
- Mark a question the caller must hear. If the caller never heard it, the call goes to a step you choose instead of moving on.
- Set a global intent, such as cancel or ask for a person, and it is live at every step you do not exclude. When the caller's reply does not answer the step's own question, a matching intent sends the call to the step you set for it.

## How Do I Test for Made-Up Answers Before Go-Live?

[Test before you deploy](https://docs.speko.ai/agents/testing)

- Speko writes 4 to 12 test cases from the agent's instructions, tools and knowledge.
- AI-judged checks, such as No made-up facts, flag invented facts and broken rules. They are advisory.
- Each case shows Pass, Fail, Inconclusive or Incomplete. Open a result to hear the call and read the evidence for each check.
- When an AI-judged check fails, Improve the prompt drafts a fix for you to review.

## How Do I Catch Problems on Real Calls?

[Monitor real calls](https://docs.speko.ai/agents/monitoring)

- After each real call ends, an AI judge checks it for made-up facts, broken rules, unanswered questions, repeating loops and bad endings.
- Speko also fails a call from its record when the caller waited 8 seconds or more for a reply, the call dropped or the audio failed.
- Alerts go to Slack, email or a webhook when calls start failing.
- A failed call becomes a test case in one click.

## Do Benchmark Scores Show How My Agent Will Behave?

[LLM benchmark](https://speko.com/benchmarks/llm)

- Speko's public LLM board measures how often a language model invents a policy or fact it was never given.
- The benchmark ranks models, never agent platforms.
- A model score does not tell you how your agent will behave, so test the agent on your own script.

## What Are the Limits of Keeping an Agent on Script?

- Speko publishes no agent-level rate for made-up facts or script adherence yet.
- AI-judged checks are advisory and never block a deploy; only rule checks do.
- Blocking deploys on a failed check is off by default.
- Realtime agents say each line in their own words, not word for word.

## Notes

- Every fact is from the Speko page linked beside its section, checked November 19, 2026.
- The product panel at the top is an illustration with made-up data.

## Questions

### How Do I Keep a Voice Agent on Script?

Build the call as workflow steps on a standard agent, where each step says its line word for word and stores the answer in a typed slot. Test each draft with simulated callers before it ships. After each real call, an AI judge checks it for broken rules and made-up facts.

### How Do I Test a Voice Agent for Hallucinations Before Go-Live?

Generate test cases from the agent, add a caller who asks about something the agent was never told, and pick the No made-up facts check. Open each failed case to hear the call and read the evidence. Simulated test runs are free.

### Can a Long Prompt Keep an Agent on Script?

A long prompt asks the model to remember its agenda while it improvises. A workflow puts the agenda in the structure, so the same call runs the same way twice.

[Get started free](https://platform.speko.ai/sign-in) [Book a demo](https://speko.com/book)

You get $5 in credit when you sign up. Calls then cost $0.09 per minute, plus carriage on Speko numbers.

Related

- [AI Phone Interviews and Candidate Screens](https://speko.com/ai-phone-interviews)
- [AI Intake Calls for New Clients and Patients](https://speko.com/client-intake-calls)
- [AI Phone Surveys and Follow-Up Calls](https://speko.com/ai-phone-surveys)
- [How to Move a Voice Agent From Vapi or Retell](https://speko.com/migrate-voice-agent)
- [Speko Agents](https://speko.com/agents)
- [Voice agent testing with simulated callers](https://speko.com/voice-agent-testing)
- [Voice agent monitoring for every real call](https://speko.com/voice-agent-monitoring)
- [Voice agent workflow builder](https://speko.com/voice-agent-workflow-builder)
- [Multilingual voice agents](https://speko.com/multilingual-voice-agents)
- [Voice agent API and SDKs](https://speko.com/voice-agent-api)
- [Voice agent platforms compared](https://speko.com/voice-agent-platforms)
- [Vapi alternatives](https://speko.com/vapi-alternatives)
- [Voice agents by business](https://speko.com/solutions)
- [Speko pricing](https://speko.com/pricing)
