Best Practices

Customer flows

Improv vs. Scripted

The single biggest decision in a customer flow is the authoring mode, and it maps directly to what you’re testing. Improv — You write a free-text brief describing who the customer is, what they want, and how they behave. The simulator improvises a fresh conversation from that brief on every run — different words, same intent:

A first-time caller wants to book an appointment for tomorrow around 2pm.
They're friendly but a little unsure of the process and will ask what
information they need to provide.

Improv is the default choice for most testing. Because the wording changes each run, it exercises your agent’s ability to handle natural language variation for the same intent — the thing that actually breaks in production. Scripted — You author the conversation step by step on a graph canvas, writing the exact customer lines:

"Hello, can I make a booking for tomorrow at 2pm?"

Scripted mode is for precise control: verifying your agent parses a specific date format, detects a keyword, handles a DTMF sequence, or recovers from a deliberate silence.

Use Scripted when the exact utterance matters — phrasing, keyword detection, slot filling, keypad input. Use Improv when you want varied, realistic conversations that test the same intent from a different angle every run.

Write improv briefs like backstories

The customer setup brief is where the simulated customer comes to life. Behavioral color — emotional state, context, quirks — belongs in this prose, not in the persona (personas carry voice and speech characteristics; the brief carries the situation). Good briefs give the simulator context that drives nuanced, realistic behavior: Bereaved customer

James recently lost his wife and is calling to cancel her phone line on
their shared plan. He is soft-spoken and may become emotional. He doesn't
fully understand the account details and may need things explained gently.

Skeptical professional

Priya is a software engineer who immediately suspects she's talking to an
AI. She will ask pointed questions like "Are you a real person?" and "Can
you transfer me to a human?" She becomes frustrated if the agent can't
directly answer her questions about a billing discrepancy.

Impatient multitasker

Carlos is calling during a short break at work. He has 5 minutes. He'll
give short answers, may mishear things, and will ask the agent to repeat
or speak up. He needs to reschedule a delivery that requires a signature.

Structure scripted flows around paths

Scripted flows are graph-based — a DAG of agent and customer steps. Every unique path from start to leaf becomes its own variant automatically: the path engine derives one variant per path, so structuring the graph well is how you get coverage without duplication.

Start with a happy path

Begin with a single expected route — the happy path where everything goes as planned:

Add branches for edge cases

Once your happy path works, branch at points where the conversation can diverge. Branches inherit everything above them, and each new path shows up as an edge case in the variant rail: Focus your branching where:

The agent asks the customer a question (customers respond in unexpected ways)
The agent could go on a tangent or lose track of the conversation
Tool calls or lookups might fail or return unexpected results

If you want the same path exercised by a different caller, don’t duplicate the branch — use Add variant for path to run that path again with a different persona or environment.

Structural graph edits (adding or removing steps and edges) pause variant editing until you save — the flow needs to refresh its paths before you can add or remove variants.

Compose flows instead of duplicating them

Two composition tools keep shared sequences in one place:

Flow link (scripted step) — splices another scripted flow’s steps in at that point. Keep a shared IVR tree or authentication sequence in its own flow and link to it; when the menu changes, you update it once.
Preceded by (improv variant) — runs a scripted flow before the improv segment. Use it to deterministically navigate an IVR menu, then hand off to an improvised conversation. Improv has no fixed ending, so it can only come last — you can’t link a scripted flow after it.

Templating with variables

Variants carry key/value variables that are handed to the customer-side model, and text fields support {{variableName}} references:

"Hi, my name is {{patientName}} and I need to reschedule
my {{appointmentType}} appointment"

Persona-scoped properties use the {{persona.*}} prefix and resolve from the variant’s persona at runtime:

"Hi, my name is {{persona.name}}"

This keeps a single flow reusable across many test cases, with only the key details changing between runs. You can also use Apply a test profile in the variant’s variables editor to fill the entries from a saved profile in one click. See the Variables guide for the full lifecycle.

Generate flows with Ask Roark

You don’t have to author flows from scratch. The New flow page (/customer-flows/new) offers three Ask Roark starting points under “Generate”:

From your calls — pick real production calls and let Ask Roark draft flows from them. This is the fastest route to representative coverage: the drafts reflect how customers actually talk to your agent.
From a transcript — upload or paste a transcript from any source (another platform, a QA review, a bug report) and generate a flow from it.
Describe what you want — describe the situation in plain language and Ask Roark drafts the flow for you.

Whichever card you start from, treat the draft as a baseline: review the generated brief or steps, then extend it with edge cases covering paths that didn’t occur in the source material but could happen in production. Generated variants carry a Generated badge until you review them.

Personas

Personas model who is calling — the voice and speech profile pinned to each flow variant. A good persona strategy tests your agent across a range of realistic caller profiles.

Diversify voice and speech

Build a set of personas that vary:

Language and accent — battle-test your transcriber’s accuracy across accents (US, British, Indian, Spanish, and more) and languages
Speech pace and response timing — slow, rambling talkers and quick, clipped ones ensure your agent neither interrupts customers nor times out waiting for them
Clarity and disfluencies — vague or rambling callers with natural ums and false starts stress your agent’s understanding far more than clean studio speech

Background noise is no longer a persona property — it’s the variant’s Environment chip in the flow editor (Office, Coffee shop, Driving, Airport, and more). Vary it per variant to verify your endpointing holds up in non-ideal audio.

Test difficult customer types

Combine persona fields (base emotion, clarity) with a matching customer setup brief to build challenging callers:

AI Skeptic

Base emotion Skeptical, plus a brief where the customer tests the agent with trick questions and asks for a human

Hostile Caller

Base emotion Frustrated, plus a brief with a rude, escalating customer — verify your agent stays polite and professional

Sensitive Situation

A brief describing a customer in a difficult moment (bereavement, financial hardship) — ensure your agent is empathetic and considerate

Rapid Switcher

Base emotion Distracted, clarity Rambling, plus a brief where the customer changes topics frequently

The persona sets the delivery; the variant’s customer setup prose sets the situation and behavior. Keeping them separate means one “Frustrated rambler” persona can stress-test every flow in your library.

Plan configuration

A plan composes agent endpoints, attached flows (with a variant selection per attachment), metrics with Pass/Fail checks, and run settings. How you configure it depends on what you’re testing.

Common patterns

Load testing

Use the Load testing template rather than configuring this by hand. Its Volume panel exposes exactly the two knobs you need — Concurrent calls and Total iterations — plus the flow to test.

Pick a flow whose happy path matches your target call duration
Set total iterations to your target volume and concurrent calls to how many should hit the agent simultaneously

This reveals how your agent performs under peak load without over-provisioning your regular test plans.

Adaptability across personas

Test how your agent handles different voices, accents, and speech styles on the same conversation.

Attach a single flow with the default variant selection (your happy path)
Attach it multiple times with a different persona override per attachment, covering a wide spread of accents, paces, and base emotions

This isolates persona-driven variation from flow complexity, making it easy to spot which caller profiles cause problems. For language coverage specifically, the Multilingual template does the fan-out for you — each attached flow runs once per selected language.

Instruction following

Agents are non-deterministic — verify they don’t go off-script or hit loopholes.

Build flows with multiple edge cases: scripted branches for the paths you can enumerate, improv edge cases for the ones you can’t
Write agent expectations for the behaviors that must hold on every variant, and additional expectations per edge case
Focus edge cases on points where the agent might go on a tangent or fail to recover

The Flow adherence template ships with metrics and checks tuned for exactly this.

Red teaming

Use the Red teaming template to test resilience against adversarial inputs. It sources flows carrying the Adversarial label from your library, or generates 3–10 adversarial edge cases for you at easy, medium, or hard difficulty — covering prompt injection attempts, PII extraction, and social engineering.Label your own hand-authored adversarial flows with the Adversarial system label so the template picks them up automatically.

See Templates for the full catalogue and what each one preconfigures.

Keeping simulations under control

Setting	Default	Recommendation
Max duration	15 minutes (up to 24h)	Set to ~110% of your average call duration. Prevents runaway calls if the agent goes on a tangent.
Silence timeout	30 seconds (5–300s)	Ends calls after sustained silence, catching cases where either side stops responding.
End-call phrases	`goodbye` (max 10)	Specific phrases that end the simulation immediately when matched. Add phrases that signal a call has gone off track.
End-call reasons	none (max 10)	When you can’t pin down exact phrases, describe the condition instead — an LLM evaluates each turn and ends the call when the reason is met.
Iterations	1 (max 100)	Runs per test case. Fixed at 1 for outbound runs.
Concurrency	5	Keep at the default or below for regular runs to avoid unnecessary load on your agent and manage costs. Reserve high concurrency for the Load testing template.

Next Steps

Customer Flows

Author improv briefs and scripted graphs

Personas

Create diverse caller profiles

Templates

Start from a goal with preconfigured plans

Run Plans

Configure and execute simulation plans

Getting Started

Observability

Metrics

Simulations

Recipes

Integrations

SDKs & Libraries

Resources

Customer flows

Improv vs. Scripted

Write improv briefs like backstories

Structure scripted flows around paths

Start with a happy path

Add branches for edge cases

Compose flows instead of duplicating them

Templating with variables

Generate flows with Ask Roark

Personas

Diversify voice and speech

Test difficult customer types

AI Skeptic

Hostile Caller

Sensitive Situation

Rapid Switcher

Plan configuration

Common patterns

Keeping simulations under control

Next Steps

Customer Flows

Personas

Templates

Run Plans

​Customer flows

​Improv vs. Scripted

​Write improv briefs like backstories

​Structure scripted flows around paths

​Start with a happy path

​Add branches for edge cases

​Compose flows instead of duplicating them

​Templating with variables

​Generate flows with Ask Roark

​Personas

​Diversify voice and speech

​Test difficult customer types

AI Skeptic

Hostile Caller

Sensitive Situation

Rapid Switcher

​Plan configuration

​Common patterns

​Keeping simulations under control

​Next Steps

Customer Flows

Personas

Templates

Run Plans

Customer flows

Improv vs. Scripted

Write improv briefs like backstories

Structure scripted flows around paths

Start with a happy path

Add branches for edge cases

Compose flows instead of duplicating them

Templating with variables

Generate flows with Ask Roark

Personas

Diversify voice and speech

Test difficult customer types

Plan configuration

Common patterns

Keeping simulations under control

Next Steps