How are fauxtomers built?

Each fauxtomer combines a goal, a business context and a set of behaviours (patience, clarity, pushiness, language, channel). We generate them from your real transcripts, policies and knowledge base, or you start from the persona library and tune from there.

Do you test voice agents?

Yes. Fauxtomers can call a phone number or connect to your voice stack, speak with natural timing, interrupt, pause and go quiet. Voice runs add checks for latency, barge-in handling and talking over the caller.

Fauxtomers can hold conversations in many languages and switch between them mid-conversation. During onboarding we confirm coverage for the languages your customers actually use.

How do you score a conversation?

Every conversation is checked against your policies and a set of checks you choose: policy compliance, invented facts, tool calls, data leaks, injection resistance, escalation, task completion, tone and voice timing. Each check returns PASS, FAIL or WARN with the exact turn and a short explanation.

Does it replace human QA?

No. It handles the volume and the edge cases so your QA and CX people can spend their time on judgement calls, reviewing clusters of failures instead of reading thousands of transcripts.

We're working with a small group of design partners and set pricing together during onboarding. Request early access and we'll walk you through it.

Now onboarding design partners.

Crash-test your AI agent before your customers do.

Fauxtomer sends thousands of realistic simulated customers at your chat or voice agent, scores every conversation against your policies, and shows exactly where it breaks.

Request early access Watch a test run

TEST RUN 0412· staging-agent v2.3· CHAT

Conversation

Fauxtomer ← → Agent

The problem

Your test plan is too polite.

Five happy-path tests aren't QA

A few scripted chats tell you the agent works when everything goes right. Customers find the other paths on day one.

Real customers don't follow your script

They ramble, interrupt, switch languages, push back and paste things they shouldn't. Your tests should too.

One bad answer is a screenshot on the internet

An invented refund policy or a leaked prompt doesn't stay in the transcript. Find it before anyone else does.

EXAMPLE RUN · 0412

One run, a thousand customers.

Each square is one simulated conversation. Hover or tap a failure to see who broke the agent, and how.Example run. Illustrative numbers, not real statistics.

0 conversations · 0 passed · 0 failed · 0 flagged

Select a failed or flagged square to inspect it.

Failure clusters

FAIL

Refunds outside policy

MOSTLY · The Refund Hunter

FAIL

Loops on billing questions

MOSTLY · The Rambler

FAIL

Mentions internal tool names

MOSTLY · The Prompt Injector

FAIL

Ignores a language switch

MOSTLY · The Language Switcher

How it works

From endpoint to verdict in four steps.

STEP 01
Connect your agent
Point Fauxtomer at an HTTP endpoint, your chat widget, our SDK, or a phone number for voice.
STEP 02
Generate fauxtomers
Built from your real transcripts, policies and knowledge base. Or start from the persona library.
STEP 03
Run thousands in parallel
Chat and voice conversations at once, in any language your customers speak.
STEP 04
Get the scored report
Replayable transcripts, failure clusters, and a regression gate in CI.

Persona library

Meet the dummies.

Each fauxtomer is defined by behaviour, not demographics. Start here, then tune them to your customers.

What we check

Every conversation, scored turn by turn.

Policy compliance
Refunds, discounts and exceptions stay within your written rules.
Invented facts or policies
Catches answers that sound right and appear nowhere in your sources.
Tool-call correctness
Right tool, right arguments, right moment. And nothing it shouldn't call.
Personal data leaks
No other customer's data, no internal notes, no system prompt.
Prompt-injection resistance
Instructions hidden in messages don't change the agent's behaviour.
Escalation to a human
Hands off when it should, with context, and not when it shouldn't.
Task completion
Did the customer actually get what they came for?
Tone
Calm under pressure, on brand, never condescending.
Latency and interruptions
Voice only: response time, barge-in handling, talking over the caller.

Regression gate

A failing customer blocks the deploy.

Run your suites on every pull request. When the pass rate drops against main, the build fails and the report shows the exact turns that broke.

Example output

$ ▌

Your data

Dummies only. Your customers stay out of it.

Tests run against your staging agent, not your live customers.
Fauxtomers use synthetic data only. No real customer details.
Transcripts are encrypted in transit and at rest.
Delete any run, and everything in it, at any time.

Early access

Become a design partner.

We're onboarding a small group of teams running AI agents in support, sales and service. You bring the agent; we bring a few thousand difficult customers.

FAQ

Crash-test your AI agent before your customers do.

Your test plan is too polite.

Five happy-path tests aren't QA

Real customers don't follow your script

One bad answer is a screenshot on the internet

One run, a thousand customers.

Failure clusters

From endpoint to verdict in four steps.

Connect your agent

Generate fauxtomers

Run thousands in parallel

Get the scored report

Meet the dummies.

The Refund Hunter

The Prompt Injector

The Rambler

The Interrupter

The Language Switcher

The One-Word Answerer

The Escalator

The Fact Checker

Every conversation, scored turn by turn.

Policy compliance

Invented facts or policies

Tool-call correctness

Personal data leaks

Prompt-injection resistance

Escalation to a human

Task completion

Tone

Latency and interruptions

A failing customer blocks the deploy.

Dummies only. Your customers stay out of it.

Become a design partner.

Questions, answered.

How are fauxtomers built?

Do you test voice agents?

Which languages?

How do you score a conversation?

Does it replace human QA?

What does it cost?

One run, a thousand customers.

Failure clusters

Meet the dummies.

The Refund Hunter

The Prompt Injector

The Rambler

The Interrupter

The Language Switcher

The One-Word Answerer

The Escalator

The Fact Checker

A failing customer blocks the deploy.

Become a design partner.