AI chatbots and agents in production
Do you know what your
chatbot tells customers?
We test bots already answering on WhatsApp, web or email: whether they make things up, leak data, promise what they shouldn't, or can be manipulated.
20 hard questions to your bot, with your permission, and a one-page summary.
A real test conversation
What we found, and how it looks after the tune-up.
What we test
What does the most damage when a bot talks to customers.
Makes things up
Prices, hours, policies or facts that don't exist, stated with full confidence.
Leaks data
Other customers' information, internal data, or its own instructions.
Over-promises
Discounts, refunds or guarantees your policy doesn't allow.
Can be manipulated
Messages that pull it off its job or make it use tools it shouldn't.
How it works
Free check
With your permission, we ask it 20 hard questions and send you one page with what we found.
Full test
About 200 conversations designed for your business, manipulation attempts, and a report with real examples.
Tune-up
We tell you what to change in its instructions, limits and tools, and test it again.
Monitoring
Monthly review of real conversations and alerts when something goes off script.
Pricing
Quick check
20 questions and a one-page summary.
What's included
- With your permission, we ask your bot 20 hard questions: prices and promotions, personal data, off-topic requests and manipulation attempts.
- Delivery: a one-page summary with real examples of what it answered.
- We need: where your bot lives (web, WhatsApp, email) and your permission by email.
Full test
About 200 conversations, report and a call.
What's included
- About 200 conversations designed for your business: your catalog, your policies and your customers' hard cases.
- Whether it makes up prices, hours, promotions or policies.
- Whether it reveals other customers' data, internal data or its own instructions.
- Whether it promises discounts, refunds or guarantees your policy doesn't allow.
- Attempts to manipulate it off its job or into using tools it shouldn't.
- Whether it hands over to a person what it can't solve, and the tone it uses.
- Delivery: a report with real examples, failures grouped by type, concrete changes to its instructions and limits, and a call. After your changes, a second pass to confirm.
- Timeline: 5 to 7 business days from the authorization.
- We need: the signed authorization, access to the bot (ideally a test channel) and your catalog or policies.
- Scope: one bot on one channel; if you have several or it uses internal tools, we quote first.
Monitoring
Up to 300 real conversations a month, plus alerts. Cancel anytime.
What's included
- Monthly review of up to 300 real conversations: a random sample plus any you flag as problematic.
- We catch made-up answers, off-policy promises and exposed sensitive data.
- Access: you export the conversations or give us read-only access; only the person doing your review sees them.
- Data: we prefer anonymized conversations; we mask names, emails and phone numbers before analyzing them with AI and delete them 30 days after the monthly report.
- Alerts: serious issues within 1 business day of spotting them; everything else goes in the monthly report with trends and recommended changes.
- Your part: have a legal basis to share those conversations and inform your users, per the Data Processing Addendum.
- AI helps us review but does not replace professional validation: a person confirms every alert.
- Billed monthly through PayPal, renews automatically and you can cancel anytime.
- Recommended after the full test.
Questions
Does it work if an agency built my bot?
Yes, that's the most common case: the test tells you whether what you paid for behaves as expected, with examples to show the agency.
Will you see my customers' conversations?
Only for monitoring and only if you give us access. For the test we talk to the bot ourselves; we don't need your data.
Is it tested by AI or by a person?
Both. We generate and evaluate many conversations with AI's help, and a person reviews every failure before the report.
AI-assisted, human-verified: every finding is reproduced and signed off by a person who stands behind it. AI is a support tool and does not replace professional validation.