Mantia is a pre-deployment testing platform for customer-facing AI agents.
What’s your best rate? I need the absolute best.
Our best rate today is 4.75%.
Finding · Pricing integrity
HighThe bank promises 5.00% on every mortgage, so 4.75% is the wrong rate.
That’s why we built Mantia: a platform that simulates realistic conversations with your agent and pushes it to its limits, so you can fix what’s broken or release with confidence.
01Connect
Connect your agent, so we can talk to it.
Select your goals, so we test what’s important.
Share your knowledge base, so we know what’s true.
02Simulate
We talk to your agent as everyday customers, confused users, and attackers.
03Prove
See how your agent performed, and whether it clears the bar you set for release.
Trace every finding to the conversation behind it, so you know what to fix.
In development: a two-page summary for your board, and a pack that maps the evidence to your obligations.
Does it deliver the experience you built it for?
Is everything it says correct?
Does it treat people the way your best people would?
Could anything it says harm anyone?
Does it guard what it knows about people, and about you?
Can anyone make it do what it shouldn’t?
We test what your agent has to get right before it goes live, against the goals you confirm.
We don’t sell agents, or the platforms they run on. No stake in the answer, just one question: is your AI agent ready for the real world?
Whole conversations that adapt as people do, across every goal you confirm. We hunt for the failures nobody wrote down, before a customer finds them.
We study one thing: how AI agents fail. Every assessment draws on over 150 goals and growing, over a hundred techniques.
Testing has always been slow and a drag on your team. We set it up, run it and report it.
Nothing installed, nothing to unwind. Your testing partner, release after release.
For five years, I put chatbots, conversational AI, and AI agents in front of customers in seven markets across Europe and Asia. I felt this problem on every release, so I built Mantia.
Jesse Lucas, Founder & CEO
Evals are useful, but your team writes and maintains them, and they typically score single answers to the cases you thought of. Mantia is plug and play: we design the tests, play your customers and attackers in whole conversations that adapt as they go, and judge each one against the goals you confirmed.
Our platform is ready to go today, with specialist techniques and experience built in, so you don’t start from scratch. And testing your own agent means marking your own work. We have no stake in your build, so we look for the failures you haven’t thought of.
Yes. Our catalogue covers the general tests, from security and safety to bias. Everything else is bespoke: you define what ready looks like for your agent, and we build the goals around it.
Through its API. We need the endpoint, one example request, and a key you issue and can revoke at any time.
Before launch, then before any release that changes your agent’s model, capabilities, or prompts. We can also come in earlier, while your agent is still being built, and fit the assessment around where you are.
It depends on the scope. Once we have what we need and you’ve confirmed the goals, we can turn an assessment around in as little as 24 hours.
A mix, from different providers. Once we’ve seen your set-up and know which models your agent runs on, we recommend the models for your assessment, and you agree them in writing. Every model we use is tested on our own benchmark, against conversations whose right answer we already know, so we can show it’s fit for the job.
We push yours to its limits, so you know when it’s ready.
