Why we show you our AI's mistakes

On our front desk page there is a section describing two mistakes our assistant made with real customers. Publishing your own product's failures on the page that sells it is unusual, so this piece explains the thinking.

The fear is specific

Business owners are not vaguely worried about AI. They are worried about one specific scene: an assistant, wearing their name, confidently saying something wrong to a customer. Inventing a price. Contradicting a policy. Softening on a position the owner holds firmly. Every glossy AI receptionist demo carefully avoids this scene. That avoidance is exactly why the fear persists.

What actually happened

Our assistant at the academy read an internal price list that included a casual session rate, and twice told a parent that kids could do casual sessions. That is not the policy for kids. Separately, when a parent asked whether their child would train with the opposite sex, it softened and implied an arrangement might be possible, when the business's actual position is that classes are mixed, said plainly and warmly, with no debate.

Both times, a real person got a wrong or wobbly answer. Both times the fix went into the policy layer that governs the assistant. The internal price list now carries an explicit rule: it is not evidence of what we sell, it exists only so you never invent a number, and you must never reason from a price to a policy. The mixed classes position is now stated the way the owner would state it, with instructions not to apologise for it and not to offer workarounds.

Why publishing them helps you

First, it proves the system was trained on reality. Rules like never reason from a price to a policy do not come from a brainstorm. They come from shipping, watching, and fixing. An assistant that has never been wrong in production has usually never been in production.

Second, it shows you the failure handling you are actually buying. Any vendor can tell you their AI is accurate. The useful question is: what happens when it is not? Our answer is visible: the mistake is caught, the policy layer grows a new rule, and that class of mistake stops. The prompt behind our academy assistant is about 17,000 characters, and nearly all of the hard-won part is exactly this kind of rule.

Third, it sets the standard we invite you to hold everyone to. If a vendor will not show you a real failure and the fix that followed, they are either not being straight with you or they have not shipped to real customers yet. Either answer is worth knowing before you sign.

The line we hold

The same honesty applies to what is not built. Our system does not do real-time voice conversations. It does not have an owner inbox yet. It does not answer Instagram DMs. Those gaps are listed on the sales page too, because the first client who discovers a gap you papered over is the last client that referral chain produces.