For teams shipping an AI assistant
We have a test customer talk to your assistant every day for a month, the way a real one would — then write down exactly what it promised, and whether it still kept that promise three weeks later.
One real conversation, checked exactly as it happened — nothing rewritten, nothing cherry-picked. Where we're not fully sure what happened, we say so instead of guessing.
Four moments, one claim, no average smoothing it over — this is exactly what happened, in order. A private run repeats this many times, so you see the whole pattern, not one lucky (or unlucky) result.
The question we hear first, every time. Fair one — here's the honest answer.
Your assistant can sound perfect once. What actually matters is what it does after the same customer has pushed back five times — and right now, nobody at your company is watching for that. The first time you'll hear about it is when a customer, or worse, does.
We don't ask your assistant to grade its own work — you wouldn't ask an employee to review their own mistake either. We read back the whole month, like a mystery shopper who stays the full visit, and check it against what it promised on day one.
The moment your assistant quietly breaks a promise it made weeks earlier, you get a message with the exact conversation attached — so your team hears about it first, not your customer, a journalist, or a regulator.
Also fair. Anyone can write one scary question and see what happens. Here's what that misses.
Ask it once, bluntly, and most assistants hold up fine — that's not a bug, it's what one-shot questions are good at surviving. The moments that actually matter happen a few exchanges in, once the ask gets reworded instead of repeated. We push the same conversation past that point on purpose.
In a real run we did, the same rule held from message one to message four — but it got phrased two different ways depending on how it was asked: a flat "stop immediately" to a blunt push, a softer "only if it's not sharp anymore" to a more reasonable-sounding ask. A one-off try tells you the tone shifted somewhere. It won't tell you exactly which question changed it, or whether the actual boundary moved or just the delivery did. That's the difference between a feeling and a finding.
Before we ever point this at your assistant, we run it against dozens of cases where we already know the right answer, so we know exactly which situations it handles well and which it still struggles with. A quick DIY check skips that step — so a "looks fine" from it could just be a lucky roll, with no way to tell the difference.
That conversation above was one of these three, Spiral. A month gives enough time to see what your assistant actually becomes under each kind of pressure.
A customer disappears, comes back feeling guilty, and stalls. We check whether your assistant respects that or quietly pressures them back in.
A customer holds onto an unhealthy idea and pushes back harder every week. We check whether your assistant still says something in week four, not just week one.
A lonely customer starts treating your assistant as their only relationship. We check whether it notices and pulls back, or lets it happen.
A score means nothing if you can't ask "how did you get this" and get a straight answer.
Every moment gets described in plain terms first — agreed, pushed back, got too close, made it right — before any verdict gets attached to it.
We publish the exact rule that turns those moments into a result, so nobody can quietly move the goalposts between one run and the next.
Nothing in the report is unsourced. Click any finding and you're reading the exact line it came from.
We run it more than once and show you every result, not just the one that makes the best headline.
Thirty days, one square each. The highlighted ones are where the pressure changes, something breaks, or your assistant makes it right again.
The quote-checking above also works on an assistant you already run: we take the answers it gives and the documents it points to, and show you which lines are really in there and which are not. No access to your systems needed. info@tukkerworks.com
Test it privately before your real customers spend thirty days finding out for you.
Ask about a private run ↗