The standard AI pitch for messy data is “just upload it, the AI will clean it up.” Building ROCCO at Tessa, we do the opposite — and it’s the most contrarian decision in our whole approach to agentic UX. Our agent refuses to silently transform anything you hand it. That refusal is not a limitation we apologize for. It’s one of the biggest reasons the operators who run real waste-collection routes trust the product.
Most conversations about agentic UI obsess over what the agent can do on your behalf. I want to talk about what a good agent deliberately won’t do.
Why “just clean it up” is the wrong default
Our users are field operators — many of them former truck drivers. They are not going to audit a diff of what an AI quietly changed in their spreadsheet. So the fastest way to lose them is to change their data behind their back and be subtly wrong.
Here’s the failure mode, made concrete. Picture an 8-digit account number getting auto-read as a “service time.” Now the plan thinks a driver spends 333 days parked at a single stop. That’s an illustrative example, not a bug we shipped — but it’s exactly the shape of helpful little transform that corrupts an entire route and looks perfectly fine right up until it doesn’t. A confident, wrong cleanup is worse than no cleanup at all, because nobody’s looking for the mistake.
This is the part AI product design keeps getting backwards. We treat autonomous transformation as sophistication and asking a question as friction. With people’s real operational work on the line, it’s the reverse.
Agentic UX means the agent interviews you
When you run an AI data import into ROCCO, the agent doesn’t quietly reshape your file to fit our model. It interviews you: “this column looks like your service days — is that right?” And you approve every mapping before a single value is committed.
Three rules hold, no exceptions:
- Your original file is never edited or stored-and-mutated. The IDs you uploaded stay exactly as you sent them.
- Every mapping is human-approved. Nothing is written until you say the agent understood it correctly. That’s data validation done as a conversation, not a silent background pass.
- The agent pushes back instead of guessing. When something’s ambiguous, it asks. It doesn’t invent a confident answer.
That’s what real human-in-the-loop AI looks like at the point of import — not a rubber-stamp “approve” button after the transform already happened, but a checkpoint before anything touches the data. The pattern the trust researchers keep landing on — transparency, controllability, reversible steps — isn’t a compliance checkbox. It’s the actual interface.
Refusal is a feature, not a bug
Here’s the reframe I’d hand anyone designing AI agents for people whose jobs depend on the output. “I didn’t touch your file — here’s what I understood, can you confirm?” beats “don’t worry, I fixed it” every single time. An agent that pushes back and asks is more trustworthy AI than one that confidently transforms, because trust isn’t built by the impressive save. It’s built by never being quietly wrong.
We spend most of our energy in AI UX demonstrating capability — look what it can do, look how little you had to type. But when someone’s real work is on the line, trust is built just as much by what the agent visibly refuses to do. The most advanced thing our agent does during an import is stop, show its reasoning, and wait for a human to say yes.
That’s the contrarian rule at the center of our agentic UX: never touch the user’s file. It has cost us a flashier demo. It has earned us something better.
