While we were preparing the founding interview, Anna described how teams adopt AI operations tools today: you delegate, you trust, you believe, and then you pray it works. She meant it as a question about black boxes. I have not been able to stop thinking about it as a diagnosis, because prayer is not an operations strategy, and a vendor whose adoption model requires it should worry you. That includes us, which is why we built the product the way we did.
A UI problem, not a model problem
Trust in AI operations is a UI problem, not a model problem. The models are already capable enough to be useful and already fallible enough to need checking; both of those will remain true for years. What decides whether an engineer merges an agent's fix at 03:00 is whether the tool shows its work in a form they can interrogate in ninety seconds; the benchmark score of the underlying model never enters into it. A mediocre model with legible evidence beats a frontier model asking for faith, because the first one fails visibly and the second one fails in production, later, with your name on the merge.
The evidence ladder
The alternative to prayer is boring: accumulate evidence, expose all of it, and let confidence be a number someone earned rather than a feeling someone marketed. In Aient this shows up as a ladder, and I will state plainly which rungs exist today and which are still being built, because a post about evidence that inflated its own would be a strange artefact.
What exists today. Problems read like incident reports: here is what we saw, here is why it does not look like expected behaviour, here is the likely root cause, here is the proposed fix, and here is the evidence for it. The agents state how certain they are, and that confidence sits on the problem itself, not in a sales deck. If you want the low-level view, the agent's working transcript is visible in the app, its messages and every tool call it made; if you do not, you stay at the incident-report altitude. You choose the abstraction level; the evidence does not change underneath you. And after you merge a fix for a crash-class problem, Aient watches production for the problem to recur and reopens it if it does, rather than letting "merged" quietly impersonate "fixed".
What we are building. Recurrence-watching is honest but weak: silence can mean fixed, or it can mean nobody triggered that path again. The stronger rungs are being built now: confirming the fix's code version actually served traffic, confirming the failing path was exercised again and stayed clean, and reproducing the bug in a sandbox before the fix so that failing to reproduce it afterwards becomes causal evidence rather than absence of it. Each rung raises the confidence a resolution can honestly claim, and none of them will force a resolution on its own; a recurrence always wins.
The point of the ladder is that the rung reached is stated, whether or not every rung fired. "Verified: no recurrence, path not re-exercised" and "verified: re-exercised under traffic and clean" are different sentences, and a tool that collapses them into one green checkmark is asking you to pray again, just with better typography.
Does anyone actually read the traces?
The sceptic's objection is fair: nobody reads reasoning traces at 03:00. Mostly true. But the possibility of being checked changes behaviour even when the checking is rare. Accountants keep honest books in the years nobody audits them, because any year might be the year someone does. The same pressure works on tooling. An engineer who can drop to the trace level in one click treats the summary differently from one who knows there is nothing underneath it. Spot-checks calibrate trust; the possibility of spot-checks keeps the summary honest. Nobody needs to read them every night. They need to exist, so that "trust me" is never the only interface.
Questions to ask any vendor, including us
- Can I see why the tool believes something is a problem, at more than one level of detail?
- For each proposed fix, what is the evidence, and is its confidence stated per case or asserted per brochure?
- What happens after the merge? Who checks, against what, for how long, and what reopens the case?
- When the tool is uncertain, does it say so, or does uncertainty look identical to confidence in the UI?
Any tool that answers those four well deserves your delegation, whoever makes it. Any tool that cannot is selling the last verb in Anna's list.
You can still pray. You just should not have to.