What Happens on the Bad Day? AI Oversight for Leaders

Rana Gujral is the CEO of Behavioral Signals, which uses voice AI to detect intent, emotion, and risk in real time. He has been on both sides of enough AI deals to know which questions buyers ask, and which ones they should.

“A lot of responsible AI talk lives at the altitude of principles. Fairness, transparency, accountability. They’re beautiful words. They don’t survive contact with a Tuesday morning product decision. So when I think about what it actually looks like in practice, I try to push it down to the level of design choices and habits.”

“The question I almost never hear buyers ask, and it’s the one I’d lead with, is what does the system do when it’s wrong? Not how accurate is it, not what’s the benchmark score. What happens on the bad day?”

That depends on what the system assumes going in. Every model is trained on data that reflects someone’s world, not necessarily yours. When it hits your users, your edge cases, your context, it has no way of knowing it’s out of its depth.

Ask vendors to articulate those assumptions and you get silence. That’s data. And if the answer on rollback is a pause followed by “well, in practice that hasn’t really come up,” walk away. If you can’t undo an automated action, you don’t have oversight. You have hope.

Treat AI as a feature and it will behave like one

When AI rollouts fail, leadership teams tend to reach for the same explanations. Data quality. Talent shortages. Regulatory uncertainty. Integration costs. Rana thinks those are real. They’re comfortable to talk about, but they’re not what’s stopping change.

The businesses that fail start with the model, then go hunting for a problem worthy of it. That approach “almost always ends in a pilot that demos well and dies quietly.” Everyone has access to roughly the same models, the same APIs, the same vendors. The difference is somewhere else.

“It’s that organisations treat the AI as a feature instead of as a participant in the workflow. A team buys or builds a capability, plugs it into an existing process, and assumes the surrounding humans, incentives, and feedback loops will absorb it. They don’t.”

But the workflows organisations drop AI into are made up of hundreds of small human judgments that nobody has mapped. Human cadence, escalation paths, patterns. Drop in a system that fails differently, at a different rate, in different places, and the seams show. People don’t know when to trust it, when to override it, when to escalate. They either over-defer or route around it. One produces bad outcomes; the other produces no outcomes.

You can’t insert AI meaningfully into a loop you don’t understand, and the loop is already moving under you.

AI is moving upstream of your decisions

“AI isn’t taking over our thinking through some dramatic rupture. It’s slowly moving upstream of it. From answering our questions to shaping what we notice to influencing what we want before we know what we wanted.”

Intelligence without experience is hollow. A system can produce a convincing answer with no memory of being wrong, no update from consequence, no skin in the game. That’s the argument at the centre of The AI Instinct, Rana’s forthcoming book, which introduces the concept of artificial general experience, or AGE. Using AI either sharpens your judgment or delivers a stream of outputs that “feel like thinking without actually building any thought process.”

Most people picture AI’s influence as the moment you ask a chatbot something and it answers. That’s the top inch of what’s happening. Perception, pattern recognition, and emotional weighting happen largely below conscious awareness, and that’s the layer AI is increasingly touching.

It’s personalising persuasion, mirroring tone and adjusting to hesitation until influence stops feeling external.

Consent drift follows. Every frictionless tap turns consent into performance rather than deliberation: you’re not choosing anymore, you’re confirming a prediction the system already made about you. Agency is a practice muscle, and if every path is optimised for comfort, the capacity to push back disappears.

If the AI layer went away tomorrow, could your organisation still think?

The question Rana puts to every leadership team: can analysts still analyse, writers still write? If not, something has been outsourced that wasn’t meant to be.

“The skill that’s about to be scarce isn’t prompting, it’s verification.” People who can look at a confident output and know what to stress-test, what to override, what to trust. That comes from reps, not training videos, so give teams real decisions with stakes and feedback now, before the systems get more persuasive.

Treat every AI output as a probabilistic aid, not an oracle. Teams should ask what evidence would contradict a given output, and nobody, including the CEO, gets to use “the model said so” as a final answer. That costs nothing and is the strongest protection against the drift that sinks deployments.

A model starts as a suggestion engine and six months later, if nobody has named the limits, nobody is overriding it. aibl sees this consistently in mid-market organisations.

Rana’s starting point: pick one decision and map it. Who decides, what inputs do they use, what does good look like, what does failure cost. “You can’t govern what you haven’t named.” Where stakes are high and reversibility low, AI is advisory only; where stakes are low and reversibility high, let the system run. Either way, write down the decisions that stay human regardless of how good the model gets: hiring, firing, capital allocation, anything tied to organisational values.

A quarterly review keeps it honest: where did AI make us faster, where did it make us worse without anyone noticing, and what did we learn that we couldn’t have learned without it.