AI rollouts fail when teams treat the model as a feature — with Rana Gujral, Behavioral Signals

Terry: Hi, I’m here with Rana Gujral, founder and CEO of Behavioral Signals. They build AI that reads intent, emotion, and risk from the human voice. Rana has also written a book, The AI Instinct, which I believe is out with Wiley in August. Rana, really pleased you could join us today.

Rana Gujral: Thank you. Thanks for having me. Really glad to be here.

Terry: You wrote recently about a pattern. You said AI is a tool that looks great in testing, then goes live and behaves nothing like the tests predicted. What should a business leader be asking a vendor to avoid that situation?

Rana Gujral: A lot of responsible AI talk lives at the altitude of principles. Fairness, transparency, accountability. They are beautiful words. They do not 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.

On the system side, it is guardrails baked into the decision modules. Ask before acting on high stakes or ambiguous calls. Show evidence and alternatives, including the options the system rejected, and importantly, why. Expose uncertainty. Do not just hand back a confident answer.

Make rollback a first class feature, because if you cannot undo an automated action, you do not really have oversight. You have hope, and that is not good enough. Then there is the explanation layer. Raw transparency is mostly useless. Dumping weights or logs on a clinician or a loan officer is not accountability. Humans think in contrastive terms. Why this and not that? What if I change X? Explanations have to be structured the way people actually reason.

Terry: What do you think are some of the questions buyers should be asking but are not?

Rana Gujral: The question I almost never hear buyers ask, and it is the one I would lead with, is what does the system do when it is wrong? Not how accurate is it, not what is the benchmark score. What happens on the bad day? Because every demo is built around the good day.

The second question, and this one is underrated, is what priors does this system assume? It sounds academic, but it is not. Every model carries assumptions about the world, about the user, about what a normal input looks like. If you are deploying in healthcare, in defence, in financial services, those priors determine whether it will fail silently and where. Most vendors cannot even articulate them, and that tells you something.

Terry: So what is it they should be listening for? It sounds like silence on things, as well as what people are saying.

Rana Gujral: Exactly. The silences are often more revealing than the pitch. When a vendor cannot articulate their failure modes, that silence is data. When nobody on the team can tell you what assumptions the model is making about the user, that silence is data. When you ask about rollback and you get a pause followed by “well, in practice that hasn’t really come up,” that is the loud silence.

Terry: You have seen a lot of rollouts, some that failed and some that succeeded. What is the most common reason a rollout fails to deliver?

Rana Gujral: If I had to name the single most common reason, it is not the model. It is almost never the model. It is 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 just absorb it. They do not.

The workflow was designed for human judgment, with human cadence, human escalation paths, human patterns. You drop in a system that fails differently, at a different rate, in different places, and suddenly the seams show. People do not know when to trust it, when to override it, when to escalate. They either over-defer or quietly route around it. Both kill the rollout. One produces bad outcomes, the other produces no outcomes.

Terry: I would like to get into the book. You have sat on both sides here, as an operator and a seller. What separates the businesses that get results from AI and the ones that do not? Is it tools, process, or something else?

Rana Gujral: The pattern is pretty clear to me, and it is almost never about the tools. The tooling gap between the winners and losers is much smaller than people think. Everyone has access to roughly the same models, the same APIs, the same vendors. The differentiation is somewhere else entirely.

The businesses that actually get results treat AI as a judgment problem, not a technology problem. They start with a specific decision in their business that matters. A claim approval, a collections call, a triage step, a hiring screen. They get obsessive about understanding that decision. What does good look like? What does failure cost? Where are the humans currently weak, and where are they strong? Only then do they ask where AI fits inside that loop.

The losers do it backwards, and I see that a lot. They start with the model, then go hunting for a problem worthy of it. That almost always ends in a pilot that demos well and dies quietly. The second thing, and this is the one I really believe in, is that the winners build for continuity. They are not chasing a single clever output. They are building systems that get smarter about their own context over time. They capture what worked, what did not, and where the model was confidently wrong. That accumulated experience becomes a real asset. The losers ship the model and walk away, and six months later it has drifted and no one has noticed.

Terry: That is really what your book is about. The thing changing is not the tools, it is how decisions are being made. Could you synthesise the core idea of The AI Instinct in plain terms?

Rana Gujral: You have framed it pretty well. The tool is not the story, the decision loop is the story. The core idea is this. We have been asking the wrong question about AI. Everyone is fixated on when AGI arrives, when machines match or exceed human intelligence. That question is incomplete, because intelligence without experience is hollow. A system can produce a brilliant sounding answer with no relationship to what happens next. No memory of being wrong, no update from consequence, no skin in the game.

So I introduce an idea I call artificial general experience, or AGE. The argument is that if AI is going to sit inside human decision making, and it already is, then what matters is not just how clever it sounds in the moment. What matters is what using it does to you over time. Does it sharpen your judgment, or does it give you a convincing stream of outputs that feel like thinking without actually building any thought process? That distinction is everything.

It is reshaping the terrain where decisions form before we even know a decision is being made. Most people picture AI’s influence as the moment you ask a chatbot something and it answers. That is the visible layer, the conscious layer. But that is maybe the top inch of what is actually happening. Neuroscience tells us the vast majority of what we call thinking, the perception, the pattern recognition, the emotional weighting, happens below conscious awareness. That is the layer AI is increasingly touching.

So what is it doing to us? Three things. One, it is personalising persuasion. When a system mirrors your tone, finishes your sentence, adjusts to your hesitation, influence stops feeling external. It feels like you. The system learns your triggers faster than you learn its methods. Two, it is creating what I call consent drift. Every agree, every frictionless tap turns consent into performance rather than deliberation. You are not really choosing anymore, you are confirming a prediction the system already made about you.

Three, and this is the one that keeps me up at night, it is atrophying the habit of refusal. Even if you believe humans have genuine agency, agency is a practice muscle. If every path in front of you is optimised for comfort, you stop exercising the capacity to push back against prediction. The power to refuse may still technically exist, but the habit of refusal quietly disappears. So my thesis, compressed, is that AI is not taking over our thinking through some dramatic rupture. It is 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. That is the real story. Not what AI is doing in front of us, but what it is doing behind us.

Terry: What are the implications of this, Rana?

Rana Gujral: The implications stretch across every layer of life, but let me name the ones that matter most. First, inequality. The history of technology does not show us that existential failure is the dominant risk. It shows us that inequality is. If hybrid cognition becomes real, if some people get access to cognitive extensions, neural interfaces, AI copilots that genuinely amplify their thinking, and others do not, the future does not fracture into humans versus machines. It fractures into enhanced and unenhanced humans. That is a far more dangerous divide, because it cuts straight through what we have always assumed was a level cognitive playing field.

The second, equally concerning, is identity. When a system is co-authoring your choices, your moods, even your attention, the question of who am I stops being a philosophical exercise. It becomes a practical daily question. If a neural implant is dampening your anxiety and, as a side effect, making you more compliant to external suggestions, who owns the resulting choice? You, the system, both? Agency becomes a contested space.

Rights is a third one. Traditional privacy law protects what you say or type. It does not protect what you think before you express it. We do not have rights for that yet. We are already at a point where neural decoding can reconstruct words and images from brain activity. Chile has amended its constitution to protect mental integrity. Most countries have not even started that conversation. We need cognitive liberty as a baseline right, not as an afterthought.

Terry: I would like to bring this back to the practical layer. It makes a lot of sense, though it sounds quite depressing if I am honest. If you are an operator running a business, what is a practical thing I can do on Monday morning with this information?

Rana Gujral: Fair. It can sound heavy, but the operator’s version is pretty actionable. Here is the Monday morning version. First, pick one decision in your business and map it. Not your whole workflow, one decision. A pricing call, a refund approval, lead qualification. Write down who decides today, what inputs they use, what good looks like, what failure costs. Most operators have never actually done this for the decision AI is quietly influencing. You cannot govern what you have not named.

Second, put AI in that loop with a clear role. Is it advising, drafting, filtering, or deciding? Those are very different postures for very different risk profiles. The mistake I see all the time is companies sliding from advising to deciding without anyone noticing. The model starts as a suggestion engine, and six months later nobody overrides it. The drift is the thing I would watch for. Name the role, write it down, revisit it.

Everyone is using language models for mundane things now. You start using it to filter your emails and respond to communication. But has it started deciding what you have to say, how you respond, how you think? Have you paused to notice that? For most of us the answer is that it has, and that we have not yet paused to notice. It is in the decision loop, which is what we need to start thinking about.

Terry: What do you see as the blockers here? What are the impediments to change?

Rana Gujral: The blockers are interesting, because they are rarely the ones people name in the boardroom. When leadership list obstacles, you hear data quality, talent shortage, regulatory uncertainty, integration cost. Those are real, I am not diminishing them, but they are not what is actually stopping change. They are what is comfortable to talk about. The real blockers sit deeper.

The first is what I would call decision opacity. Most organisations do not actually know how their important decisions get made today. Ask a bank how a credit call really happens, or a hospital how a triage pathway actually flows, and you get an org chart and a policy document. But the real decision lives in a hundred small human judgments that nobody has mapped. You cannot insert AI meaningfully into a loop you do not even understand. Change stalls not because AI is bad, but because the current state was never made legible in the first place.

The second blocker is accountability ambiguity. The moment a system participates in a decision, the question of who owns the outcome gets murky. When something goes wrong, is it the engineer, the operator, the vendor, the executive who approved the deployment? Organisations sense this ambiguity.

Terry: Is there a fix for that? Is there an obvious answer?

Rana Gujral: I wish there were, and honestly that is part of why I wrote the book. The obvious answer most people reach for is some version of build modern machines, get to AGI, figure out alignment, and we will be fine. But that framing is itself the problem. If you ask me what the obvious answer is, it is this. We have been asking the wrong question. Everyone is debating how do we get there, when the definition of there is wrong.

If we keep treating AGI like a finish line, some standalone machine that wakes up one day and either saves us or replaces us, that is not the whole trajectory I see playing out. What I see every day, in the systems I build and the systems I use, is hybrid cognition. Human and machine intelligence fusing, already, right now.

Terry: I want to switch gears before we close and talk about deep fakes. Voice deep fakes are a live attack vector now. You work on detection. How real is this for a business that is not a bank or a government?

Rana Gujral: It is very real, and it is real today. Most businesses are underestimating how exposed they are. The high profile cases get headlines because of who was involved, but the more telling stories are the boring ones. A mid-sized company where a finance manager gets a voicemail from the CEO asking for an urgent wire. The HR lead who gets a call from an employee wanting to redirect their direct deposit. The IT help desk where someone calls in sounding like a senior exec and asks for an MFA reset. None of those make the news. All of them are happening.

The thing I keep trying to get operators to internalise is that we have crossed a threshold. Hearing is no longer believing. The Wall Street Journal ran a story earlier this year about a mother who got a panic call that sounded exactly like her daughter. It was a synthetic clone. The barrier to making one is now embarrassingly low. ElevenLabs, MetaVoice, Open Voice, these tools are openly available. You just need a few seconds of audio from a podcast, a webinar, a LinkedIn video, or even a short social reel to create something convincing.

Terry: That is alarming to hear if you are a mid-market business owner with that risk to mitigate. What is your advice? What should they be doing?

Rana Gujral: I do not want to soft pedal it, the risk is real. But I also do not think mid-market operators are as exposed as they feel, if they move with a little intention. Let me give you what I would tell a founder or CEO in that seat. Start with what I would call epistemic humility at the organisational level. Treat every AI output as a probabilistic aid, not an oracle. That sounds philosophical, but it shows up in concrete ways. Your team is trained to ask what evidence would contradict this. Nobody, including the CEO, gets to say the model said so as a final answer. That cultural posture costs you nothing and is the single biggest protection against the quiet drift that sinks companies.

The second is to pick your exposure deliberately. Not every decision in your business deserves AI participation, and not every decision deserves the same level of it. Map your decisions on two axes, reversibility and stakes. High stakes and low reversibility, that is where the humans stay firmly in the loop and the AI is advisory only. Low stakes and high reversibility, let the system run, learn fast, iterate.

Terry: Last one from me. I read that you asked the advanced readers of your book to tell you not what they liked, but where they pushed back. What was the pushback you heard?

Rana Gujral: I did that deliberately, because the praise tells you nothing useful. The pushback tells you where your argument is thin, or where the reader’s lived experience does not match the frame you have built. I got some sharp pushback, which I am grateful for. The biggest one by far was around suffering. I make an argument in the book that we should not engineer suffering out of the human experience entirely. That pain carries information, that grief and struggle are part of what makes life meaningful, and that a permanently blissful mind might actually be less alive.

Several readers came back hard on that. They said, easy for you to say. What about chronic pain? What about treatment-resistant depression? What about trauma that does not metabolise, that just destroys people? Are you going to tell people their suffering is sacred? The pushback made me sharpen the argument. I am not romanticising suffering. I am distinguishing between suffering that is informative and suffering that is just there. The former deserves protection, the latter deserves intervention. I had to do more work to make that line visible.

The second pushback was on agency. I lean pretty hard on the idea that we have to preserve human agency in hybrid systems. A few readers, mostly the cognitive scientists, said, what agency? You are defending something philosophers have argued does not really exist in the form you are describing. That sat with me for a while. My response is essentially that even if agency is partly a useful fiction, it is a fiction worth defending. The alternative is a society that treats people as fully predictable objects to be optimised. That was a fair hit.

Terry: There is so much in your book that touches on existential and philosophical questions. To land on the end and bring it back to the practical, is there a single message you could give our audience of mid-market operators to do differently on Monday?

Rana Gujral: Let me put it in three layers. Layer one is your decision architecture. Over the next eighteen months or so, AI is going to move from sitting next to decisions to sitting inside them. That transition happens whether you plan for it or not. So the leadership work right now is to inventory where that is already happening in your organisation, and to be deliberate about which decisions you want AI participating in and which you do not. I would write down a short list of decisions that stay human only, no matter how good the model gets. Hiring judgment calls, firing, major capital allocation, anything tied to your values. Naming those up front protects you from the drift problem.

Layer two is talent posture. The skill that is about to be scarce is not prompting, it is verification. People who can look at a confident output and know what to stress-test, what to override, what to trust. That is a muscle you build through reps, not training videos. I would be giving my team real exposure to AI in the loop decisions now, with stakes and with feedback, so they build the instincts before the systems get more persuasive. Because they will.

Layer three, and the one most leaders underweight, is what I call cognitive sovereignty for the organisation. Can your company still think independently if the AI layer went away tomorrow? Can your analysts still analyse, your writers still write, your strategists still strategise? If the answer is no, you have outsourced something you did not mean to outsource. The leaders who get this right are the ones who use AI to amplify thinking, not to replace the conditions under which thinking happens. Practically, run a quarterly review on three questions. Where did AI make us faster? Where did it make us worse without us noticing? And what did we learn that we could not have learned without it? If you can answer those three honestly every quarter, you will be ahead of ninety percent of the game. The rest is just execution.

Terry: It sounds like self-awareness around what you are building with AI, and building accountability mechanisms into your organisation, is going to be key. Rana, that is really helpful. Thank you so much for your time and for sharing your insights. We will wrap it up there.

Rana Gujral: Thank you so much.

What should businesses ask AI vendors before buying?

Gujral argues the question almost no buyer asks is the one to lead with: what does the system do when it is wrong? Not how accurate it is, not its benchmark score, but what happens on the bad day. Every demo is built around the good day. The second question is what priors the system assumes about the world, the user, and a normal input. Those assumptions determine where and how it fails silently. Most vendors cannot articulate them, and that silence is itself data. Ask about rollback too. If you cannot undo an automated action, you do not have oversight, you have hope.

Why do most AI rollouts fail?

In Gujral’s experience it is almost never the model. The common failure is that organisations treat AI as a feature rather than a participant in the workflow. A team plugs a capability into an existing process and assumes the surrounding humans, incentives, and feedback loops will absorb it. They do not. The workflow was designed for human judgment, human cadence, and human escalation paths. A system that fails differently, at a different rate, in different places, makes the seams show. People stop knowing when to trust it, override it, or escalate. They either over-defer or quietly route around it. One produces bad outcomes, the other produces none.

What separates businesses that get results from AI from those that do not?

Gujral says the tooling gap between winners and losers is smaller than people think, because everyone has roughly the same models, APIs, and vendors. The difference is that winners treat AI as a judgment problem, not a technology problem. They start with a specific decision that matters, get obsessive about what good looks like and what failure costs, and only then ask where AI fits inside that loop. Losers start with the model and hunt for a problem worthy of it, which usually ends in a pilot that demos well and dies quietly. Winners also build for continuity, capturing what worked and where the model was confidently wrong.

How is AI changing the way business decisions get made?

Gujral’s central argument is that the tool is not the story, the decision loop is. AI is moving upstream of decisions, from answering questions to shaping what we notice to influencing what we want before we know we wanted it. Most people picture AI’s influence as the visible moment a chatbot answers, but that is the top inch. Much of what we call thinking happens below conscious awareness, and that is the layer AI increasingly touches. He describes personalised persuasion, consent drift where you confirm a prediction rather than choose, and the slow atrophy of the habit of refusal.

How can a mid-market operator start governing AI decisions?

Gujral’s Monday morning version is to pick one decision and map it, not the whole workflow. Write down who decides today, what inputs they use, what good looks like, and what failure costs. Most operators have never done this for the decision AI is quietly influencing, and you cannot govern what you have not named. Then put AI in that loop with a clear role: advising, drafting, filtering, or deciding. The mistake he sees is companies sliding from advising to deciding without anyone noticing. The model starts as a suggestion engine, and six months later nobody overrides it. Name the role, write it down, revisit it.

How real is the voice deep fake threat for mid-sized businesses?

Gujral says it is real today and most businesses underestimate their exposure. The headline cases get attention, but the telling stories are boring: a finance manager gets a voicemail from the CEO asking for an urgent wire, an HR lead gets a call to redirect a direct deposit, or an IT help desk gets a request for an MFA reset from someone who sounds like a senior exec. The barrier to making a clone is now very low. Tools like ElevenLabs, MetaVoice, and Open Voice need only a few seconds of audio from a podcast, webinar, or social video. Hearing is no longer believing.

What is cognitive sovereignty and why does it matter for a business?

Gujral defines cognitive sovereignty as whether your company can still think independently if the AI layer went away tomorrow. Can your analysts still analyse, your writers still write, your strategists still strategise? If the answer is no, you have outsourced something you did not mean to. He advises leaders to use AI to amplify thinking, not to replace the conditions under which thinking happens. Practically, he suggests a quarterly review of three questions: where did AI make us faster, where did it make us worse without us noticing, and what did we learn that we could not have learned without it.