Most human in the loop designs I see are theater. There's a checkbox, someone clicks approve, but the interfaces already frame the answer.
SPEAKER_00Am I still deciding or am I just approving?
SPEAKER_01If you want to audit whether your team is still deciding or just approving, don't ask them. They'll tell you they're deciding. Look at what they produce. If the doc reads like restatement of what the system suggested, they approved. If it reads like the system said A, I checked B and C, and here's why A still wins, they decide it.
SPEAKER_00How does somebody distinguish, like, when should I really get this serious about it and track my decision making?
SPEAKER_01Yeah, I think the threshold is reversibility. If the thing you're about to ship can be undone cheaply in an hour with an apology and a resent, you don't need the whole reasoning trail. But the moment decision becomes hard to reverse, the calculus flips, right? Hard to reverse means it changes what somebody else does next.
SPEAKER_00In five years, where are these senior people going to come from if AI is doing a lot of the heavy lifting?
SPEAKER_01Yeah.
SPEAKER_00Welcome to Using AI at Work. My name is Chris Daigle, and I'm the host. And before we get started, I just want to take a moment to say thank you to all of our listeners. Uh, we just got notified that we're in the top two and a half percent of podcasts globally, which considering when I started this, it was just gonna be something that I did for fun and to now realize that we've impacted hundreds of thousands of non uh non-technical business professionals who are interested in leveraging AI and their personal and work lives. Uh, that's uh a huge uh win for us. So I just want to say thank you to all the listeners and everybody that uh has listened and forwarded an episode and that sort of thing. And today, this episode is not going to disappoint. So most executives that I talk to, they're attempting to track AI usage, certainly like how many seats have a license, how many people are logging in, and how frequently, how how often they're using the tools. But nobody that I've spoken to is really tracking whether their people are still the ones deciding. Is AI taken over? Are they still leveraging the instincts that they've built in those roles? Now, my guest today has spent two decades on that question from inside the technology. Ronnie Gougerol ran behavioral signals where he built systems that read intent, emotion, deception risk, and those sorts of things out of human voice, which is not an easy task. And it ended up being deployed in financial services, defense, trust, and safety. He's had uh a long history in the space with uh uh numerous successes. But the reason he's here today is because his new book is out. It's called The AI Instinct, and it's out from Wiley with a foreword by known thought leader in the AI space, Kai Fu Lee. His argument is that AI has moved into a layer where our our business instincts used to live. And that now AI is shaping what we notice and what we trust before our our our reasoning catches up. Now, we're gonna get into what that means for how you assign accountability, what a rule looks like when you've got 50 people using AI in your company every day or more. And uh Rana doesn't know this yet, but I'm gonna be asking him about three things that he does personally to make sure that he isn't outsourcing or delegating all of the reasoning and instinct and thinking to AI and that he is still the one making the call. So, Rana, uh congratulations on the book and welcome to the show. Um, anything that you think that the listeners need to know before we get started?
SPEAKER_01Yeah, thank you. Um I mean, I'm excited for this conversation. Um, responsible deployment and practice uh looks a lot less like policy document and a lot more like a design constraint uh you would feel every day when you ship. Um excited to get into it. Let's do it.
SPEAKER_00Nice. So um before we kind of dive into my questions, if uh maybe give me a version of your thesis that you would say to a CEO who has, I don't know, the elevator pitch, the 30 seconds, and doesn't necessarily care about AGI or those sorts of timelines.
SPEAKER_01Yeah, I mean, um it 30 seconds, yes. So I'd say AI has moved from being a tool your company uses to being a system your company thinks with. It's inside the loop of attention and judgment now, right? So your product managers are drafting with it, your analysts are framing decisions with it, your executives are testing arguments against it. That's not a productivity software anymore. That's cognitive infrastructure. And the thing nobody's pricing correctly is what does that do to your organization over time? Because a system that sits inside your judgment either builds better instincts in your people or it produces a very convincing stream of outputs that feels like thinking without actually being thinking. I mean, those are two completely different companies five years out. One has sharpened its taste and the other has quietly outsourced it and doesn't know yet.
SPEAKER_00Interesting. So you you call the title of the book uh is The AI Instinct. Now, instinct is something that gets the it's a biological that gets developed over uh years, especially in business, of uh an individual doing the thing. Why do you think that instinct is the right word for software like AI? And what would have been wrong with calling it maybe judgment or intuition instead?
SPEAKER_01Instinct is the right word because it's the layer underneath judgment and intuition. You know, judgment is deliberate. I mean, intuition is felt. Um instinct is what fires before either of them shows up. It's the precognitive pull that makes you turn your head, flinch, trust someone, or distrust someone. And that's exactly the layer AI is now operating at in our lives. Right? So think about what happens when you open your phone, right? Before you've made a judgment about what to read, before you've had an intuition about what matters today. Um a system has already shaped what's in front of you. I mean, it's acted on you at the instinct layer, it's decided what your attention would snap to. That's not judgment territory. That's older than judgment, right? And there's a second reason. Yeah. Instinct in humans is what you get when experience compounds into something automatic, right? So a trader who's been through three crashes doesn't um reason her way out of the fourth one. I mean, her body knows before her brain catches up. That's decades of um, you know, consequence baked into a reflex. The interesting question with AI is whether these systems are developing something analogous, like a compressed residue of um everything they've been trained on that fires as kind of a machine reflex, um, not really reasoning, right? Not intuition in the felt sense, something more primitive and I and I think a lot more powerful. So if I'd called it the AI judgment, I would have sounded like a book about decision-making frameworks, right? So judgment implies uh deliberating self, like uh weighing options, and that's not what's actually happening inside these models or you know, inside the humans using them, frankly. So intuition would have been closer, but intuition still lives in the domain of feeling, and it's specifically human, at least, you know, so far. And it carries a warmth I didn't want to project into software. So I think instinct is colder, it's mechanical in a way. I mean, it's what happens when a system, uh, biological or otherwise has absorbed enough pardon that response uh, you know, proceeds through, uh, precedes thought, right? And so that's the layer where AI is quietly, I guess, reorganizing human life. And it's the layer where if you're not careful, we stop noticing that our reflexes aren't fully ours anymore. So I think that's why the word matters. It names the place the action is actually happening.
SPEAKER_00Yeah, I love it. So a formula that you reference in the book is human plus tools plus rules. Now, most of the companies that we work with, they've definitely got the first two, but they don't necessarily, I don't even know if they need to have the third. In in your lexicon, your context, what does a rule look like that uh, you know, a company that was had 50 people, let's say, using AI today, what would something, what would your definition of that rule, that third part of this mechanism, look like for just a traditional organization that maybe wasn't necessarily software?
SPEAKER_01So role clarity, I mean, I think that's where it starts. And um, to be honest, I mean, most deployments completely skip it, right? So when a human and AI system work on a decision together, uh you have to be able to answer out loud before anything ships who sets the aim, who supplies the evidence, who has the veto, and who gets blamed when it goes wrong. So if you can't answer those four questions in a sentence each, you don't really have a deployment. You have an accident waiting for a press release. Um I'll give you in a failure mode I see most often, right? It's called a moral crumple zone. I mean, something goes wrong in a human AI system and the blame collapses into the onto the human operator, like the nurse, the loan officer, the driver, the analyst, who actually had very little control over what the model did. The AI is a black box, right? I mean, the org chart says the human is accountable, and so the human absorbs the fault for a system they couldn't meaningfully steer. You know, responsible deployment means you design that out. You don't put a human in a chair, hand them uh model output, and call their signature oversight. That's just theater. The second thing, and I think this is quieter, is uh automation bias, right? When people work next to a confident system long enough, they stop deferring. I mean, they sign off, they stop checking. And then the signal they used to have, the thing that let them catch the model's uh mistakes, uh, are trophies. It just are trophies, right? So Aviation figured this out decades ago and called it automation complacency. You know, uh pilots who fly an autopilot for years struggle when they have to hand fly through a crisis. I mean, the same thing is happening right now in medicine, in credit, in hiring, in content moderation, and we're not naming it. So in practice, audit trails you can actually read, um, supervised autonomy, meaning the model runs, but uh human can override, um, and uh, you know, and the override is cheap and fast. I mean, transparency about what the system considered. Um, and and this is the one people resist deliberate and efficiency, right? You keep uh humans uh doing some work by hand, even when the machine could do it, because you need the skill to stay alive in the org. So I think that's the part that makes it responsible deployment and keeps your people sharp instead of quietly hollowing them out.
SPEAKER_00So that's interesting, that last point that you brought up, because uh most people, you know, I think the the the average non-technical business professional who was using the tool isn't thinking about that way. Like I still need to keep my they they understand human in the loop and those sorts of things, but they don't necessarily understand that for their own benefit, they need to continue to get those repetitions in. Like you said, the AI could do it, but I, as in your position, I as the human should take some of that back from AI and continue to do it manually. What's the what's the takeaway or the win for the the the power user, let's say, who's really got you know maybe multiple screens going and that sort of thing, and they have started to defer more to the model and they've lost that that that pilot going from autopilot back into manual controls.
SPEAKER_01Yeah, the the win is uh that you stay dangerous, and I mean that in the best sense, right? So here's what happens to the power user who's running six stabs, three models, and has basically outsourced their first draft uh of thinking to the system. I mean, they start to feel incredibly productive. I mean, output goes up, uh turnaround shrinks, and the dashboards look great. Um, but something quieter is happening underneath, right? The small friction that used to make them curious, the moment where you uh go, hey, wait, that doesn't sit right. The moment where you rewrite your own sentence because you don't quite believe it yet, those moments get smoothened over, right? And so the copilot fills the gap before the gap can do its work on you. And the gap is where judgment gets built. Like judgment's uh is an accumulated in relationship with consequences. Um you get it by being wrong, noticing you were wrong, and updating. If the system's always catching you before you fall, you never develop the muscle that uh that tells you the floor is about to give. So the practical takeaway for the power user is this, right? So pick the decisions in your week that actually matter. The ones uh where being caught uh wrong costs you something real and deliberately do those manually first, right? Draft your own take before you ask the model, write your own hypothesis before you pull the data. I mean, argue with yourself a bit before you argue with the co-pilot. And then bring the AI in as a challenger, uh, not as a starter. I mean, ask it what evidence would would contradict what you just wrote. Ask it for alternative you reject it. So I think the other habit I push is uncertainty hygiene. When the model gives you an answer, uh force yourself to say out loud how confident you actually are, independent of how confident it actually sounds, right? So confidence in these systems is stylistic. I mean, it's not calibrated. If you don't check your own gauge, you'll just simply inherit theirs, right? So I think the pilots who survived the automation crisis are the ones who still hand flew the plane on clear days. Same principle, right? I mean, to do the manual reps when the stakes are low. So the instinct's there when the autopilot drops out.
SPEAKER_00I think that's fantastic advice for the listeners. Like, don't fall into the trap of more is better necessarily, right? Now, you said that organizations should stop treating AI as a separate capability and start treating it as part of the team's cognitive architecture. So what does an org chart look like when a leader starts to uh apply that? Like, where does the where do the roles change? How do the role definitions change? What roles need to get added to an org chart in this environment?
SPEAKER_01The org chart stops being a map of humans and starts being a map of decisions. I think that's the shift, right? So today most org charts answer the question who reports to whom? And when AI is generally part of your cognitive architecture, the chart has to also answer who owns the decision. I mean, what tools sit inside it and what rules governs how they interact. The unit isn't the person anymore. I mean, the unit is people plus tools plus rules, right? I mean, working on a specific class of decision, pricing decisions, hiring decisions, diagnostic decisions, underwriting decisions. I mean, each one is a little cognitive system and needs to be drawn as one. So practically a few roles change and a few roles uh show up that weren't there before. I mean, your your senior operators, your VPs, your directors, I mean, their job shifts from being the smartest person in the room to being the designer of the room. They're deciding which decisions get AI in the loop, at what depth, you know, what with what veto rights. Um, that's a design job, not a management job. And most of them have never been trained for it. So that's the problem, right? And then you need roles that didn't exist five years ago, right? Someone who owns model behavior and production, not just model performance, someone who owns what the system is allowed to remember about your people and your customers and what it's required to forget. Someone who runs what I call the renewal cycle, because any human AI arrangement should have an inspection date and a right to withdraw built into it. I mean, that's a real function. Call it whatever you want. You know, cognitive operations, uh, AI stewardship. The name matters less uh than the fact that a human being wakes up on Monday um responsible for it. And I think you retire something too. I mean, you retire the fiction that the tool is just a tool and therefore doesn't need a seat on the chart. I mean, if a system inside your loop of attention and uh judgment uh is actually actually inside your loop of attention judgment, and it's shaping what your team believes before they finish forming the belief, it belongs on the map. I mean, leaving it off the chart doesn't make it less powerful, it just makes it unaccountable.
SPEAKER_00So you mentioned something and in in the interviews that I've been doing, I've started to see a pattern. Uh recently I spoke to somebody and they said that at the individual level, the person is a doer, but in kind of the environment that we're entering into where AI is becoming a generative AI is becoming much more a part of everybody's role, that they need to go from a doer to a planner. And that's similar to this concept that you just uh described where uh rather than me being the smartest person in the room as an executive or a leader of a team, I need to now be the designer. How does somebody develop that skill? Where do you learn that? Because you're right, it's not something that's been part of the role.
SPEAKER_01Yeah, I mean, I think you'll learn it by running small, cheap experiments on decisions you already own and treating the design of the decision as the artifact, not the outcome. Um so here's what I mean, right? So pick one recurring decision of uh on your team, not a giant one, something you you would do weekly, a hiring screen, a pricing exception, a content approval, something like that, right? So write down on one page how that decision gets made today. Who sees what, in what order, with what evidence, and where the judgment call actually happens. I mean, most people have never done this even um you know, for even one of their decisions. I mean, when you write it down, you immediately see the scenes. I mean, you see where a model would draft, where a human has to weigh, where the veto sits, uh, where the audit trail is missing. And that one page is the muscle. Uh, do it for 10 decisions, and and you're sort of like a designer, right? So the second thing is that this is the part people skip is you have to get fluent enough with the tools uh to have taste. Not to build them, to have taste. You've never sat with Claude or GPT or Gemini for a weekend and try to get it to do something real inside your domain. Um, you know, you cannot design uh a system around it. You you'll either overtrust it or undertrust it, and both are expensive. So I tell executives uh spend 24, uh 24 hours, you know, uh not 24 hours of demos, um, 24 hours of you alone trying to make the thing useful on a problem you actually care about. You'll come out the other side with the physical sense of where it's sharp and where it hallucinates, and what um, and that sense is what lets you draw the room. So the third the third piece is you study people who already do this, right? So product managers uh at good software companies have been designing human plus system uh workflows for 20 years. Air traffic controllers, ICU teams, trading desks. Go read how they think about handoffs, alarms, override authority. I mean, that literature exists. I mean, most business leaders have never touched it because it wasn't their job. Now it is. You know, so nobody's going to hand you a certification for this. You build it by doing it on something small in public with your own team watching.
SPEAKER_00Yeah, that's fantastic advice. And for the listener, one of the ways that we've kind of seen it at Chief AI Officer is that before we would certify a chief AI officer, there was something that I was looking for. And it wasn't that they were using AI like a drive-thru window. Hey, I need an answer, got it, I'm out. It was that individual who spent, as Rana's suggesting, an entire weekend in one thread, really going deep and paying attention to stuff and seeing how like, and if you have never done that as a listener, I would encourage you to rather than just bounce between threads, go deep and see the difference, and you will understand the model behavior for your role and your use in a way that you never have before. So, one of the concepts, Rana, that that's talked about in the book is this idea of attention sculpting. Can you tell me what you mean by that?
SPEAKER_01Yeah, attention sculpting is what happens uh when a system decides on your behalf and usually invisibly what you notice, right? And because what you notice becomes what you value, the thing shaping your attention is quietly shaping your preferences. So think about the mechanics, right? So every feed you touch, every recommender, every autocomplete, um, every rank search result is making a claim about salience, right? It's saying out of the ocean of things you could be looking at right now, um, these are the ones worth your milliseconds. I mean, you experience uh you experience that as a convenience. What's actually happening is curation with a point of view. The system has a model of view and it's pulling on the levers of what surfaces and in what order with what emotional temperature. The the reason I call it sculpting and not uh say filtering is that filtering sounds subtractive, like like a like the system just removes noise. Sculpting is additive and it's shaping, right? A sculptor doesn't just take marble away, they reveal a specific form. I mean, these systems are revealing a specific version of the world to you over and over, and the version they uh reveal trains what your attention learns to reach for the next time. You know, TikTok is the crude, obvious example, right? But the same mechanic runs inside uh a Gmail priority inbox, inside a code assistant suggestion ranking, inside which three documents your enterprises search puts on top, right? So now pair that with what you uh you were just describing, right? The person who spends the weekend in one thread. That person is doing the opposite of being sculpted. They're pushing back on the system's sense of uh salience and imposing their own. I mean, they're saying, I decide what deserves that PA. Right? Not the interface. That's the muscle that atrophies when you live in the drive-thru mode. I mean, you stop noticing that something else is doing the noticing for you. So where this gets uh, I guess, ethically loaded is that attention sculpting doesn't uh ask for consent in any meaningful way. There's no moment where the system says, Hey, I'm about to ship what feels important to you for the next four hours. Is that okay? Right? It just happens. And in a hybrid setup, where the assistant finishes your sentences and uh mirrors your tone, the the sculpting stops feeling external. It just feels like you had the thought. That's the part leader should be losing sleep over.
SPEAKER_00Now, in that in that vein, I guess, uh, there's this self-check of am I still deciding or am I just approving? So how would you suggest a leader kind of audit? I mean, maybe as an individual, you can stop and you can ask yourself those questions. Maybe you've got a sticky note that says, you know, am I deciding or am I just approving? But how do you audit that in the team's behavior and their use of the tools?
SPEAKER_01The tell isn't in the decision, it's the it's in the artifacts around the decision, right? So if you want to audit uh whether your team is still deciding or just approving, don't ask them. I mean, uh they'll tell you they're deciding. Everyone thinks they're deciding. Look at what they produce, right? The three things that actually inspect, right? First, the reasoning trail. Um, when someone on your team green lights and AI assisted recommendation, is there a written record of why they agreed and specifically what they considered and rejected? Uh, not a summary of the model's output, their reasoning. I mean, if if if the doc reads like a restatement of what the system suggested, they approved. If it reads like the system said A, I checked B and C, and here's why A still wins, they decide it. You can spot the difference in about 30 seconds per artifact. Right? I mean, the second disagreement uh rate, right? And so track how often your team overrides the tools. I mean, if if a team is running a copilot for three months and the override rate is under 5%, something is wrong. I mean, either the tool is superhuman, which it isn't, or the humans have gone quiet. I mean, healthy hybrid work has friction in it. You want to see people pushing back, asking for alternatives, requesting the counterfactual. If everyone's nodding, uh nobody's thinking. Sure. And I think the third, and I think it's the one most leaders won't do is run a periodic unplugged decision. Like pick a recurring call once a month, one person on the team makes it without the AI in the loop. Then compare. Not to prove the human is better, but to see whether the human still can. I mean, the moment someone says, I don't know how I'd even start without a tool, you've found your uh you found your de-skilling. I mean, that's just signal. Yeah, yeah. So the meta movement underneath all three is that you're you're building an audit trail of human cognition, not of AI output. Most companies log what the models did, but almost nobody logs what the human thought flip that, right? So the question isn't what the system would recommend. Uh the question is what did you weigh and what would have changed your mind? And if your team can't answer that on demand, um, they're just approving.
SPEAKER_00So, you know, uh I guess there's a strata of usage. Some might be, hey, I need the email rewritten, some might be I need to do strategic action for the company and I need I need the support. That approach is sound, however, it seems like at some point on that strata of I guess uh difficulty or complexity of the use of the model, it would be cumbersome at the low end of the strata and extremely valuable at the high end of the strata. How does somebody distinguish, like, when should I really get this serious about it and try my decision making and be prepared to defend the reason why I presented this artifact as compared to something else?
SPEAKER_01Yeah, I think the threshold is reversibility. That's the single question that tells you where you are on the strata, right? If the thing you're about to ship can be undone cheaply in an hour with an apology and a resent, you don't need the whole reasoning trail. Um, you know, rewrite the email, move on, right? Life is short and cognitive overhead is real. So if if you audit every keystroke, you'll grind to a halt and your team will start hiding their AIUs from you, which is worse than any of the harms you were trying to even prevent, right? So so the moment uh but the moment uh decision becomes hard to reverse, the calculus flips, right? Hard to reverse means it changes what somebody else does next. A pricing move, competitors will match, a hire, a public commitment, um, you know, a model going into production against real customers, um, the a strategy dock that becomes the frame everyone else plans against for the next two quarters. I think those are the decisions where you owe yourself uh and the people downstream of you are a record of why this and not that. So I I kind of, you know, I use a rough um three-band test in my own work. Like band one, um reversible in under a day, no artifact needed, just ship. Band two, reversible but expensive, meaning it'll cost you a week or a relationship to unwind. You write two lines, what the system suggested, what you actually chose, and why they differ. Two lines. That's it. You know, uh band three, functionally irreversible or high stakes for someone else's life or money, you write the contrastive version. I mean, I chose A or B because of X and Y, even though B offered Z, and you keep it, right? Um the other trigger uh separate from reversibility is um asymmetry of harm. Um if the downside lands on someone or somebody who wasn't in the room when you decided, um, that's a band three, no matter how small the decision feels to you. Like, you know, a collection script, a screening filter, a routing rule, uh small to you, but enormous to the person on the other end. So the honest answer to when it gets serious is like, you know, uh when future you or a regulator or a person uh affected would reasonably ask you to show your work, um, you know, that's when. So if you show it now, it's cheaper than reconstructing it later.
SPEAKER_00Heck yeah. Okay, that's very, very helpful. Now I'm gonna shift gears a little bit because I I want to kind of dig into how you actually use it. So um maybe take a second and walk me through what you do before you open the model on a strategic decision. Like what would you write down? How how long would that process take?
SPEAKER_01Yeah, I mean, um before I open the model, I write three things down on paper actual paper or a blank doc with nothing else on the screen, right? This takes me about 10 minutes, and it's the most important 10 minutes of the whole session, right? First thing I'd write, what am I actually trying to decide in one sentence with the decision maker named? Not think about pricing. Something like I need to decide by Friday whether we raise enterprise pricing 15% for new logos only, and I'm the one signing off. If I can write that sentence, I'm not ready to use the tool. I'm ready to journal. Those are two different activities. Right? And the second thing, what's my current instinct and why? Right? Two or three sentences. Like this matters because once the model gives me a fluent, well-structured answer, my prior gets quietly overridden. I won't remember what I actually thought coming in. So I pin it. Later I can compare. And um, if I've drifted, I want to know whether I drifted because of a real argument or because the output just sounded confident. Right? Third, what would change my mind? This is the one people skip, and it's the one that does the most work. I write down two or three specific pieces of evidence and arguments that, if true, would flip my instinct, right? Churn data over a threshold, a competitor move, whatever that is, like a customer segment I hadn't considered. Now I have a test. Um, now I have a test the model has to pass instead of a why but it has to match, right? Um so then you open the model and I don't ask it what to do. Um, you know, I ask it to argue against my instinct using a criteria I just wrote, and I asked it to um find the evidence I said would change my mind and tell me honestly if it exists. I mean, I ask it to name what I'm probably not seeing. The whole pre-work is maybe 10 minutes, the session after is 30, 40. And the reason the 10 matter so much is that it makes me a client with a brief, not a user with a prompt. You know, the system's now working on my judgment instead of substituting it, and that's the whole game.
SPEAKER_00I love it. So that that really feeds into what you were talking about earlier, like forcing the repetitions of the human still participating in the process. And listeners, if you're if you were to follow that process, I think it becomes obvious that like you are very involved. You're not just absorbing output, you're now actively participating as a partner in the thinking rather than outsourcing that thinking to, hey, what should I do? Open-ended kind of question. What would be some, you know, we we talked about this. One of the concerns that we we hear across a bunch of industries is that particularly like legal and things where the years of being in that role allows you to be the person who's now capable of training that junior, training the person that's gonna re that you can delegate to or that may eventually replace you as you you move on or or ascend in the organization. In 10 years, where are these or five years, where are these senior people gonna come from if if AI is doing a lot of the heavy lifting, even though we're participating, as you just suggested, you know, the the pre-work before we go into the models? How do people with that concern, how do they build something today that makes sure that they're not gumming down the human as AI gets more capable and becomes more of a part of their operational modus operandi, I guess?
SPEAKER_01Yeah. Um, you know, replacement is a story we tell uh to make budget decisions easier, right? Augmentation is what actually happens on like a Tuesday afternoon, right? So I give you the frame I actually use, right? So every job is a bundle of tasks, and every task sits uh somewhere on a spectrum from pure pattern recognition to pure judgment under ambiguity, right? So AI is coming for the pattern recognition and first, you know, hard and it's already there in a lot of places. Uh look at what happened in customer support. Like Klarna said in early 2024, their AI assistant was doing the work of 700 agents. That's real. But if you actually look at what stayed, it's the escalations, the weird ones, the calls where somebody's angry and the policy doesn't quite fit. You know, those got harder, not easier, because the easy ones stop landing in human cues. So the line I draw isn't augmentation versus replacement. It's which tasks in this role are compressible and what's left when you compress them. And then critically, is the leftover job actually a job a human wants to do? Or have you built a role that that's nothing but exception handling and blame absorption? You know, and I think the last part is where I think most executives are getting it wrong right now. They're modeling headcount reduction and calling it augmentation because it sounds gentler. But if you strip the routine 80% out of the out of a role and leave a person doing nothing but the hard 20, you created a job with no on-ramp, no recovery time, and no way to build intuition. I mean, junior people can't grow into it uh because the practice reps are gone. And senior people burn out because there's no rhythm, right? There's that's not augmentation. That's that's a role that will quietly fail in 18 months, and you will blame the person, right? So the honest version of augmentation is when the human's decision quality goes up and the human is still doing enough uh of the work, it stays sharp uh to stay sharp at it. You know, uh radiologists reading with the eye assist and catching more, uh, not radiologists rubber stamping a queue of 400 scans and a shift. I mean, those are different features, and we're choosing between them right now, mostly without admitting we're actually choosing.
SPEAKER_00So as we as you come to the end of this, this has been awesome. I'm looking forward to getting the book. Uh, and we're gonna have, for the listener, we're gonna have links to uh order the book uh right there on the show notes, and I would encourage you to do it because I think that if nothing else, maybe you've you've watched a lot of tactical trainings on how to use the tools better, how to build a skill, whatever those things might be. Maybe you've had uh external experts come in and work with the team to help them start to integrate generative AI into the operations and the workflows and the processes and that sort of thing. But if you're listening to this, you are a decision maker, right? And I think that if you I mean, I I I'll admit, some of what you've talked about as not necessarily best practice for the human uh AI loop, I've been guilty of for sure. Um so as a listener, I would encourage you to practice those three steps, the pre-work that Rana shared earlier, and see if the experience and particularly the quality of the output that you're getting improves. I have no doubt that it will. Um, and again, I'll admit I've been guilty of not necessarily doing that uh exclusively on the strategic work. So that was a big takeaway for me. Now let's talk about tomorrow for the listener or Monday morning, you know, if they're listening to this over the weekend. Let's say it's a CEO, they they they listen to this, they're running a $50 million company, maybe they've got Chat GPT business deployed for their teams, but there's no real rules like we talked about earlier with human plus tool plus rules. Maybe they don't have that audit trail where an individual who's providing output that was synthesized from the models is able to defend it. And maybe they don't even have any idea of whether or not they haven't they haven't had a uh an open discussion about this idea of deciding versus approving. What would you suggest be one of the first steps that they do to start to introduce this into um how their company behaves when using the models?
SPEAKER_01Yeah, so the cleanest test I know, can you point to the exact moment in your workflow where a human has enough context, enough time, and enough authority to say no to the model? And does that moment actually happen or is it just theater? Right? Because most human in the loop designs I see are theater. I mean, there's a checkbox, someone clicks approve, but the interface has already framed the answer, the queue has already prioritized speed, um, and the reviewer has like 30 seconds uh and 40 items. That's not oversight. I mean, that's just laundering, right? So the responsible deployment starts by taking that seriously and designing against it. Uh, and I a few things I look for when I'm evaluating the system ours or someone else's um are like one, uh calibrated confidence, right? So the model needs to tell you how sure it is, and that number needs to actually mean something across the distribution. If it says 90% and it's write 60, you've built a machine that manufactures false trust. You know, calibration isn't a nice to have, right? It's it's a it's the substrate uh under every downstream decision a human makes uh uh about whether to defer or push back. Two, um, audit trails that uh that a non-engineer can read, not logs, trails. Um the what, the why, the inputs, the alternatives that were considered and rejected. Because when something goes wrong, and it will, I mean, you need to be able to reconstruct the decision without a forensic team. And if you can't, you don't really have accountability. You have an accountability gap. Um that's yeah, that AI, uh the AI is a black box, um, and and the operators uh you know says they were just following it, and nobody's actually responsible. Um and then it becomes a failure mode. Um three, watch for what I call the moral crumple zone, which we discussed earlier. When the systems fail, blame tends to land on the human operator who had the least real control. The clinician who accepted the flag, the analyst who signed off. So if your deployment design lets that happen, you haven't deployed responsibly. You've built a liability sheet with the person shape dent in it. So the last one, the fourth, plan for de-skilling. You just have to plan for it. I mean, if the human override exists on paper, but the humans never actually override in 18 months, they've lost a skill too. You know, uh we talked about aviation, which figured out decades ago and it called it automation complacency. You have you have to force the muscle to stay alive. I mean, structured disagreement, um, rotation cases where the humans decide without seeing the model first and then compares. Um, those things become really important. Otherwise, the loop is closed and you just haven't noticed. So I think, you know, responsible isn't like a certification. It's whether any of that um uh any of that is actually happening on a Wednesday afternoon when nobody's watching.
SPEAKER_00Wow. Um a whole different way of thinking about AI, dear listener, for sure. And going from, oh, now I can move faster to wait a minute, maybe I need to slow down for a second on some of these activities to make sure that the output that I'm getting isn't just slop, as they call it, right? So, Rana, final question. What is a question that you wish more executives or or potential readers would ask you? And what would be your answer to that?
SPEAKER_01Every founder I talk to uh can tell me what AI is doing uh for their output. You know, faster tickets, uh cheaper content, better code review, great. What almost none of them can tell me is what six months of using these tools have done to how their people think. You know, whether their designers uh still sketch before prompting, whether the PMs still write the spec themselves or just edit the one the model gave them, whether the their engineers can still read a stack trace cold, or whether they've quietly become dependent uh on pasting it into um the chat window. So I think that's where the actual risk lives. I mean, it's not in the tool, it's in the drift. Um so and here's why I'd give this advice over uh above everything, anything else. As a founder, you have maybe um 18 months to build a team whose judgment you can trust when things get weird, and and and things always get weird. You know, the fundraiser blows up and the customer turns and the the model, the provider uh changes pricing overnight. Um in those moments, you need people who can reason from first principles, not people who can prompt well. And if you spend your first year optimizing for velocity through AI, you may find you you've built a company that can execute beautifully on well-defined tasks and falls apart the moment task isn't well defined. Right? So I think um practical version, I guess, you know, once a quarter, pick a real problem in your business and have your team solve it without any AI assistance, not as a hazing ritual. Yeah. You know, as a diagnostic, right? Watch what happens. And if a room goes quiet, if people can't get started, um, if the quality falls off a cliff, you've learned something important about what you've actually been building, and you still have time to correct it, hopefully, right? The founders who win the next decade won't be the ones who um adopted AI fastest. They'll be the ones who noticed early what it was doing to the minds of the people around them and made deliberate choices about it.
SPEAKER_00Wow. That's a very interesting exercise, and I think I might present that to some of our clients that we've worked with for you know longer than six months. It's fantastic. So, Ron, again, congratulations on the book. Uh, the book is The AI Instinct. It's out by Wiley. It's out now. You can get it um anywhere that you're you're buying books, including Amazon. Um, and a number of takeaways from me personally, um, I'm going to be introducing several of these concepts into the trainings that we're doing for organizations because we have been more focused on the speed, the efficiency, the quality productions that are possible. And it's been kind of an open loop for me. Well, what does that do to the human? And I think that this model that's been uh discussed today and is discussed in detail in the book, obviously, uh is gonna be a new addition to perspectives that we introduce to the leadership as they're uh seriously considering uh AIFI and their team. So, Rana, again, thank you. I wish you all the success with the book. And uh listeners, we're gonna have all of the information related to the book and how to get it, and more information about how to get in touch with Rana in the show notes. And uh once again, thank you for uh being a listener of the show. I really appreciate it. If this episode, if you got some as many takeaways as I did from this, please feel free to forward it to anybody else who's leading a team, who's asking these questions about how do we make sure that we're using AI at work correctly. So, Rana, again, thank you so much.
SPEAKER_01Thank you, Chris. Uh, it was a real pleasure. Thank you for having me. You're welcome.
SPEAKER_00Thanks for tuning in to using AI at work. Don't forget to subscribe for more conversations about how to use AI at work. And a special thank you to our sponsor, Chief AI Officer, for empowering businesses with AI education and training. Visit their website for a free AI readiness assessment and AI strategy guide to help you get started using AI at work. That's www.chiefaiofficer.com. Follow us on Twitter at the handle UsingAI at work and visit www.usingai at work.com for free resources to help you harness AI in your role.