Past the Prompt: Why AI Skill Demand Is Shifting From Using to Building, and How Leaders Can Close the Skills-Support Gap

[00:00:00] Remember when being quote-unquote good at AI just meant you knew how to write a really clever prompt? Oh, yeah, definitely. You know the era I'm talking about, like tweaking a few words to get that perfect, slightly less robotic email draft, or generating a cool image for a slide deck. Right, and it felt like magic back then. It really did. Well, if that's all you've got on your resume today, you are already falling behind. It is genuinely startling to see how fast, I mean, the ground is shifting beneath our feet here. Yeah.

[00:00:30] We went from celebrating basic access to these tools to, you know, expecting complete mastery, essentially in the blink of an eye. And that is exactly what we are focusing on today. I was recently reading some brand new research from Dr. Jonathan Westover, published just this month, October 2026. Brand new. Yeah, hot off the presses. He put out this piece called Past the Prompt, and he is basically arguing that the skills we thought were like totally future-proof a year ago are already becoming obsolete.

[00:01:00] Already, yeah. So our mission for this deep dive is to figure out what the new gold standard skill actually is and, you know, how you can avoid getting swallowed up by the widening gap between what workers need to know and what companies are actually prepared to teach them. Right. Okay. Let's unpack this. Let's dive in. Because to truly understand why the era of basic prompting is effectively over, we have to look at the numbers. Yeah, the data.

[00:01:27] And the speed of the shift here is what really stands out, honestly. The numbers are wild. Westover highlights a report from Workday, also from this month, by the way, that looked at job requisitions across roughly 550 major enterprise employers. That's a massive sample size. Huge. And according to their data, the demand for simple prompt engineering peaked back in January 2026. Then over the next few months, it just plummeted by 25%. Wow.

[00:01:57] But over that same exact period, the demand for hands-on architectural skills, things like building AI tools, automating complex workflows, AI engineering, that surged by 51%. Right. The entire market just flipped upside down inside of a single year. What's fascinating here is the underlying driver behind that massive flip. I mean, it is a textbook case of rapid commoditization. Okay. How so?

[00:02:22] Well, think about your daily workflow right now. AI is built into almost every mainstream software application on your desktop. Right. It's everywhere. Yeah. Your email client auto-drafts, your word processor summarizes, your messaging apps suggest replies. We have data from Microsoft and LinkedIn showing that 75% of knowledge workers were already using AI a while ago. Oh, right. They were mostly just bringing their own tools to work. Right. They call it BYO AI. Bring your own AI. Precisely.

[00:02:48] And when a skill becomes that universally accessible, when literally anyone can do it with zero formal instruction, it, well, it immediately stops differentiating you as a candidate. That makes sense. Employers stop putting prompting in job descriptions for the exact same reason they don't ask if you know how to use a web browser. It is just expected now. Right. It goes from being this rare superpower to being just absolute table stakes. Exactly.

[00:03:13] So, to map out where we are actually heading, Westover introduces this concept of a four-rung capability ladder. The first rung is where we used to be, which is literacy. Right. The basics. Yeah. That is just knowing how to give basic instructions, writing a simple prompt. And we should note the OECD has warned that globally we might not even have enough training infrastructure to get everyone to that basic literacy level. Which is concerning.

[00:03:39] Very. But the leading edge of the market has already abandoned that rung entirely. Right. Which brings us to rung two, which is fluency. And this isn't just using the tool. It's integrating AI into your everyday tasks. Yeah. But more importantly, as the paper defines it, fluency is having the judgment to know when to trust the AI and when to turn it off. That's a crucial distinction. Definitely. And then rung three is building. This is where that 51% surge in demand is happening.

[00:04:08] It means creating custom assistance, automating multi-step workflows, and connecting AI to your team's specific proprietary data. Right. And then there is the fourth rung, which is really where the future is heading, and that's stewardship. Stewardship. Okay. Yeah. This is about governing the risk of these systems, completely redesigning how your team operates, and actively training others. Wow.

[00:04:31] Microsoft's latest Work Trend Index expects that within the next five years, over 40% of leaders will be looking for quote-unquote agent bosses. Agent bosses. Yeah. People who can manage and train autonomous AI agents just like they manage human staff. I want to make sure we're really grounding this for everyone listening. If I use an analogy here, literacy is just knowing how to drive a car. You know where the gas pedal is and how to steer. Right.

[00:04:56] But building is like being the mechanic who actually tunes the engine or customizes the car for a very specific racetrack. I like that. But wait, let me stop you there because building sounds exactly like software engineering. Are we saying everyone needs a computer science degree now? Do we all need to learn Python? That is probably the most common misconception right now, and it is crucial to clarify. No, building does not mean traditional coding. Okay. Phew.

[00:05:21] Right. The enterprise platforms available to workers today, they're designed to let non-engineers configure assistants or chain together highly complex steps without writing a single line of traditional code. So what does that actually look like in practice then? It requires an entirely different mental model. It is closer to process design or systems architecture than it is to conversation. Okay.

[00:05:44] You are not just asking a chatbot a question. You are thinking about inputs, handoffs, conditional logic, and automations. Right. For example, you might build a workflow where an AI reads incoming vendor emails, extracts the invoice data, cross-references it with your internal spreadsheet, flags discrepancies, and then automatically drafts a reply for you to approve. Oh, wow. Yeah. You didn't code that, but you architected the system.

[00:06:09] Okay. So it's less about knowing a programming language and more about being an architect of the workflow itself. You have to understand how the work actually gets done step by step. Exactly. And this ties directly back to that second rung on the ladder, which is fluency. Right. The judgment piece. Yes. Specifically, the judgment to know when these tools fail. Well, Westover highlights a really brilliantly designed study by Delacqua and colleagues. Okay.

[00:06:35] They took over 750 consultants from Boston Consulting Group, very smart, highly capable people, and gave them real-world tasks to do with the help of AI. And I'm guessing the AI helped them completely crush those tasks. Well, yes and no. This is where the mechanism of how we interact with AI becomes so important. The researchers mapped out what they called the jagged technological frontier. The jagged frontier. Okay.

[00:07:00] Yeah. Inside this invisible boundary are tasks where the AI is naturally strong, things like synthesizing information or generating ideas. On those tasks, the consultants working with AI were 25% faster and completed 12% more tasks than those without it. Those are massive productivity gains. They really are. But here is the catch. The researchers deliberately included tasks that fell just outside that jagged frontier. Oh, interesting.

[00:07:26] Yeah. These were tasks requiring highly contextual, nuanced human reasoning, things AI is currently bad at but will confidently pretend to be good at. Right. The hallucinations. Exactly. And on those tasks, the consultants using AI were actually less likely to produce correct solutions than the consultants who were given no AI assistance at all. Wait, really? The AI made these top-tier consultants worse. How does that even happen?

[00:07:52] It's a phenomenon called automation bias. The AI generates an answer so quickly and with such overwhelming confidence in polished language that the human expert essentially falls asleep at the wheel. Oh, I see. They stop applying their own hard-earned expertise. They just blindly trust the machine, accept the output, and get it completely wrong. Wow. So true fluency, true capability is having the discernment to spot that invisible, jagged boundary. Literacy without human judgment can destroy value just as fast as it creates it.

[00:08:22] Which is a really sobering thought. I mean, imagine a whole team of people blindly prompting away, generating very confident, highly polished, and entirely incorrect work. It's a disaster waiting to happen. But this brings us to a really uncomfortable disconnect in the modern workplace. If building and fluency are the new non-negotiable rungs on the ladder, how is the average worker supposed to learn this stuff? That is the multi-million dollar question right there.

[00:08:49] And the data points to a glaring systemic gap between what leaders expect and the reality on the ground. Let me share this tension from the Workday report because the numbers really paint a picture here. Here, 79% of workers say they know exactly what skills they need to succeed. They aren't confused about the direction the wind is blowing. Right. But only 66% say their employer actually helps them develop those skills. That's a huge drop-off. It is.

[00:09:15] Meanwhile, 40% of business leaders openly state they expect AI to help them squeeze more productivity out of their current workforce rather than just cutting headcount. So leaders demand more output. Workers know they need more advanced skills. But the bridge connecting them, employer-supported training, is incredibly flimsy. And we're not just talking about abstract professional development here either. The financial stakes for the listener's own career are incredibly real. The wage premium is huge.

[00:09:44] Lightcast analyzed over 1.3 billion job postings. And they found that 51% of jobs requesting AI skills are now completely outside of traditional IT and computer science roles. Wow. Yeah. They are in marketing, HR, operations, finance. And those jobs carry an average salary premium of $18,000. That's about 28% higher than comparable roles without those skills. That's a massive jump.

[00:10:09] And PwC did their own analysis and cited an even steeper wage premium of 56% for AI-exposed jobs. So the financial incentive to climb this ladder is clearly there. The incentive for the individual is crystal clear. But the organizational support structures, well, they seem to be moving in the exact opposite direction. Here's where it gets really interesting and honestly a bit baffling. In that same workday data tracking those job requisitions, mentions of management and leadership skills actually fell by 7%.

[00:10:40] Interesting. Yeah. And mentions of training skills fell by 13%. Let me put that in perspective with our earlier analogy. It is like a gym selling thousands of advanced memberships, telling everyone they urgently need to become elite mechanics of their own bodies, but then firing all the personal trainers right when people walk through the door. If we connect this to the bigger picture of how humans actually learn, that is a fatal design flaw by these companies. Yeah, it doesn't make sense.

[00:11:06] Adult learning in the workplace relies heavily on what organizational psychologists call training transfer. The research consistently shows that having a socially supportive work environment, specifically a supportive manager and peer group, is the single strongest predictor of whether a newly learned skill actually sticks. Because learning process design and system architecture, I mean, it isn't easy. You are going to fail. Precisely. You have to experiment, you have to break things, and you have to iterate. If you thin out the middle managers and the peer trainers,

[00:11:35] you remove the psychological safety net. Wow, yeah. You destroy the very conditions that make complex re-skilling possible. And what is the human cost of that flaw? What happens to the people caught in the middle of all this? They stall out. Or they simply burn out. Microsoft's 2025 survey found that 80% of the global workforce already says they lack the time or energy to do their current day jobs. 80%? Yes. So if leadership suddenly demands that they also learn complex process design

[00:12:02] on top of that full workload, but without offering protected time or a manager to guide them, all that uncertainty just metastasizes into anxiety. And that leads to a massive divide, right? It does. There is a very real risk right now of creating a two-speed workforce. Boston Consulting Group found that while 72% of people overall use AI regularly, regular use among frontline workers has completely stagnated at 51%. Oh, man.

[00:12:30] The desk workers who have more autonomy over their time, they pull ahead and become builders. The frontline workers get left behind, and the economic inequality gap just widens further. Okay. So seeing that gap laid out, it's fairly intimidating. But the research doesn't just leave us dwelling on the problem, thankfully. Westover outlines specific evidence-based solutions, showing how top-tier companies are actually fixing this and helping their people cross that bridge. Right. There is a playbook here.

[00:12:59] So if you are leading a team, or, you know, if you're just trying to figure out how to navigate your own upskilling, how do you actually do it? It starts with visibility. You cannot close a skills gap if you don't even know what skills your team currently has versus what they actually need. Makes sense. Most organizations still talk about AI skills as this vague, monolithic thing. Companies need to build a live skills map, what HR calls a skills-based architecture.

[00:13:26] But what does a live skills map actually mean in plain English? It means moving away from static, yearly job descriptions. The paper points out that a rigid taxonomy of skills built in early 2025 would be completely useless by mid-2026. Right. Things just move too fast. A live map defines exactly which rung of the capability ladder each specific role requires, and it updates as the technology changes. AT&T and Amazon are great examples of doing this, right? How so?

[00:13:55] Well, AT&T entirely overhauled their workforce by mapping specific future state digital skills to distinct roles, and then providing self-service pathways for employees to get there. Okay. And Amazon took it a step further, investing $1.2 billion in tiered training. They recognized that a frontline warehouse worker needs an entirely different pathway and different support than a corporate data scientist. Which naturally leads to the next massive hurdle, which is just finding the time.

[00:14:25] Right? And I was reading about how Walmart is handling this, and it actually surprised me. Oh, yeah? Yeah. They announced a free AI certification and built in partnership with OpenAI, and specifically for their U.S. frontline and office-based associates. That's huge. It is. That is a tangible example of a company, you know, funding the training and theoretically protecting the time for employees to do it. But, you know, the research also points to the darker flip side of this coin.

[00:14:53] Ah, you're thinking of the Accenture example. Exactly. Accenture managed to train 550,000 people in generative AI fundamentals incredibly fast. Yeah, lightning fast. But their CEO was very open about the fact that they were actively pushing out people who couldn't adapt to the reskilling on their compressed timeline. Right. It shows that while you can scale training rapidly, it can take the form of brutal efficiency. It treats human beings like interchangeable parts. If you don't climb the ladder fast enough, you're out.

[00:15:22] It is a stark reminder that the broader market is not pausing to wait for anyone to feel perfectly ready. But if we look at how to actually train people effectively without just threatening their jobs, we have to look at how people learn to become builders. Right. You cannot learn how to architect an AI workflow by passively watching a webinar. It requires deliberate practice in the flow of actual work. And PwC's Digital Accelerators program seems to be the gold standard for this.

[00:15:49] Instead of just putting people in a room for a weekend hackathon that produces a bunch of sticky notes and no actual tools, they took a couple thousand staff members, gave them intense virtual learning, and then tied their projects to real daily work problems. Yes. And the key difference there is the output. These workers didn't just learn a concept. They built reusable assets. Right. They created a digital lab with over 5,500 assets, automations, data workflows, custom models.

[00:16:16] The research noted one specific validation workflow built by a participant that saved eight hours per run. Just think about that. Saving eight hours doesn't mean they wrote a faster email. It means they completely redesigned an entire business process so a human didn't have to spend a full day doing manual data entry. That is the ultimate definition of return on investment. And they achieved that by giving people safe sandboxes to experiment in.

[00:16:44] They validated their employees' skills based on the working products they actually built, rather than just having them check a box saying they completed a multiple choice quiz. But if you're a massive global company, how do you scale that level of hands-on coaching without your HR department having a collective breakdown? You can't have central HR holding everyone's hand. No. You solve it by empowering the people who are already doing it. Moderna is the fascinating case study here.

[00:17:13] They wanted full AI adoption and they wanted it fast. Okay. So instead of relying on top-down training, they ran an internal contest to find their top 100 power users. They took those people and officially designated them as generative AI champions, embedding them in every single business line. And because they are embedded, they actually understand the daily pain points of their coworkers. An HR trainer might not know how the supply chain team works, but a peer champion does. Precisely.

[00:17:41] And the result of that peer-to-peer model? In just two months, Moderna employees built 750 custom GPTs. That's incredible. By identifying the people who are already geeking out over this technology, protecting their time to teach others, you create a local, socially supportive environment where it is safe to fail. JPMorgan Chase did something very similar, embedding super users, roughly 10 to 20 percent of their staff, directly into teams as local experts. Okay.

[00:18:10] I am following you completely on the training and the peer champions, but I have to push back when it comes to measuring if any of this is actually working. Okay. Sure. The research talks about aligning metrics with outcomes, but honestly, if you were a manager, wouldn't it just be infinitely easier to check the software dashboard? Just track how many AI prompts an employee writes a day or how often they log in. Why overcomplicate the measurement?

[00:18:35] This raises an important question, and it is a trap that so many well-intentioned companies fall into. Measuring raw usage is incredibly easy, but it is dangerous. Dangerous? Yeah. Think back to the jagged frontier we discussed earlier. If you reward an employee simply for generating a high volume of AI outputs, you are actively incentivizing them to use the tool even when they shouldn't, when the task requires nuanced human judgment. Ah, I see.

[00:19:02] They will use the AI to write a one-line internal message just to hit their quota, even if the output is bloated garbage. Exactly. They are optimizing for the metric, not the work. The paper contrasts two deeply revealing examples of this dynamic. Take Shopify. Their CEO, Toby Lutke, took a radical approach. What did he do? He told his staff they actually had to prove a job could not be done by AI before they were allowed to request additional headcount. Oh, wow. Yeah. He forced them to deeply analyze their own workflows.

[00:19:31] He even asked teams to imagine what their department would look like if autonomous AI agents were already functioning members of the team. That forces you out of the prompting mindset and immediately into that stewardship and building mindset. You have to think about the system. Right. Now, compare that outcome-based thinking to Duolingo. Their CEO, Luis Van On, initially announced they would evaluate employees based specifically on their daily AI usage. Which sounds logical, right?

[00:19:59] Encouraged adoption. But a year later, he had to completely backtrack. Why? What happened? Because the metric was forcing people to shove AI into tasks where it simply didn't fit, just to look good on their performance reviews. They were measuring the tool, not the capability. That makes total sense. To truly succeed, leadership has to measure the outcome, better quality work, faster cycle times, and most importantly, the human judgment to know when to turn the machine off. That is such a vital distinction.

[00:20:28] Okay, as we pull all of these threads together, Westover's paper closes by looking at the long game. What is the ultimate takeaway for building durable capability in this environment? He synthesizes it down to three core pillars. First, continuous skills intelligence. A static yearly skills map is dead on arrival. You need continuous infrastructure feeding real-time market signals into your internal training. Right. Keep it live. Second, we have to rebuild the teaching organization.

[00:20:56] We must explicitly value and protect the time of the managers and peer champions who actually coach others through the messy process of learning. Stop firing the personal trainers. Exactly. And third, and perhaps most crucially, we have to cultivate human judgment and stewardship above everything else. Because the specific software you learn today will inevitably change tomorrow. But the habit of continuous learning and the wisdom to govern the risk that holds its value indefinitely.

[00:21:24] So, what does this all mean for you listening right now? It means you need to stop viewing basic AI literacy as the finish line. Look at your current role, no matter what department you are in. Ask yourself honestly, are you just prompting the machine or are you designing workflows? Are you stewarding the process and helping your peers adapt? Exactly. Because the gap between what companies need and what they provide is real.

[00:21:50] And the market is paying a massive premium for the builders. And I want to leave you with one final provocative thought to ponder. The paper mentioned a fascinating detail from a study by Brynjolfsson. They looked at over 5,000 customer support agents using a new AI assistant. Okay. What they found was that the AI tool effectively absorbed the quote-unquote tacit practices, the unwritten secret sauce,

[00:22:16] the unique problem-solving techniques of the absolute top performers on the team. Wow. It then distributed that elite knowledge to everyone else, rapidly lifting up the lower-skilled workers. Which sounds phenomenal for the company's bottom line and great for the lower-skilled workers who suddenly get a huge performance boost. It absolutely is. But here's the question for you to mull over today. If you are the top performer right now and a builder inside your company observes your workflow,

[00:22:46] captures your hard-earned brilliance, and turns it into an automated assistant that is handed out to the entire team, what becomes your next competitive advantage? Man, that's heavy. When your current brilliance is commoditized and distributed by AI, how do you stay indispensable? Wow. Brings us right back to that initial analogy. If the car is driving itself and the machine has learned all your best routes, you better know how to tune the engine.

[00:23:14] The era of the prompt is in the past. It's time to start building. Thanks for joining us on this deep dive. We'll catch you next time.