What It Really Takes

There's a recurring frustration in enterprise AI right now: companies are investing, pilots are proliferating, but the ROI isn't matching the ambition. Sharon Goldman opened the conversation with exactly that tension. Where are we, really? What's working and what isn't?

Andrew Ng's answer reframed the problem. Bottom-up AI adoption isn't the issue, he's a proponent of it. The problem is what it produces. "Bottom-up innovation tends to give you point optimizations," he said. A team using AI to shave an hour off a loan approval process is a real gain. But the bank that used AI to restructure the entire workflow — market, apply, get approved in ten minutes — created a new product. That's a different order of magnitude, and it requires a top-down motion to pull off. "These top-down innovation motions are hard," Ng acknowledged, "but more and more firms are figuring this out, and this will drive more business growth rather than just efficiency."

Greg Hart connected this directly to what Coursera is seeing on the demand side. Enrollment in AI courses doubled in pace — by the numbers, someone enrolled every four seconds in 2025. But the more telling signal wasn't the volume. It was the mix. Enrollment in critical thinking courses grew nearly 200% year over year. "The things that will distinguish and differentiate a given employee," Hart said, "are going to be the human skills — critical thinking, communication, teamwork." Technical fluency gets you to the table. Judgment is what you do once you're there.

The case for coding… and why it's more controversial than it sounds

The moment that drew the most reaction was Ng's argument that everyone should learn to code. He was direct about expecting pushback: senior business leaders have been telling people not to bother, on the grounds that AI will automate it anyway. His response was equally direct. "We'll look back on that as some of the worst career advice ever given."

His reasoning isn't about writing syntax. It's about leverage. AI is making it easier for anyone to build custom software, and the people who know how to direct that — marketers, recruiters, finance professionals with coding fluency — are starting to pull away from those who don't. "On my teams, the marketers, the recruiters, the finance professionals who know how to code start to outperform, and the gap is growing." The nuance matters: he's not talking about hand-writing code. "Don't write code by hand. Get AI to do it for you, and you'll be more powerful that way."

Learning by doing, and why courses alone aren't enough

Goldman pushed on something that comes up consistently in enterprise AI conversations: the individual dimension of skill development. AI adoption is personal in a way that traditional corporate training isn't. You have to use the tools to understand what they're capable of — and where they break down.

Hart's response drew on what Coursera sees in learner behavior. Courses with high interactivity — ones that force experimentation rather than passive absorption — produce better outcomes. But the biggest differentiator is engagement with Coursera's AI tutor. Learners who use it don't just perform better on assessments. They develop a working mental model of the technology's actual capabilities and limits. "That's so critical," Hart said, "as they're trying to think about how they could apply this in their role."

The enterprise data backed this up. A large global technology company saw employees who engaged with Coursera retained at a 50% higher rate than those who didn't — a result nobody had anticipated. The explanation: those employees were becoming more effective in their roles and advancing in their careers. Hart also noted that 46% of Coursera learners report a salary increase after taking an AI course, and 91% report a positive, tangible career outcome within six months of completing one.

The PM bottleneck nobody's planning for

One of the more unexpected observations came from Ng's description of a structural shift playing out on his own teams. As engineering accelerates, the constraint moves. Work that previously required fifteen engineers over three months he now expects two engineers to deliver in a month. The response on his teams isn't cuts — it's more hiring, because the list of worthwhile projects grows when execution gets cheaper. But it creates a new problem: product management hasn't kept pace.

Engineers finish and ask what's next before PMs have the answer. "We're increasingly bottlenecked not by the act of building, but by deciding what to build." His teams are responding by hiring more product managers and asking engineers to take on more PM work. The implication for workforce planning is significant: the skills that become scarce in an AI-accelerated organization are often the ones nobody thought to develop.

The entry-level question

Goldman raised the anxiety that's generating the most headlines: college graduates struggling to find jobs, the gap between what universities are teaching and what the market needs. Ng's answer was blunt. "We can't hire enough smart, fresh college grads that are skilled in these tasks." The demand is there. The supply isn't, because higher education adapts curriculum slowly — faculty expertise, committee approvals, institutional inertia. Universities that are moving fastest are often turning to Coursera and DeepLearning.AI to fill the gap, augmenting their own curricula with external credentials in areas where they don't have in-house faculty.

Hart's closing advice applies to graduates and to everyone navigating this shift: be flexible, use AI in everything you do regardless of field, and think of yourself as a lifelong learner. Not a credential-earner — someone who demonstrates actual mastery of actual skills. The market signal supports this: 91% of employers say a candidate with a micro-credential reduces their training costs by 10 to 30%, making demonstrated mastery a concrete hiring advantage.

The session covered more ground than most hour-long panels. The through-line was consistent: the talent gap is real, the tools to close it exist, and the organizations pulling ahead are the ones that have stopped waiting for a stable playbook and started building the capability to keep learning as the ground shifts.

Watch the full session on-demand. 

Keep the conversation going. Register now for HumanX 2026 Amsterdam (September 22–24, 2026) or HumanX 2027 Las Vegas (March 7–10, 2027). Be sure to subscribe to The Output newsletter and save yourself from bad AI takes (and save on your HumanX ticket).