Sam Altman admitted that AI adoption is slower than he thought. We have not had " the iPhone moment, " he said, meaning we have not reached the point where the technology stops being a tool and simply becomes the way we work. He blamed economic inertia: humans are too change resistant. One day later, news surfaced that Meta had quietly abandoned its plan to restructure its workforce around AI agents. The internal Project that was called “OT” (e.g. office transformation) would have cut some teams by up to 60%, replacing them with small human pods supervising AI systems. Mark Zuckerberg stopped it.

Two of the industry's biggest players, days apart, admitting the same thing. Deployment is harder than the pitch decks say.

Does More Code Means More Value?

None of this means the underlying promise is fake. AI creates more every week. More content, more code, just more. Delivery Hero recently claimed its "HeroGen," a team of AI coding agents, produces output that matches roughly 130 developers. The question I always ask myself is whether that means more code or more value, an important distinction. The pattern I keep seeing is that AI creates more text, more emails, more content. But does it create decisions? Often not. AI has no conviction. As I put it after teaching an eCornell workshop on this exact gap, the winners in AI will not be the ones with the flashiest demos, they will be the ones who turn messy work into reliable workflows.

This is not a doomer argument. I've spent the last three years applying AI in marketing and e-commerce, and I've written before about the mistakes brands make chasing this shift . AI does simplify workflows when it's done well. A workflow that took a team weeks now runs semi-autonomously. Suddenly companies can manage far more products and far more differentiated channels than their team ever could by hand. The clear winners are the ones using AI to win new channels and build an AI-supported operating model, not the ones bolting AI onto what they already had.

So why did Meta’s own pilot collapse? The same reason only 6% of companies say they fully trust AI agents to run core business processes, according to a 2026 Harvard Business Review survey . Germany’s adoption numbers of AI tell a similar story. The Ifo Institute found AI usage among German firms jumped from 40.9% to 54.5% in a single year. That’s not the AI takeover some people feared. It’s the opposite. Companies are using AI but only slowly. They just don't trust it yet.

Sam Altman, and I rarely agree with him , is right about why. It's not the models' fault. Adopting AI into a workflow is hard because most workflows carry a lot of undocumented exceptions, the workaround everyone quietly does that never made it into a single SOP. For AI to actually help, someone has to do the unglamorous work of mapping what a company actually does onto what AI can actually do.

The Tale Of The Self-Driving Car

I used to think my kids would never need to learn how to drive. It was taking longer than anyone expected. Early self-driving cars knew quickly the rules of the road, but they still could not merge onto a real freeway, because humans don't drive by the rules. We drive by habit, courtesy and a thousand small exceptions. Engineers had to sit in the car and watch, over and over, to map out how people actually behave before the system could handle the street.

FDE - Forward Deployed Engineer

That mapping work now has a job title: Forward Deployed Engineers (FDE). FDE’s postings are up 729% year over year. Call it "solutions engineer" with a new name: solutions engineering built for machine learning. Palantir pioneered the role. OpenAI, Anthropic, Google, Salesforce and, OMMAX are all now hiring for it at scale. An FDE embeds inside the client, not to ship a generic product, but to sit inside the actual workflow long enough to learn what the org chart and the vendor contract leave out.

From Engineer To Operator

How much FDE an organization needs versus traditional engineering is already being debated. Go back to the self-driving car. At first it took a trained engineer behind the wheel, not to drive, but to watch, correct and feed real human behavior back into the system. Driving data alone was not enough. The gap was too wide, and engineers adjusted the system itself because every edge case mattered. As the technology matured, the seat needed less engineering and more ordinary judgment. I made this same point after talking autonomy warfare with Swarmer’s Alex Fink: the real question is never AI versus no AI, it’s how much autonomy a system should have , and that applies as much to any business AI agent as it does to a warrior drones.

I expect FDE roles to follow the same arc. Today, closing the gap between a company’s workflow and an AI system takes someone close to a machine learning engineer. As the tooling and governance layers catch up, that seat should need less pure engineering and more domain fluency: a marketer trained in AI, a finance person trained in AI, not an ML engineer parachuted in from outside.

We don't yet know if forward deployed engineers, or whoever replaces them as the role matures, fully close AI's deployment gap. What Altman and Meta just confirmed, in the same week, is what does not work: assuming the model takes over and the org chart rearranges itself. The blueprint, lower cost and more channels, is real and worth building toward. But right now the only credible way there runs through someone sitting inside the workflow, doing the unglamorous work no platform will do for you.