At HumanX in San Francisco earlier this year, Andrew Ng made a point that reframed how many in the room were thinking about enterprise AI. Ng built the AI infrastructure at Google Brain and Baidu before founding DeepLearning.AI and Coursera, which now serves roughly 148 million learners globally. He is someone whose read on where AI creates organizational stress has earned its credibility.
His teams, he said, now expect two engineers to deliver in a month what previously required fifteen engineers over three. They are responding by hiring more engineers, not fewer, because the idea backlog has outrun the capacity to execute.
The implication is easy to miss. The bottleneck has moved. It is no longer in engineering capacity. It is decision-making speed. Engineers finish work and ask, “now what?” Product managers, not developers, have become the scarcest resource in the AI era. And the organizations pulling ahead are not the ones with the most AI tools. They are the ones that figured out how to make faster, better decisions about what to do with what AI produces.
This is more significant than it first appears. For three years, the dominant enterprise AI conversation has been about execution: Which tools to deploy, how to drive adoption, how to manage risk. Those are real questions. But they are not the binding constraint anymore. The binding constraint is judgment. How fast can the organization decide what to scale, what to fix and what to stop?
The visibility problem is a decision problem
What Lanai sees in customer data makes this concrete. One revenue operations team had 140 reps using AI across three regions. One rep had built a renewal outreach workflow that outperformed the team average by 110 times. Leadership had no idea it existed. There was no system to surface it, no way to connect it to the pipeline and no path to replicate it.
Once the workflow was visible, the team extracted it, deployed it as a governed agent and rolled it to all 140 reps across three regions in 72 hours. The result was 11.4 FTE of reclaimed capacity and $2.8 million in the affected pipeline. The technology was not the challenge. The decision was and that was only possible once someone could see what was actually happening.
The same pattern appears on the cost side. A customer in IT and security discovered 23 AI tools running across six departments. Nine were completely ungoverned, with customer PII flowing through personal accounts. Three enterprise licenses sat at under 8% utilization. The tools existed. The spend existed. What did not exist was a single place to see what was critical versus redundant and make a call. Once that visibility existed, they consolidated to 14 governed tools and cut $340,000 in shelfware. Not by deploying new technology. By making a decision they previously could not make because they lacked the information to make it confidently.
Coursera’s own retention data points in the same direction from a different angle. Employees who completed AI training were retained at 50% higher rates than those who did not. The most valuable output was not task efficiency. It was that people who understood what AI could do started generating better ideas about what to do with it. The upside was clearer thinking about direction, not faster execution of the same tasks.
The work itself is changing faster than the org chart
There is a deeper structural issue underneath the visibility problem. The org chart and the income statement, the two systems enterprises use to understand who does the work and what it costs, were both designed in an era when the answer to “who does the work?” was so obvious nobody bothered to say it out loud: a human being.
McCallum’s 1855 railroad org chart mapped thousands of people across miles of track. Pacioli’s double-entry system evolved into the profit-and-loss account so merchants could see what was left after paying people, not processors. The industrial revolution added depreciation to admit that machines do work over time, but even that assumed workers were either people or large pieces of hardware. It never imagined a world where software itself is on the shop floor, doing the work as operating labor.
AI is operating labor in a software costume. And neither the org chart nor the P&L was built to see it.
When an agent handles ten thousand support tickets, it appears on the P&L as software expense. When a human did that work, it was labor. The substitution is real but registers nowhere official. Based on observed activity across Lanai B2B SaaS customers, here is where knowledge work actually sits today, and where customers predict it is going by 2028:
| Level | Definition | Today | 2028 |
| L1 | Work only a human should own: accountability, trust, novel judgment | 45% | 20% |
| L2 | Human does the work; AI assists and accelerates | 35% | 30% |
| L3 | Agent executes; human reviews the output | 15% | 35% |
| L4 | Agent runs end-to-end; no human required | 5% | 15% |
Source: Lanai customer data, 2026. Prediction: Lexi Reese, CEO Lanai.
The L2 peak-and-decline finding is the most counterintuitive and the most important. L2 does not decline because AI assistance gets worse. It declines because the best-adopted L2 workflows get promoted out of it. The question for any organization is not how to stay in L2. It is which L2 workflows are ready to move, and whether the organization has the data to make that call deliberately rather than accidentally.
By 2028, L3 will become the modal form of knowledge work. The most common configuration will be an agent executing a task while a human decides whether it was done right. That is a fundamentally different job description than what most knowledge worker roles were designed around. And it is arriving faster than most org designs are prepared for.
The management problem nobody budgeted for
Most enterprise AI dashboards are not built to surface any of this. They track adoption rates, active users and tasks completed. Those metrics measure activity. They do not measure whether the organization is converting AI activity into decisions, and decisions into results.
The gap shows up most visibly when the CFO asks what the company got for its AI spend. Token spend that was $5,000 a month eighteen months ago is $40,000 today in many organizations, with no clean story attached. The executives who survive that conversation are not the ones who spent less. They are the ones who built the translation layer between spend and outcome before anyone demanded it. Tokens to threads. Threads to tasks. Tasks to time saved. Time saved to business result. That chain exists in the data. Most organizations have not assembled it.
The CIOs who will have the clearest story for their boards in 2026 are the ones who treated AI deployment as a management problem from the start, built the systems to connect AI activity to the business outcomes they are already accountable for, and developed the organizational habit of acting on what they see.
Execution stops being the constraint. Judgment becomes the scarce resource. The organization that was slow because humans could not execute fast enough is now slow for a different reason: Not enough people who can make the right call under genuine uncertainty. Most org charts are designed to consume judgment, not develop it.
Leaders approved the tools. The ones pulling ahead are the ones who decided to own the outcomes.
This article is published as part of the Foundry Expert Contributor Network.
Want to join?
Read More from This Article: The real AI bottleneck isn’t what you think
Source: News