This article from Vice President of Technology Michael Lesher was originally published to his LinkedIn page on October 22, 2026.

Two or three years ago, one number got repeated everywhere: AI would drive a 40% increase in productivity. It’s worth holding that promise up against what’s happened since.

I led the rollout of AI tools across our team at ITX, watched closely to see where it failed, and built the governance to close those gaps as they surfaced. Some of what we found about AI productivity lines up with the research. But also, a lot of it doesn’t.

Four specific predictions got made about how AI would change work, on productivity, coding speed, return on investment, and autonomous agents taking on complex tasks. Two years later, the record is mixed — especially at enterprise scale, where complexity has a quality of its own.

Claim #1: AI Will Increase Productivity.

In 2023, McKinsey projected a 15-40% increase in productivity from artificial intelligence, estimating AI could add $2.6 trillion to $4.4 trillion in annual corporate profits globally.

Did it hold up? Not quite. MIT’s 2025 report, The GenAI Divide: State of AI in Business, found that 95% of enterprise AI pilot programs failed to show any measurable impact on profit and loss. That’s enterprise organizations specifically, not small, early-stage startups.

McKinsey wasn’t wrong so much as incomplete. The report didn’t account for the logistics of implementing AI at scale. If 95% of enterprise initiatives failed while startups often succeeded, the difference comes down to a narrower product built around one specific problem, and a workforce with less institutional processes to unlearn.

Overhauling a workflow that already works, however imperfectly, is slower and harder than most leadership teams expect. Retraining people and rebuilding trust in a new process, not buying the software, is where the real cost shows up.

Claim #2: AI Coding Assistants Will Make Developers Faster.

This was one of the easier claims to rally behind, rooted in a specific number, and not built from expectations in the flashiness of new technology. In 2023, a controlled experiment within GitHub Copilot found developers with access to AI completed a task 55.8% faster than those without.

Two years later, Google’s 2025 DORA report backs up that claim at a larger scale. AI adoption is now positively linked to software delivery throughput, (this is, interestingly, a reversal from their 2024 findings, showcasing how much we are learning). But delivery stability moved the opposite direction, hiding costs of AI that then showed up later.

I see this constantly. AI output reads as more confident and more finished than it is, which is what makes it dangerous to trust immediately. On day one of a project, a developer using AI looks like they are flying. By day five, most of that lead is gone, spent fixing what looked right the first time. Few people stopped to ask whether the work was being checked at that new speed.

Claim #3: Investment in AI Will Be Funded Through Its ROI.

By 2023, McKinsey wasn’t just projecting a productivity bump, it was putting a price on it: up to $4.4 trillion in annual corporate profit. That kind of number gets AI projects greenlit.

Perhaps too fast, in fact. That confidence hasn’t held up under enterprise scrutiny. MIT’s research found that companies poured budget into sales and marketing tools while back-office automation went underfunded, even though it delivered the strongest returns. Many also tried building their own tools in-house rather than buying, and those efforts succeeded only about half as often as the ones they paid a vendor for.

We’ve seen a version of this ourselves. Someone spent real time and budget building a tool that already existed elsewhere in the business, and the result went mostly unused. Without a clear target, AI can erode productivity rather than add it, once you count the wasted spend and the cleanup that follows.

Claim #4: Agentic AI Will Handle Complex Work Autonomously.

OpenAI’s CEO Sam Altman posted a reflection to his personal blog in early 2025, mostly a victory lap on how far OpenAI had come. Buried in it was a specific, bold prediction:

“We are now confident we know how to build AGI as we have traditionally understood it. We believe that, in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies.”

AGI (that’s Artificial General Intelligence) isn’t the same as the narrow AI agents most people use today. It describes a system that can handle multiple tasks and expand its own skills, going beyond skimming a document or summarizing an email.

Six months later, Gartner estimated that over 40% of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls.

I’ve watched this pattern before: AI moving faster than the structure built to support it. Agentic AI is an accelerator, since the whole point is doing more without someone checking every step. Fewer checkpoints mean a mistake gets further before anyone notices.

What’s the Connection?.

My takeaway from the on-the-ground reality of AI implementation in 2026 is that you’ve got to spend real time thinking through what your AI project will deliver and how.

The positives that got the most attention came from small startups and individual power users, and that gave everyone rose tinted glasses. Nobody stopped to ask whether an enterprise organization could duplicate those efforts and get the same outcomes. The conditions were never the same.

Enterprises pay a higher price than startups or individuals: people switching processes, training on new tools, or learning a workflow that takes real time to perform well is expensive and time-consuming. On top of that comes the expectation to deliver more while every output gets checked more closely. That combination is already showing up as burnout, and the burden eventually travels up to leaders managing both the business goals and the people trying to hit them.

A large organization can’t expect an entrepreneur’s results without an entrepreneur’s conditions. Without guardrails built for the enterprise context specifically, the gap between individual and organizational outcomes is expected, not surprising.