AI at work isn’t a technology problem.

If you’ve kicked off an AI initiative in the last eighteen months and it hasn’t exactly set the world on fire, you are in overwhelmingly good company.

A July 2025 study from MIT’s Project NANDA looked at 300 enterprise AI deployments and found that 95% of pilots delivered no measurable financial return. Of the $30–40 billion companies have poured into generative AI, only about 5% of integrated deployments are producing real P&L impact.

Before you blame the models: don’t. The technology is fine. Claude, GPT, Gemini — these are wildly capable tools that were writing passable marketing copy in 2023 and are now drafting legal briefs, debugging code, and sitting in on sales calls. The models aren’t the bottleneck.

The bottleneck is us.

The real problem: you’re bolting AI onto 2005

Most mid-market AI rollouts follow a familiar shape. Someone at the top gets excited about a demo. IT or Ops buys a tool. It gets launched in an all-hands. A few early adopters love it, most of the team clicks around for a week, and three months later the license is paid for and barely used.

That’s not a technology failure. That’s a workflow failure.

MIT’s researchers called this “the GenAI Divide” — the gap between the 5% getting real value and the 95% who aren’t. And the divide doesn’t run along the line of who bought the best model. It runs along the line of who bothered to redesign how work actually gets done.

This is the part worth pausing on: AI doesn’t improve a bad process. It just runs it faster, politely, at scale.

Here are three things to think about if you want to land on the right side of that 95/5 split.

1. Redesign the work, not just the tool

PwC calls it the 80/20 rule of AI: the technology delivers about 20% of an initiative’s value. The other 80% comes from redesigning how work actually flows through the organization.

Consider a UK banking group Bain profiled recently. They had a product-change process taking 60–100 days with more than ten handoffs. They could’ve slapped a chatbot on top and shaved off a few days. Instead, they asked a better question: what if we did this in one day? Four months later, they had — by rebuilding the entire workflow around what AI now made possible.

That’s the move. Not “where can we add AI to what we already do?” but “what should stop, simplify, or completely change because AI exists now?”

For a mid-market company, this doesn’t mean a grand enterprise-wide reinvention. It means picking one workflow — onboarding, invoice processing, quoting, sales follow-up, customer support triage — and asking whether the steps in it still need to exist at all.

Most of them don’t.

2. Your people are the project

Here’s the most interesting finding in all the 2025 research, and it should quietly rearrange how you think about your own company:

Only about 40% of companies have official AI subscriptions. But inside those same companies, roughly 90% of workers report using personal AI tools like ChatGPT or Claude for work tasks every day.

Your team is already using AI. They’re just not using yours.

This shadow AI economy isn’t rebellion. It’s a tell. It means your people want to use these tools. They just don’t trust that your official rollout is going to help them, protect them, or stick around long enough to be worth learning.

What actually moves the needle:

• Training. McKinsey found that 48% of U.S. employees would use AI more often if they had formal training, and 45% would use it more if it were integrated into their daily workflow. Those are the two biggest motivators employees name, and both are things leadership directly controls.
• Champions, not mandates. Identify the enthusiastic early users in each function and give them extra support, air cover, and public recognition. Top-down AI adoption has a worse track record than most top-down anything.
• Psychological safety. If your team thinks AI is how you plan to replace them, they will quietly sabotage it. If they think it’s how you plan to let them stop doing the worst part of their job, they will champion it. Tell them which one it is — clearly, and more than once.

3. Go narrow before you go wide

The most visible AI failure pattern of 2025 was what analysts started calling “perpetual piloting” — dozens of proofs-of-concept, a big-deck AI strategy, and not a single production system actually running at scale.

The fix is boring: pick one vertical workflow, solve it end-to-end, prove the ROI in dollars, and then extract the reusable pieces for the next one.

This sounds obvious, but it runs directly against how most mid-market companies are approaching AI. They want an “AI strategy.” They want an “AI platform.” They want a tool that does a little bit for everyone. What they end up with is a tool that does a little bit for no one.

One workflow. End to end. With measurable outcomes. That’s the real unit of AI progress in a mid-market company. Not a platform. Not a strategy deck. A workflow that used to take three days and now takes three hours, with the savings showing up somewhere the CFO can see them.

Then the next one. Then the next one.

The part nobody wants to hear

AI rollout isn’t a procurement problem, an IT problem, or a model-selection problem. It’s a change management problem. And change management is famously hard — which is probably why so many companies are quietly hoping this turns out to be a technology problem instead. It would be so much more convenient if it were.

It isn’t.

The good news is that if you treat AI like any other major operational change — real executive sponsorship, workflow redesign, training, champions, clear metrics — you will quietly end up in the 5% while your competitors spend another year running pilots that go nowhere.

The technology is ready. The question is whether the organization is.

---

If you want a sober look at where your AI efforts are stuck — and what it would take to get them unstuck — that’s what we do at Stealthy Good. No grand strategy decks. No AI theater. Just a clear plan for the one or two workflows where change would actually matter. Let’s talk.

Sources

• MIT Project NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025)
• PwC, 2026 AI Business Predictions
• Bain & Company, Want More Out of Your AI Investments? Think People First (February 2026)
• McKinsey, Reconfiguring Work: Change Management in the Age of Gen AI (August 2025)
• BCG, AI at Work 2025: Momentum Builds, but Gaps Remain
• RSM US, Middle Market Firms Rapidly Embracing Generative AI (2025)