The happy path proves nothing

Hot take to start: if your AI only works for most of your cases, you haven't proven it understands your business. You've proven the technology works. Those are not the same thing.

A LinkedIn thread this week sharpened this for me. AI scientist and Cohorte founder Charafeddine Mouzouni argued in a post that an AI agent can only join a team that knows how its own work gets done. Robert Heckel, CEO of Computer Visionaries AI, went further in the comments: documenting the process isn't enough; you have to capture why experienced people break it. Then, replying to my comment, he put it better than I could: the happy path proves the technology works. "The exceptions prove whether it understands the business."

He's right. And the data agrees with him.

Ready models, unready companies

In January 2025, Sam Altman wrote on his blog that 2025 might be the year the first AI agents "join the workforce." This August, Deloitte published a survey of 501 US leaders, from senior managers through the C-suite, conducted this spring. Only 5% said their business processes were highly prepared for AI agents. Only 15% had scaled agents across functions. Business processes ranked as the least-prepared area of all.

The detail that stopped me: most leaders Deloitte interviewed said they're layering agents on top of existing processes instead of redesigning them. That's the old way of thinking in a new outfit.

What I've seen, in private companies and federal agencies

I've watched the same pattern play out in both the private and federal sectors. One team finds a tool that solves its problem. Another department buys its own. Then another. Every implementation is a success on its own terms. Every team gets its dashboard, its workflow, its win.

But the systems don't talk to each other. The data lives in one place, the context in another. And the people who need the full picture most—executives and leadership—only ever see slices of it. They're asked to make organizational decisions from segmented, fragmented information. Nobody gets to look at the business with an organizational mind.

This isn't just my experience. GAO, the federal government's nonpartisan watchdog, has published a report on fragmented, overlapping, and duplicative federal programs every year since 2011. Its 2026 edition credits about $774 billion in financial benefits from fixing those issues so far, with more than $100 billion still on the table. And McKinsey's 2026 State of AI survey, which included 1,719 leaders across 97 countries, shows the same pattern playing out with AI. Eighty percent of respondents say AI has improved their individual productivity. Yet only 37% can point to any impact on their organization's bottom line, a share that hasn't moved since last year. The wins stay local. The enterprise doesn't feel them.

Here's the part that should worry all of us: we're repeating the old mistake with AI. The people are getting faster. The business isn't. Department-by-department AI is just department-by-department software with a faster engine. The roughly 6% of companies McKinsey counts as AI high performers do it differently: nearly three-quarters of them fundamentally redesign their workflows around AI, compared with about a quarter of everyone else.

Why the outliers are the whole point

In that thread, I said that focusing on outliers runs against our current culture of Time to First Value and MVP. I stand by it. We're so focused on getting it to work for most cases that we never teach it the cases that matter. Then we wonder why it can't go three levels deep.

The routine cases prove AI works. Any good demo can show you that. The outliers are where your business actually lives: the franchisee with the unusual agreement, the policy exception someone approved years ago, the rollout that never fit the playbook. Those cases carry the complicated design and policy decisions your most experienced people make without even noticing. Mouzouni described work like this as running on folklore. If the folklore never leaves people's heads, no agent, however capable, can learn your business.

That's where the human comes in. AI can now work through deep, time-consuming problems in seconds, the kind that used to mean weeks of pulling reports from five systems and stitching them together. But it can't decide which cases matter. A human has to point to the outliers because the outliers are what prove AI has learned the model and the business, not just the majority.

Think AI-forward, not AI-added

The old playbook: buy per team, implement per team, measure adoption per team, launch for the majority, and handle the edge cases "later." Later never comes.

An AI-forward approach flips it:

• Start with the enterprise question, the one leadership can't answer today, not a single team's pain point.
• Design around the exceptions first. Ask your most experienced people Robert's question: When do you break the process, and why?
• Connect before you automate. An agent that sees one department's data will give you one department's answer.
• Measure outcomes, not adoption. Logins weren't value before AI. They aren't value now.

Notably, 75% of the leaders in Deloitte's survey said they see more value in humans and agents working together than in agents on their own. The human isn't the bottleneck in this model. The human is the teacher.

AI won't fail on your easy cases. It will fail on the ones you never showed it. Implementation isn't an EVENT; it's an education.

So tell me: what's the one exception in your business that only your most experienced person knows how to handle, and has anyone ever written it down?