I can’t remember the last time I sat down with a management team that didn’t mention AI somewhere in the first ten minutes. A copilot for customer service, a drafting tool for the sales team, an internal chatbot trained on the company’s own documents. Almost everyone has a pilot running somewhere. What almost nobody can show me is the line in the P&L that pilot actually moved. That gap, between using AI and AI changing the numbers, is where most of the disappointment sits right now. It’s also where most of the real opportunity still is.

Why so many pilots stall out

The pattern repeats across industries often enough that it’s worth naming plainly. A pilot gets built around a tool instead of a workflow, so it speeds up one step without touching the process wrapped around it. The team running the pilot has no authority to actually retire the manual version of the task, so both versions keep running side by side and the cost never really drops. Or the pilot works, but only in the one team that built it, and nobody owns turning that early win into something the rest of the company can use.

What’s really going on underneath this

None of this is really an AI problem. It’s an old, familiar operations problem wearing a new coat. Most companies still fund technology projects and operating changes through two separate budgets, run by two separate teams that rarely talk. Add to that the pressure every management team feels to show a board that “we’re doing something with AI,” and you get a lot of activity that was never designed to hit the income statement in the first place.

What the companies that get it right do differently

• They start from the cost or the constraint, not the tool. The brief is “cut the cost of handling a support ticket” or “let the current sales team cover more accounts,” never “go deploy a copilot.” That framing ties the initiative to the P&L from day one and gives it an owner with a budget to defend.
• They redesign the workflow, not just the task. Automating one step inside a process that otherwise stays the same almost never frees up real headcount or cost. Rebuilding the approvals, handoffs, and staffing around the new capability is what actually changes the operating model.
• They measure the before and after, on purpose. Cost per unit, cycle time, output per employee, tracked with the same discipline as any other operating metric. That’s what lets a board tell a real productivity gain apart from a good demo.

An example that plays out constantly

Take a mid-sized services company that rolls out an AI tool to draft first responses to customer emails. Six months later, leadership proudly reports that agents are using it every day. Sounds like a win, until you ask the follow up question: has headcount in the support team changed, has average handling time actually dropped, has the cost per ticket moved at all? Usually the honest answer is no, because the tool sped up drafting but the agent still reviews, edits, and sends every message the same way they always did, and the team is still staffed for the old process. Compare that to a company that used the same tool to fully resolve routine tickets without a human touch, then reassigned two agents to handle the complex cases that were previously getting backlogged. One of these is a pilot. The other is a P&L result.

The bottom line

“We’re using AI” tells you almost nothing on its own, because nearly every company can say yes to it right now. The question worth asking in diligence or in a portfolio review is narrower: which specific cost line or capacity limit was this meant to move, what was it before, and what is it today.

The lesson

A management team that can answer those three questions clearly, cost line, baseline, and current number, is usually showing you an operating discipline that shows up everywhere else in the business too. That discipline is worth more than the AI tool itself.

Assessing AI-driven operating improvement in a portfolio company, or building the business case for one? Discuss a mandate →