There's an old phrase I can't stop thinking about. "Paving the cow-path."

Picture it. A cow wanders across a field, taking the long, winding route. Over the years a dirt track forms where the cow walked. And then one day, someone comes along and pours concrete over that exact winding track — and calls it a road.

It's smoother. It's faster. But it still goes the long way round, because nobody asked whether the path made sense in the first place.

That is what most AI adoption looks like right now. We are pouring very expensive, very clever concrete over cow-paths.

The bolt-on trap

Here's how it usually goes. Someone identifies a painful process. It has too many steps, too many hand-offs, too many people involved. The natural instinct: let's throw AI at it. Automate the steps. Speed up the hand-offs. Make the whole thing run faster.

And it works — in the narrowest sense. The process that used to take three days now takes three hours. The report that took a team of four now takes one person and a prompt. Everyone feels clever.

But six months later, the underlying problem is still there. The process was broken before AI touched it. AI just broke it faster.

BCG puts this bluntly. When you layer AI onto your existing structures — the same workflows, the same approvals, the same forms — you get only incremental gains. A few minutes here. A slightly faster draft there. Their advice: remove the unnecessary work before you digitise it.

Most organisations do the opposite. They digitise, then wonder why nothing changed.

The question nobody asks

The uncomfortable question is this: should this process exist at all?

Not "how do we make it faster?" Not "can AI do step four?" But: what would happen if we just stopped doing this?

I've watched teams agonise over automating a monthly report that nobody reads. I've seen sprints dedicated to building an AI pipeline for data that was never accurate enough to trust in the first place. I've been in meetings where someone finally asks "who actually uses this?" and the room goes quiet.

The most powerful AI move is often the one that writes zero lines of code. It's the decision to kill the process entirely.

What redesign actually looks like

When IBM describes what they call AI-first operations, they draw a clear line. In a redesigned process, the agents run the execution — the gathering, the sorting, the drafting, the routine moves. And the people are freed up to do the judgment — the deciding, the checking, the exceptions, the things that need a human who is accountable.

This isn't "humans do the same job, a bit faster." It's "the work gets re-divided." Machines take the execution. People move up to judgment and oversight.

That is a redesign. Not a bolt-on.

The test

Here's a simple test for any process you're tempted to point AI at. Three questions:

One. If you removed every step that exists only because of an old constraint — a system limitation, a legacy approval, a "we've always done it this way" — what's left?

Two. Of what remains, which steps actually require human judgment? Not human presence — human judgment. A decision where the consequences of getting it wrong matter.

Three. For everything else — the gathering, the sorting, the drafting, the routing — what happens if you just let the agent do it, and put a human at the decision points?

The answer to question one is usually the most revealing. In most processes I've worked on, somewhere between a third and half of the steps exist only because of constraints that no longer apply. They're fossils. And if you automate fossils, you just get faster fossils.

Pour the concrete where it matters

AI is extraordinary. But it's not a process designer. It will happily automate the workaround without ever questioning the thing the workaround was working around.

That's your job. Before you ask what AI can do for your process, ask whether your process should exist at all. Kill what you can. Redesign what's left. Then bring in the tool.

Don't pave the cow-path. Redraw the map.