What Big Enterprise is already doing with AI (what it means for you)

I spent ten years selling software into big companies. Which means I spent ten years in the offices where the enterprise version of whatever's coming next gets sold first — the boardroom pitch, the pilot, the rollout, the post-mortem eighteen months later when it didn't do what the slide said.

The useful thing about having sat on the vendor's side of that table is that you learn to read the play early. What the big end is doing this year, the mid-market gets sold next year, usually worse and more expensive. So it's worth knowing what's actually happening, not the headline, but the reality, because it tells you what's heading for you and what to do before it arrives.

Here's the honest version, stripped of the noise.

The big companies have moved past playing with AI. Two years ago it was experiments — a chatbot here, someone in marketing trying things. Now it's in production, running real work: agents that read the documents, pull from the systems, apply the rules, and actually do the task rather than just answer a question about it. Salesforce, Microsoft, ServiceNow — the platforms your bigger competitors buy from — have rebuilt their whole model around it. Microsoft alone has hundreds of thousands of these agents running across companies right now. This isn't coming. It's here, at the top end, today.

Now the part nobody puts on the slide. Most of these projects are struggling, and not for the reason you'd think. They're not failing because the AI isn't clever enough — the AI is plenty clever. They're failing on three things, and I want you to read these carefully, because they're the whole point of this piece:

They're failing because the data was a mess. The AI is only as good as what you feed it, and most big companies discovered their information was scattered, inconsistent, and half-wrong the moment they tried to build on it.

They're failing because nobody owned the edge cases. The agent handles the standard job fine, then hits the one-in-five job that's unusual, and there's no clear answer to "who decides what happens now?" So it guesses, and gets it wrong, and someone finds out later.

And they're failing because nobody built the part where a human checks the work before it goes out. They automated first and asked "who's responsible when this is wrong?" second — which is the wrong order, and an expensive one.

The companies getting it right are doing the opposite of the big splashy rollout. They pick one job — one well-defined job with good, clean information behind it — put the AI on that, measure whether it actually worked, and only then do the next one. Narrow, proven, expanded. Not broad, hopeful, and prayed over.

Read those failures and that fix back, and here's what should jump out at you: the big end is spending millions of dollars and eighteen months each learning a lesson you already know in your bones.

You already know your data is messy — it's in your head and a dozen spreadsheets, but you know where the bodies are. You already know the edge cases, because you're the one who handles them; the unusual job doesn't get guessed at in your business, it comes to you. And you already know a number shouldn't go out the door without someone who knows what they're doing signing off on it — that's just how you've always run. The three things sinking million-dollar enterprise projects are three things a good family operator does by instinct.

That's the whole advantage, and it's worth saying plainly. The big end has scale, budget, and a stack of consultants, and it's still tripping over data, edge cases and accountability. You have less budget and no consultants, but you have the three things that actually matter already sitting in the business: you know your work, you own your judgment, and you're already the human who checks it. The enterprise play, done properly, is just: start narrow, on a job you understand, with you signing off until you trust it. You don't need a million dollars to run that play. You need to run it before the watered-down, oversold version of it turns up in your inbox with a logo on it — which it will.

So the one move, before that happens: pick the single job in your business where the same kind of decision gets made over and over, where you've got real history behind it, and where you'd sleep fine knowing a system drafted it as long as you still signed it off. That's your one narrow job. That's where this starts — not with a platform, not with a big rebuild, and not with whatever the enterprise vendors will eventually try to sell you. With one job you already understand better than any software company ever will.

I've been on both sides of this — the enterprise floor where the play gets sold, and the operations floor where the work actually happens. The gap between those two rooms is where most of the money gets wasted. It doesn't have to be wasted at your end.

I've built a short assessment for multi-channel family operations working out where AI actually helps — and which part of their edge is worth capturing first. Ten questions. The readout comes from me, not a machine.