IBM's CEO Just Said Most Companies Run AI at the Margin. He's Right. Here's How to Not Be One.
Arvind Krishna stood on stage at IBM Think 2026 and said the quiet bit out loud.
“Most enterprises run AI at the margin.” Source
Translation: the big processes, the ones that actually move the P&L, are barely touched. Everyone’s running tiny experiments around the edges. Chatbot here. Summary tool there. A pilot somewhere in marketing that nobody’s reviewed since July.
The thing is, he was talking to billion-pound corporates. But the diagnosis applies just as cleanly to a 20-person business in Stockport.
Maybe more so.
The dirty secret of the AI rollout
You’ve done it, haven’t you. Don’t lie.
Somewhere in your business right now there’s a Notion full of “AI experiments.” A Zapier flow somebody set up in March that nobody can remember the password for. A ChatGPT Team subscription with eleven seats and three active users. A meeting transcription tool. A “smart” CRM add-on. An invoice parser that works 70% of the time, which is the worst possible number.
You’re not alone. You’re the rule.
And the reason isn’t that you’re behind. It’s that the whole industry has been selling you the wrong shape of project.
Spare me the hype about portfolios of AI bets. That logic works for venture funds. It does not work for a business with 23 people, one finance lead, and a roadmap that already has too much on it.
What Krishna actually said
His sharper line, the one that should sting a bit, was this.
“If others are seeing success, why is your company so special that those use cases wouldn’t work?” The Deep View
Read that twice.
The honest answer for most SMEs is: we’re not that special. We just haven’t picked. We’ve been hedging because picking feels risky, and hedging feels prudent. It isn’t. Hedging across twelve half-baked AI tools is the most expensive thing you can do, because none of them ever reach the depth where they pay back.
Krishna’s prescription was almost insultingly simple. “Pick two or three areas where AI could scale massively, as opposed to doing 100 experiments.” The Deep View
Two or three. Not twelve. Not “an AI strategy.” Two or three places where you go deep enough that the thing actually changes how the work gets done.
The pilot trap
Here’s what doesn’t work.
Running a pilot. Nodding sagely at the demo. Sharing a Loom in the team WhatsApp. Promising to “look at it again next quarter.” Quietly never looking at it again.
Ray Wang from Constellation Research had a number for this one. The companies seeing 10x gains in speed and cost? They’re the ones who got past the pilot stage. The Deep View
Past it. Not in it. Not “evaluating it.” Past.
Most pilots die because nobody wanted them enough. They were ordered from the top, or sneaked in by an enthusiast in operations, or bolted on because a competitor mentioned theirs at a Chamber of Commerce lunch. There was no head of department waking up on a Tuesday genuinely wanting the thing to work.
That’s the test. And it’s brutal.
The leadership test (and why your existing rollout probably fails it)
Walk through your current AI experiments. Now ask, for each one, a single question.
Does the person who runs that function actually want this?
Not “agreed to try it.” Not “is happy to host the meeting.” Wants it. As in, has skin in the game, has told you they want it, has a metric they care about that this would move.
If the answer is no, you don’t have an AI project. You have a graveyard plot with a ribbon on it.
The reason this matters is that AI projects are not software projects. They’re change projects with software inside them. The 10/80/10 split that runs through every decent deployment, 10% human framing, 80% AI execution, 10% human refinement, only works when somebody on the human end gives a damn about both ends.
Without that owner, the 80% middle becomes wallpaper. Pretty, ignored, slowly peeling.
What “going deep” actually looks like for a UK SME
Let’s be honest about what this means in practice. You’re not IBM. You don’t have a Chief AI Officer. You probably don’t have a Chief anything except yourself.
Going deep, for a 30-person business, looks like this.
You pick one revenue process. Maybe quoting. Maybe lead qualification. Maybe the first 48 hours of customer onboarding, where the drop-off is killing you and you’ve known it for a year.
You give it to the person who runs that bit. You tell them: this is yours. You have a budget, a deadline, and a number you’re moving. Not “we’ll see what happens.” A number.
Then you stop them touching the other eleven AI things on the list. Because the entire problem is that they’re context-switching across twelve shallow tools instead of getting one thing into the bones of how the team works.
Depth means the AI is no longer a thing you “use.” It’s a thing the work runs through. There’s a difference, and the difference is where the 10x lives.
Why this is good news
Here’s the thing nobody says out loud at the AI conferences.
You’re not behind. You’re early. The corporates are mostly faking it. Krishna basically said so, on his own keynote stage. The sentiment in the C-suite right now is closer to “why isn’t this working yet” than “look at us go.”
Which means a small, focused UK SME that picks two bets and ships them properly will, by mid-2027, be ahead of most businesses ten times its size. Not because of cleverness. Because of nerve. The nerve to kill ten things so two can breathe.
At the end of the day, AI strategy for an SME is mostly subtraction.
The Practical Bit: the Two-Bet Rule
Here’s the action. Do it on Friday. Block 45 minutes.
Step one. Open a blank doc. List every AI tool, subscription, pilot, experiment, prompt library, GPT, automation, agent and “we should look into” running in your business right now. Be honest. Include the ones nobody touches.
Step two. Cross out everything except two. Maybe three if you can genuinely justify it. The criterion is not “which is most exciting.” It’s “which one, if it actually worked at depth, would meaningfully change my P&L or my customer experience.”
Step three. For each survivor, write down three things on a single line.
The named owner. A real person who wants it. Not you, unless you actually run that function day to day.
The success metric. One number. Hours saved per week, conversion rate, cost per ticket, quote turnaround, whatever.
The review date. Six weeks out. In your calendar. With the owner.
Step four. If you cannot name an owner who actually wants the project, kill it today. Cancel the subscription. Archive the Notion page. Tell the team. Don’t let it limp on draining attention.
That’s the whole exercise.
It will feel like you’re doing less. You are. That’s the point. Two bets going deep beat twelve bets going nowhere, every single time, in every business I’ve ever looked at.
Krishna’s giving you permission. Take it.
PS: If you want a second pair of eyes on which two to pick, reply to this email with your list. I’ll tell you which ones smell like real bets and which ones smell like ribbons on graveyard plots.