Only 23% of enterprises report meaningful AI ROI from their agents. That's a predicament.

Nearly four in five say they're struggling despite spending serious money. And 69% are planning layoffs they're pinning on AI - while 39% of those same companies admit they have no actual strategy to make money from the tools in the first place.

The easy take is "AI is overhyped, told you so." That's not my take, and I reckon it's lazy.

The tech mostly works. That's exactly what makes this interesting.

The thing that's actually broken

The failure here isn't the model. The model is fine. Gartner reckons over 40% of agentic AI projects will be scrapped by the end of 2027 - and the reasons listed are escalating costs, unclear business value, and inadequate risk controls. Notice what's not on that list: "the AI couldn't do it."

The capability of AI is evident. The financial results for people using it at scale? Not so much. There's a gap between those two things, and the gap has a name.

It's the Knowing-Doing Gap - Pfeffer and Sutton wrote the book on it twenty-odd years ago, long before any of this. Their finding was brutal and it has aged perfectly: giving people better information does not change what they do. Organisations are full of smart people who know the right answer and don't act on it, not because they lack data, but because the system around them isn't built to turn knowing into doing.

Swap "information" for "AI" and you've got 2026 in a sentence. The Knowing-Doing Gap is alive and well.

The survey data says it plainly. Companies have super-users pulling genuinely extraordinary results out of these tools. What they don't have is any mechanism to spread that across the business, or anything connecting those individual wins to an actual outcome. The gains are real and they're trapped. One person doing something brilliant in a corner is not transformation. It's an anecdote (and often - in isolation - a frikkin' cool anecdote too!).

Buying twelve agents doesn't make it a transformation. The average company now runs about twelve AI agents, heading for twenty by next year - and half of them operate completely alone, wired into nothing. That's not a system. That's twelve shmoes hoarding their own piles of cooler-than-cool AI-goodness.

Agent washing

My favourite stat in all of this: of the thousands of vendors out there selling you "agentic" everything, Gartner estimates only around 130 offer anything resembling genuine autonomous capability. They've got a name for the rest of it - agent washing. The enterprise cousin of greenwashing. Slap "agentic AI" on the box, charge accordingly, ship a chatbot with delusions of grandeur.

A tool is a tool. Claude, an agent, a model, whatever's next quarter's hotness - it's a thing that can do work. But a tool with no system behind it isn't scalable and is unlikely to move the needle as much... It demos beautifully, and it may even bring you some revenue M, or maybe A - in my MAD Rules of Reinvestment. But it ain't no D! You've bought the appearance of transformation and skipped the part where the business actually has to change shape around it.

I can say this because it's how we run things here, and it's the opposite of how most of it gets sold.

What we actually do with it

At The Perception Collective we use AI a lot. Claude Code is doing real work in our build pipeline - not "look, it wrote a function", but properly inside how we ship. I'm not romantic about it. To me the AI is a tool. A very good one. But it's a tool.

The thing that makes it worth anything is the system it sits inside. The process that decides what the tool touches, where its output goes, who checks it, what it connects to on either side. We didn't buy a clever toy and wait for magic. We changed how the work flows, then put the tool where it could actually compound. And a hoo-man did that!

That's the unglamorous part nobody's selling, because you can't put it in a launch tweet. And it's unlikely to be the top talking point from Anthropic or OpenAI make as they seek to go public this year, amidst exponentially increasing costs. There's no system in a box. The vendor ships you the agent. The system is yours to build, and if you don't build it, you've bought a very expensive way to generate output that goes nowhere. Or is insecure. Or incomplete. Or not ready for primetime.

A tool produces. A system compounds. Those are not the same word.

The meter is now running

But things are about to get mighty tricky, turning your cool anecdote into a money pit, fast.

For two years the AI you've been using was subsidised. Venture money was paying for your tokens so the vendors could grab the market. That era is closing. By this year, 85% of SaaS providers had shifted to hybrid or consumption-based pricing - billed on what you actually use. The free lunch is over, and a lot of people are about to find out what they were really eating.

Consumption pricing is a different animal to a seat licence, and worse in the one way that matters: it doesn't scale in a straight line. Tokens multiply geometrically in agentic workflows - an agent that reasons, calls tools, loops back, calls again, burns many times what a single question costs. EY put real numbers on it: a customer-service interaction that ran $0.04 in 2023 costs about $1.20 in 2026 once you wrap it in tools and reasoning loops. Thirty times more. Total token consumption across the market is up thirteen-fold since the start of last year.

You want it concrete? When GitHub flipped Copilot to usage-based billing on the first of this month, one developer watched their projected monthly cost go from about €67 to around €966. Same work, just with Uber Surge Pricing permanently.

Let's layer these two problems together.

If your AI sits inside a real system - measured, connected, pointed at an outcome - rising per-token cost is a line item you manage. You can see what it's buying. You can route cheap work to cheap models and save the expensive ones for where they earn it. You can also more likely measure its direct return. Not via cool anecdote, but via your bottom line.

If your AI isn't in a system i.e. instead of being genuinely integrated into your company processes, it's a big fish/small pond scenario for each individual person at your org, then consumption pricing is about to take a massive chunk outta your ass.

That's the trade that's about to go badly for a lot of people. Not because the AI failed. But because it's something that is very cool, it can do lots of stuff, but it also means it's easy to get distracted and not have it be a focused use in your org.

A small mercy, briefly

Spare a thought for the CFOs who got a stay of execution this fortnight.

The most expensive tier going - Anthropic's Fable, priced at $50 per million tokens, double what Opus runs - got abruptly pulled on the 12th of June for reasons that have nothing to do with any of this. Nobody can spin it up right now, and CFOs around the globe rejoiced: the bill that was coming for anyone who'd wired their whole operation to the priciest model on the shelf has been deferred to next quarter ;).

Enjoy it while it lasts. It's a reprieve, not a pardon. The meter's only switched off because the model's in time-out.

This Week's One Thing

Pick one AI tool you're already paying for.

Now answer two questions about it, out loud, no hand-waving:

  1. What system does its output feed into - where does the work actually go after the tool's done?
  2. What would it cost you per month if the meter doubled tomorrow? Not in your subscription fee, but your usage. And if you don't know the difference between the two, you're either small enough that it's ok, or larger, but already screwed!

If you can't answer the first one, you don't have a system - you've got a tool and a vibe, and consumption pricing is going to get chunky like consumption pricing does (hey, don't hate the player).

If you can't answer the second question, you're exposed.

I'm not saying go back to using an abacus. But, you've got to have your eyes wide open about what this increased cost is producing for your business.

Annnnnd, GO!

#BeAVillager

If you're interested in reading a little bit more about the Knowing-Doing Gap, check out an earlier piece here

References