Juniper Vale: I'm going to hand you a number before we even say hello properly — forty percent.
Mark Delaney: Uh — forty percent of what, exactly?
Juniper Vale: Gartner's forecast from June 25th: over forty percent of agentic AI projects — AI that autonomously sets goals, plans across multiple steps, executes actions without a human in the loop — canceled. By the end of 2027. Forbes called it an alarm bell.
Mark Delaney: Okay yeah, I saw that framing — like, uh, tech is in trouble, enterprises are bailing, the whole thing.
Juniper Vale: Right. And I think that framing is completely backwards. The fire alarm went off, sure — but the building was already burning before anyone installed a single agent.
Mark Delaney: Wait — so you're not worried about the forty percent. You're more worried about the sixty that make it through?
Juniper Vale: I'm saying Anushree Verma — Gartner Senior Director Analyst — put it plainly: these projects are early-stage experiments driven by hype, frequently misapplied. That's not a technology problem. That's enterprises throwing money at a concept they never actually defined. Cancellation isn't failure. Cancellation is, finally, a honest answer.
Mark Delaney: Hm — yeah, I mean, I can see that. But I kinda want to poke at it, because forty percent canceled still feels like a lot of wreckage.
Juniper Vale: Sure, wreckage — but the wreckage is actually harder to clean up than forty percent makes it sound, because enterprises can't even identify which projects were real to begin with.
Mark Delaney: Wait — what do you mean they can't identify them?
Juniper Vale: Gartner put an actual number on it — roughly 130 out of thousands of agentic AI providers are genuinely effective. Thousands of vendors, a hundred and thirty that actually do the thing. So the market is structurally counterfeit.
Mark Delaney: Okay that — uh, that number is kind of stunning. So when an enterprise goes shopping, they're mostly looking at... fake tools?
Juniper Vale: Think of it like this — you walk into a hardware store, ninety percent of the shelves are empty boxes with pictures of tools on them. You can't shop smarter if you literally cannot tell the box from the tool. That's agent washing. Vendors marketing conventional, minimally autonomous software as fully agentic. It's not a metaphor for confusion — it's a structural information problem. Buyers don't have the signal.
Mark Delaney: And so — wait, no, this actually makes the cancellation story worse, not better. Because Juniper said earlier, kill the bad projects. But how do you kill the bad ones when you can't even... I mean, you bought a picture of a hammer. You don't know that yet.
Juniper Vale: Exactly — and DEV Community ran analysis on July 7th of this year and argued eighty-three percent of AI agents are already defunct. That's already past Gartner's forty percent ceiling. The cleanup isn't coming. It started.
Mark Delaney: Eighty-three percent defunct — right, but that actually fits, kinda, if most of what got deployed was empty boxes. The casualty rate makes sense when the market was mostly noise from day one.
Juniper Vale: And that cleanup is what cracks the strategy argument open. Because Gartner — same report, same firm — also projected that by 2028, fifteen percent of day-to-day work decisions will involve agentic AI. That's not a contradiction. That's them saying: the bad deployments die, and the right ones quietly take over.
Mark Delaney: Wait — the same Gartner forecast? Forty percent canceled and fifteen percent ubiquitous by 2028?
Juniper Vale: Same firm. And that's actually the part that matters — it means the technology isn't broken. What's broken is the deployment story. There's a name for the mechanism: the capability-deployment verification gap. Pilots work beautifully in controlled environments. Then someone tries to scale it to production, and the governance infrastructure — the control plane, the oversight layer — just isn't there. The AI didn't fail. The org did.
Mark Delaney: So it's like — uh, okay, imagine a product designer who builds a perfect prototype in her studio in November, shows it to leadership, everyone's thrilled, then hands it to manufacturing in January and there's no quality control process, no supply chain, nothing. The prototype wasn't wrong. The rollout was.
Juniper Vale: That's almost exactly it. And the Claude Cowork data makes this concrete — more than ninety percent of agentic sessions on that platform are email, calendar, file management. Not autonomous decision-making. Not software development. That launched July 7th, same day as that DEV Community piece.
Mark Delaney: Hold on — ninety percent is just... organizing someone's inbox?
Juniper Vale: Ninety percent. Which means the agents that are actually sticking are bounded, augmentative — they assist, they don't autonomously run. And that's the kernel of truth in the whole cancellation story. Vendors sold autonomy. People are using assistance. The ones that survive figured out what the tool actually is — and we'll get to why the market still hasn't accepted that, which is where this gets genuinely uncomfortable.
Mark Delaney: Yeah, no, that makes sense. But — I mean, does that vindicate the cancellations, or does it indict the pitch?
Juniper Vale: Both. Look at Salesforce Agentforce — they published production numbers. Over fifty percent reduction in tail latency, up to three-point-nine times throughput improvement. But that required building out serverless and autoscaling infrastructure from scratch. That's not a model problem. That's an engineering operations problem. They had to construct the deployment discipline the model couldn't provide for itself.
Mark Delaney: So the projects that got canceled probably just... skipped that part.
Juniper Vale: Most of them skipped it, yeah — and there's actually an architectural reason why that's not fixable just by throwing a better model at it. Ashish Verma formally named it — the SIM Problem. Stateless Intelligence Memory. Agents can't maintain coherent context across sessions, across other agents, across time. Every new session, the thing basically woke up with amnesia.
Mark Delaney: Wait — that's not a training problem? Like you can't just... update the model and fix it?
Juniper Vale: That's the whole point of naming it. It's architectural. Not incremental. And that's actually why the bounded use cases survive — email, calendar management, file tasks — those don't require session memory. Each one is self-contained.
Mark Delaney: So the agents that work are the ones that, uh — I mean, they work *because* they're not trying to do the thing everyone was selling. That's almost funny.
Juniper Vale: And it gets harder — GLM-5.2, Chinese model, Marc Andreessen said publicly it matches top U.S. models. When a cost-competitive alternative shows up at that level, the ROI math on ambitious autonomous deployments just... collapses. You can't justify the infrastructure spend.
Mark Delaney: Andreessen said that — about a Chinese model? That's the kind of thing that makes the whole 'autonomous agents are inevitable' pitch feel way shakier than Anthropic, OpenAI, and Google DeepMind would ever admit.
Juniper Vale: They're all still shipping agentic products, all three of them, even as this cancellation wave builds. The trajectory is real. But the use cases that actually hold — narrower than any of them are claiming.
Mark Delaney: So the verdict is — the market didn't fail because the technology broke. It failed because 'augmentative' was always the product, and nobody wanted to sell that.
Juniper Vale: Okay, so maybe I was a little too happy about the cancellations. I kept framing the forty percent as healthy correction — and it is — but I glossed over the part where the Gartner numbers are actually sitting next to each other. Forty percent gone by 2027. Fifteen percent of daily decisions handled by agents by 2028. Those two things aren't in tension. They're a roadmap. The projects that die were built to be autonomous. The fifteen percent that lands? Built to be useful. I think I was so focused on the cancellation being honest that I undersold what the survival rate actually tells you.
Mark Delaney: Yeah — no, I think that's the cleaner way to put it. The market's question was never whether agentic AI works. It's whether anyone will just... admit what it's actually for. And uh, based on what Claude Cowork's usage data says — ninety percent inbox and calendar, not autonomous decision-making — someone already answered that. The users did. Vendors just haven't caught up to their own product yet.
Juniper Vale: That's the sting in the tail, isn't it. The users figured it out before the pitch deck did. I'll take that as where we land.
Mark Delaney: Good talk. Genuinely.
