Insight

Four reasons your AI initiative looks busy but ships nothing

In the first post in this series, we named a pattern most of the market is currently living inside: AI Theater. Budgets approved, pilots announced, steering committees formed, and almost nothing actually running in production. We called it a symptom of something deeper, Non-Adoption, and promised to open up the mechanism behind it.

Before we do, one honest note. You’ve probably seen the headline by now: “95% of AI pilots fail.” It’s the most repeated statistic in enterprise AI, it’s a year old at this point, and it has appeared in so many decks that it’s practically wallpaper. So instead of leaning on it again, we pulled the numbers that have come out since. The short version: a year further in, the picture hasn’t improved. In some places it has gotten sharper.

These are the four reasons below, and they show up in almost the same order in almost every company we’ve seen so far.

Reason one: The Platform Trap

Here’s a scene that repeats itself with almost comic consistency. A company decides it’s serious about AI. The first move isn’t a use case: it’s an architecture review. “First we need the right stack, the right models, the right governance layer, the right data architecture.” Somebody draws a diagram with six layers on it. A vendor gets flown in to present a roadmap.

Six months pass.

The infrastructure is, by most accounts, quite good. Zero use cases are in production, and the CEO is starting to ask a very reasonable question: what did the money actually buy?

This isn’t a rare misstep, and it isn’t unique to AI, though AI makes it worse. RAND Corporation’s analysis of more than 2,400 enterprise AI initiatives found that over 80% fail to deliver their intended business value, roughly double the failure rate of ordinary IT projects.(1) When those failed initiatives get tallied up individually, the number is sobering: the average abandoned large enterprise AI initiative carries a sunk cost of around $7.2 million.(2) That is not a pilot budget quietly written off. That is a platform built and shelved before it ever produced anything.

It’s the equivalent of building a factory and trying to spin up the assembly lines before anyone has designed the thing the factory is supposed to manufacture. The instinct to get the foundation right first feels responsible. It is, in fact, the opposite of how working systems get built. A company doesn’t need a massive AI platform. It needs a shipping pattern. The platform, if it emerges at all, should emerge from working use cases, not the other way around. We’ve never seen this process run successfully top down.

Reason two: Fragmented architecture

Ask three departments at almost any company how many AI tools they’re running, and you’ll get three different, confident, incomplete answers. Sales wants a deal-desk assistant. Support wants a ticket summarizer. Engineering wants a code reviewer. Marketing wants, well, exactly the kind of blog post you’re reading right now. Each request is entirely reasonable on its own. Together, they add up to something nobody actually manages.

The scale of this is larger than most leadership teams realize. Recent research puts the average enterprise at 23 distinct AI tools in active use, and only around a third of companies maintain anything close to a complete inventory of what’s actually running.(3) A separate survey of more than 500 U.S. C-suite executives found more than a quarter of enterprises now use upwards of ten different AI applications, and most admit they haven’t moved past basic, ad hoc integration between any of them.(4)

Picture ten understudies rehearsing ten different plays in ten different rehearsal rooms, with no opening night scheduled for any of them. Everyone is busy. Everyone is technically “in progress.” Nothing is finished, because nothing was ever coordinated to finish. This is why a company can have a genuinely large AI budget and, a quite some time later, still not be able to name a single thing that shipped and mattered. The money wasn’t wasted on any individual decision. It was wasted on the absence of a decision about how the pieces would fit together.

Reason three: Features versus Systems

This is the reason that does the most damage, because from a distance it looks like the exact opposite of a problem. The company has Copilot licenses. It has a chatbot on the intranet. It has an agent bolted onto the CRM. Activity is visible. Something new got added to the toolbar. And surely that counts as super progress.

The newest data on agents specifically is the clearest signal yet that it doesn’t. Forrester and Anaconda’s 2026 research found that 88% of AI agent pilots never make it to production, and the failures are rarely about the model. Most come down to unclear success criteria and governance gaps that were never resolved before the project started.(5)

The same gap shows up wherever people are actually asked to use what got bought. Independent research on Microsoft 365 Copilot puts real weekly usage at roughly 36% of licensed seats, meaning close to two out of every three paid seats sit largely idle. At $30 a seat a month, a ten-thousand-seat rollout is quietly burning well over a million dollars a year on software that’s mostly there for the toolbar, not the workflow.(6)

A feature is something you bolt on. A system is something you build into how the work actually flows through the company, tuned to your own data and your own processes, compounding a little more every quarter instead of sitting there static and slowly going stale. Features are rented. Systems are owned. Most companies right now are paying rent on capability that compounds for the vendor’s platform, not for the company’s own operations.

Reason four: The performance becomes self-reinforcing

Put the first three together and something worse than the sum of their parts shows up: the theater starts protecting itself. Ask almost any executive right now whether their company is benefiting from AI, and the answer is yes, overwhelmingly, 97% say so. Ask a sharper question, whether that’s actually shown up as real, measurable ROI, and the number collapses to 29%.(7) That’s not a rounding error. That’s most leadership teams answering yes and no to the same question in the same breath, without seeming to notice the contradiction.

And the theater isn’t holding steady. It’s accelerating. In a single year, the share of companies that quietly walked away from most of their AI initiatives jumped from roughly 1 in 6 to more than 4 in 10, abandoning nearly half of their AI proofs of concept before any of them ever reached production.(8)

A steering committee exists to report progress, so it reports on activity, because activity is the thing that’s visible and countable. Budgets get renewed because the program looks active enough. Nobody wants to be the person who says out loud, in the all-hands, that a year of work produced a demo three people use and a stack of unused licenses. So nobody does. The show goes on because ending the show would mean admitting the show was, in fact, the entire output.

This is worth sitting with, because it changes what the fix actually looks like. None of these four patterns come from a lack of talent, and they rarely come from bad intentions either. They come from decisions that made complete sense in the room where they were made. Wanting solid infrastructure before you build on it is the responsible instinct. Letting every department chase its own pilot looks like empowerment. Shipping a visible feature feels like progress, because you can actually see it. But.. when someone finally asks what shipped, that the pattern becomes visible, and by then it’s quietly woven into how the whole organization works.

That’s actually the useful part. If the problem is structural rather than personal, it can be redesigned. That’s where the real work starts. And that’s where we can help you.

Naming your own version

If any of these four sound familiar, that’s the point. Most mid-market companies are running some blend of all four at once, which is exactly why the theater feels so hard to exit from the inside. You can’t fix a mechanism you haven’t looked at directly.

The genuinely useful exercise here is diagnosis. Which of these four is actually running the show at your company right now?

Is it a platform that’s still “almost ready” eighteen months in?

A dozen uncoordinated pilots that never got a shared owner? A shelf of licenses nobody measures?

Or a reporting culture that rewards activity over outcomes? Most organizations will recognize themselves in at least two.

In the next and final post in this series, we’ll walk through what actually gets a company out, not with a bigger platform or a longer roadmap, but with a specific, sequenced way of shipping one real thing, proving it moved a number, and then shipping the next.

If you already know which of these four is yours, that’s not a bad place to start a conversation. Talk to Nebul.

Haven’t read Part 1 yet? It’s here: Busy, Performative, and Going Nowhere: The Anatomy of AI Theater.

Sources

(1) RAND Corporation, analysis of 2,400+ enterprise AI initiatives, as reported in industry coverage of AI project failure rates, 2026.

(2) Folio3 AI, “AI Project Failure Rate in 2026: What the Data Shows,” April 2026.

(3) Larridin, State of Enterprise AI 2026 research.

(4) Zapier / Centiment survey of 550 U.S. C-suite executives at companies with 1,000+ employees, October 2025.

(5) Forrester and Anaconda, 2026 enterprise AI agent research, as reported in “AI Agent Adoption 2026: 120+ Enterprise Data Points,” Digital Applied, April 2026.

(6) Recon Analytics, U.S. AI Survey, January 2026.

(7) Writer, 2026 Enterprise AI Adoption Survey (1,200 C-suite executives, 1,200 employees).

(8) S&P Global Market Intelligence, Voice of the Enterprise survey, 2025-2026.

Nebul is proud to be named a 2025 Gartner® Cool Vendor in AI specialty cloud providers

Can Big Tech be trusted with sovereign cloud? (FD article insights)