Skip to content
Fast Company Executive Board

The Fast Company Executive Board is a private, fee-based network of influential leaders, experts, executives, and entrepreneurs who share their insights with our audience.

Why most AI enterprise pilots fail

Enterprise AI doesn't fail because of bad models—it fails when AI sits beside the workflow instead of inside the decision it's meant to support.

[Photo: Getty Images]

Profile photo of Dominik Facher

Chief Product Officer · ZoomInfo

AI demos have created a bad habit. The demo runs on perfect conditions—clean data, clear permissions, no edge cases—so of course it dazzles the room, and a pilot gets funded.

In production, however, those pristine conditions disappear. Real enterprises are hamstrung by legacy systems and messy data, making deep integration difficult. To bypass that complexity, vendors default to putting the AI in a standalone chat window on the side of the screen.

While the reflex is to blame the model, the model is rarely what failed. What failed is that the AI was relegated to an external sidebar assistant—forcing users to stop their real work to prompt it—rather than being embedded directly inside the active digital environment where decisions are made.

That distinction decides everything, because enterprise buyers do not buy AI in the abstract. They buy time back, a way to prioritize, a faster path to the next action, and they feel none of it until the AI sits inside a live decision loop—like automatically flagging a churning customer or drafting a specific follow-up email right on the execution screen. The closer a product gets to that decision, the faster real adoption follows; the further it sits from it, the faster it becomes one more thing to check.

A chat window is not a workflow

The reality is that the industry fell in love with the chat window. But a chat window is merely an input method—a way to ask, not a workflow. Every time it forces a user to pause, open a box, and draft a prompt, it takes time away from their day. That is the assistant layer, and it is exactly why these tools get demoed, admired, and never opened again: they sit beside the work instead of doing the work.

What people want is the opposite: something they don’t have to supervise, that runs in the background, holds the context, and acts on its own. That is the workflow layer, the move from a copilot to an autopilot, and it is where adoption happens. It is also the harder thing to build, because no one lets a system act unwatched until it has earned that trust by proving that it can be left alone.

Build to solve, not theorize

However, none of this is theory for me. My team and I built an AI-powered sales pipeline engine based on an industry problem: go-to-market teams are rich in data and starved for orchestration. Sales reps spend their day jumping between a dozen tools and stitching context together just to make a single call, so we built something to do the stitching, surface what matters first, and suggest the next move. We built it to target exactly what a buyer is paying for: time back, prioritization, and a shorter path to action.

What happened next was the proof. It caught on faster than we expected, and from people we hadn’t built it for: account executives and account managers pulled it into whole deals, because the roles had merged and the work no longer matched the old titles. No one made them use it; they asked for it, some before we were even ready to give them a seat. That’s the kind of adoption no mandate can manufacture; it only happens when a tool lives inside a live decision and makes the next step faster. It worked so well, in fact, that our problem inverted: now the struggle was moving people off that first workflow and onto the next.

Data management is the boring work

Reaching that layer is the hard part, and the reason is always data. Our early customers had tidy, well-organized data, so the product worked well. The deeper we pushed into the enterprise, the more the underlying data broke down because legacy enterprise data is heavily siloed—until one customer had nine separate CRMs with the same account in all of them, 10 times over. Pull the wrong context and the rep stops trusting the answer, then stops opening the tool, and that is what killed early deployments. It wasn’t the model. The work that prevents failure is unglamorous: validation, ontology, matching, lineage, the boring 90%, and it is where the real moat sits.

One number made that permanent for me. Less than 1% of the useful context on any account lives in the CRM fields everyone builds on; nearly half lives in customer interactions themselves—specifically, recorded conversational sales calls and meetings—and the rest in real-time buying signals, executive hires, and market news that no one had connected. That means the order of the work has nothing to do with AI: capture the right data, get it into one place you trust, and only then run a workflow on top of it. Skip the first two, and the workflow you were trying to reach fails so quietly you barely notice.

AI gets adopted in the workflow layer

In the end, what matters is not how capable the AI is; it is whether AI lives in the workflow or beside it. Capability gets a product into the room, but living inside the decision is what keeps it there. Proving it can be trusted there is the price of admission.

That is also why my team and I at ZoomInfo built our underlying platform architecture, GTM.AI, the way we did: an open protocol that an AI agent can call with the groundwork already in place, so the heavy work is resolved before the agent ever has to ask.

The most hopeful part is also the most human. AI is making everyone a builder; a single rep can now create what once took an engineering team and a long backlog wait, as long as the grassroots building stays in everyone’s hands while centralized enterprise control and security guardrails stay in one place.

Two years from now, the enterprise AI still running won’t be the demos that dazzled a room. It will be the products that moved into the workflow, sat down inside the decision, and earned the right to stay.

Dominik Facher is chief product officer at ZoomInfo.