The bottleneck was never intelligence

I run marketing for an enterprise AI company. Every week I talk to leaders who are genuinely excited about what AI can do. And yet an MIT report says 95% of organizations are getting zero return from their AI investments. Not low returns. Zero. That number should bother all of us more than it does.

Tools are being deployed. Pilots are running. Boards are asking about it in every earnings call.

And yet the work isn't changing. Not in the ways that show up in margins or cycle times or customer outcomes.

That gap is worth taking seriously. Because it's not a coincidence, and it's not a capability problem. It's a structural one.

The bottleneck was never intelligence. It was always orchestration.

What AI actually walked into

AI arrived inside enterprises that spent decades papering over their own complexity. Every large organization has the same archaeology: systems from mergers that never got integrated, approval chains that grew longer every time something broke, and processes that work most of the time and require heroic effort the rest of the time. Nobody designed it this way. It accumulated.

An AI agent can reason beautifully within its slice. What it can't do is hand off to the next system, wait three days for a human approval without losing context, or know that the flagged exception needs to go to legal before it goes to the manager.

That connective tissue—the orchestration—was never built. And intelligence without orchestration doesn't produce outcomes. It produces impressive demos and zero change to your profit and loss (P&L).

Two kinds of work

This is where I think the conversation about AI gets muddled, and it matters enough to spend a moment on.

Not all work takes the same shape. Some work follows a fixed path: an invoice gets matched to a purchase order (PO), routes for approval, posts to the system. Each transaction looks roughly the same. The value is in executing it precisely, swiftly, and consistently.

Other work is fundamentally different. A compliance investigation. A disputed insurance claim. An onboarding where the customer is missing three documents, a system is throwing errors, and someone needs to decide whether legal gets involved before anything moves forward. This work doesn't follow a map. The right person needs to be in the loop at the right moment because the situation demands it.

Most organizations treat these as variations of the same problem. They're not.

Forcing dynamic, judgment-heavy work into a rigid process flow is exactly how you end up with the chaos diagram that appears in every AI pitch deck—the tangled web of exceptions and reroutes that nobody designed and few understand.

The question isn't "how do we add AI to what we already do?" It's recognizing that these two categories of work require fundamentally different architectures, and that intelligence has to be designed into both (not dropped on top of either).

What a decade in the enterprise actually teaches you

I was talking with a customer in healthcare recently. Their team reduced medical record review time from 70 minutes to six. Ninety percent. That result shows up in an exam room, where a clinician actually has time for the patient in front of them.

But getting there required knowing things that don't come from reading about enterprise AI. It required understanding that healthcare workflows have compliance checkpoints that aren't optional. That certain decisions need a person to sign-off (not because the AI agent can't reason but because the regulator requires a person to be accountable) and that governance isn't a layer you add at the end. Governance is the thing you design around from the start.

That knowledge comes from being inside these processes for a long time. Across financial services, manufacturing, public sector, healthcare, etc. In ‌places where the stakes are high enough that getting it wrong has real consequences.

AI cannot convey trust. It doesn't read the risk profile of a room or navigate the organizational politics of an approval chain. Only a person, at the right moment in a process, can do that. The job is making sure that person is never buried under work that machines should have handled.

That's a very different problem than building a capable model.

The architecture question nobody is asking loudly enough

Every quarter, I talk to executives who are frustrated by the same thing. The pilot worked. The rollout stalled. The P&L stayed flat. When I dig into why, the answer is almost always the same: there's no coordination layer. AI agents making decisions that systems downstream can't act on. Outputs with no clear owner. People filling gaps nobody planned for and doing it differently every time.

That's not an AI problem. It's an architecture problem. And it won't be solved by a better model.

UiPath has been working on this problem for over a decade. Not as a reaction to the AI moment. The question of how work actually moves across systems, people, and decisions was always the harder one. UiPath Maestro™ is what that looks like as a product: one layer where agents, robots, and people share context, follow real rules, and don't fall apart when things get complicated.

The patchwork era is ending.

The problem was never intelligence; it was orchestration. And the organizations that recognize that now, before the gap between pilots and performance becomes permanent, are the ones that will define what enterprise AI looks like at scale. UiPath is built for exactly this moment.

The infrastructure exists. Now it comes down to whether ‌leaders who built their careers managing complexity are willing to replace it with something that actually works.

Source: MIT Sloan Management Review, The GenAI Divide: State of AI in Business 2025, mlq.ai, 2025.