There's a version of enterprise AI that's going to fail in a very predictable way. Not because the technology doesn't work. Because it's being deployed against the wrong unit of analysis.

Most AI vendors - and most enterprise buyers - are organizing AI products and investments around functions. Customer support gets an AI agent. HR gets one. Finance gets one. It looks like progress. It's actually a more sophisticated version of the fragmentation problem enterprises already have.

Wonderful's thesis has always been different. We don't bet on specific models or modalities. Our job is to take whatever the frontier makes possible and operationalize it inside enterprises, faster than anyone else. 

When we launched, the most promising modality was voice. That made customer support the natural starting point: real-time, high-volume, unstructured, and high-stakes. If you can deploy reliably there, you've built something that works. As other modalities emerged - computer use, chat, document processing - we applied the same approach, and we’ll do that for whatever comes next. 

What stayed constant across all of it was the same structural problem. Enterprise AI keeps getting deployed against the wrong thing: the function, not the workflow. 

Here's what production taught us.

Resolving a customer issue almost never stayed inside the support function. A customer calls about a delayed shipment - the agent needs to check logistics, access inventory, pull compensation policy, trigger a fulfillment action, update the CRM. What looks like a support interaction is actually a cross-functional workflow that happens to start with the customer. 

Even as we expanded to additional agentic solutions with employee copilot agents and backoffice agents, we saw the same pattern everywhere we deployed. An employee submits an equipment request - that's HR, IT, and a manager approval chain. An insurance claim needs to be reviewed and approved - that’s support, compliance, operations, finance and risk. The entry point is always a function, but the work always crosses several.

The insight isn't that customer support shouldn’t be solved with AI. It's that customer support, or any function-specific AI, as a category, is the wrong frame.

Enterprises are organized into functions. Work is not.

Traditional software was designed around organizational structure, because software primarily stores information and enables human coordination. The org chart is a reasonable enough boundary for a CRM or an HR system like Salesforce or Workday. 

But AI doesn't store information and coordinate humans. AI performs work. And work is a workflow - a chain of actions, decisions, and system interactions that starts somewhere and ends somewhere else, crossing whatever functional lines it needs to along the way.

When you deploy function-specific AI, you're not solving that. You're adding an intelligent layer on top of an already fragmented operating model. AI for support. AI for sales. AI for HR. AI for finance. The enterprise that buys this way isn't transforming how it operates - it's automating its existing fragmentation. The companies building function-specific AI aren't wrong about the technology. They're wrong about the unit of analysis. Automating specific functions drives real value - department leaders get better numbers. But the organization as a whole doesn't change.

What enterprises actually need is an AI layer that operates across workflows, not within functions. One system that can pick up a process wherever it starts, access the systems and data that process requires, and carry it through to completion regardless of how many departments it touches.

This also changes the buyer. Function-specific AI is a departmental decision - the head of support buys a support tool. Workflow AI is an operating model decision. It has to be sponsored at the CEO level, because the value only compounds when AI is applied across the organization, not optimizing one function at a time.

It also changes what deployment actually looks like. You can't buy workflow AI off a shelf and install it. The work of deployment is understanding how processes actually run. It requires engineers - whether they are in-house or deployed from a partner and embedded - physically present in the environment where work happens, building the integrations, mapping the edge cases, and iterating on the logic until the agent is actually doing the work end to end. 

What production in customer support taught us is that solving it correctly requires infrastructure that doesn't stop at the function. The integrations have to reach across systems. The orchestration layer has to span departments. Build that, and you haven't built a support tool. You've built the foundation for an enterprise-wide AI operating system - proven in one of the hardest, highest-volume environments in the business.