Before asking where AI belongs, we should ask something more fundamental: Is the process itself worth keeping?
Most enterprise processes were not designed for today’s world. They have evolved over years, shaped by legacy systems, organizational structures, compliance requirements, acquisitions, and countless workarounds. As a result, many processes are full of unnecessary approvals, duplicate checks, manual handoffs, spreadsheets, emails, and activities that exist simply because “that’s how we’ve always done it.”
Putting AI into such a process does not necessarily transform it. It may simply make a broken process faster.
Consider a typical enterprise workflow. A request comes in, one team reviews it, another validates it, someone approves it, another team updates a system, and eventually the customer receives an answer. Now add AI. The AI can read the request, summarize it, extract information, recommend an action, and perhaps update the system automatically. That is useful. But what if three of those steps were unnecessary in the first place? What if two teams were checking the same information? What if the approval existed only because of an old organizational policy? What if the biggest delay was not the work itself, but the time spent waiting between teams?
In that situation, automating individual tasks may improve productivity without meaningfully improving the business outcome.
The biggest opportunity may not be automating a step. It may be eliminating it.
Only after that should we decide where technology belongs.
And not everything needs AI. Some activities should simply be removed. Some can be handled through traditional rules and automation. Some require human judgment. Others are genuinely well suited to AI. AI becomes particularly valuable when work involves ambiguity, unstructured information, investigation, interpretation, or coordination reading complex documents, understanding customer requests, investigating exceptions, recommending actions, or orchestrating work across multiple systems.
If a task can be handled reliably by a simple rule, there may be no reason to introduce an AI agent. The objective is not to put AI everywhere. The objective is to design the best possible process and then use AI where it creates disproportionate value.
This is particularly important when we look at handoffs. One of the biggest sources of enterprise inefficiency is not necessarily the individual task; it is what happens between tasks. Work moves from person to person, department to department, system to system, and queue to queue. Every handoff creates an opportunity for delay, duplication, errors, and rework.
This is where AI agents could become particularly interesting. Instead of giving every department another AI assistant, imagine redesigning the process around the outcome. An agent could understand a request, gather information from multiple systems, perform routine work, resolve straightforward cases, and involve a human only when judgment is required.
The transformation is not simply that one employee now works 30% faster. The transformation is that the process itself has fewer steps, fewer handoffs, and fewer waiting points.
That is a much bigger prize.
There is also a lesson here about how organizations approach transformation. The process documented in a PowerPoint is rarely the process that actually happens. Ask the person doing the work every day. They know about the spreadsheet nobody officially owns, the manual workaround everyone relies on, the system field nobody trusts, and the approval that adds three days but almost no value.
Process data can show you what is happening. Employees can often tell you why. You need both.
The people closest to the process should therefore be involved in redesigning it—not brought in only after the technology has already been selected. They understand the exceptions, the dependencies, and the real-world constraints that rarely appear in a process diagram. More importantly, involving them creates ownership of the new way of working rather than resistance to another technology initiative.
Finally, enterprises need to rethink how they measure AI success. The number of agents deployed is not the outcome. The number of prompts processed is not the outcome. Even the percentage of tasks automated is not necessarily the outcome.
The real questions are much simpler: Did the process become faster? Did the cost decrease? Did errors decline? Did customers get better service? Did employees spend less time on repetitive work? Did the business outcome improve?
That is the real measure of enterprise AI.
The future will not necessarily belong to companies that deploy the most AI. It will belong to companies that redesign how work gets done and then use AI selectively to make those new processes dramatically better.
So before asking, “Where can we put AI?”, ask something harder:
“If we were designing this process today, would we design it this way at all?”
Start there.
Then decide where AI belongs.