2. The unit of change is a process, not a department

The most common way an AI initiative dies is being aimed at something too big to finish.

“Make sales AI-native” is not a project. It is a slogan with a budget attached. It has no owner who can approve the whole thing, no single measurement that says whether it worked, and no end. Twelve months later there is a steering committee, a vendor, a pilot in three regions, and no process anyone can point at that runs differently than it did before.

The unit that actually works is one process, named narrowly enough that you can describe its input and its output in a sentence.

Not “customer support.” Instead: “triaging inbound support email into the right queue with a first response.” Not “recruiting.” Instead: “screening inbound applications against the role requirements and scheduling the first call.”

Why this size

A process at this granularity has four properties that make it finishable.

It has one owner. Somebody can say yes to changing it without convening anybody. If you cannot name that person in one try, your scope is too wide.

It has a countable unit. Tickets, applications, invoices, orders. If you cannot count the things flowing through, you cannot measure cost per unit, and you will not be able to prove the redesign worked.

It has an observable failure. You know what a bad outcome looks like and roughly how often it happens today. Without this you have no baseline, and every comparison later becomes an argument about vibes.

It fits in a quarter. Long enough to build something real, short enough that the sponsor is still in the same job when it lands.

The scoping conversation

When we assess, we push until the process is this small. The useful questions, in order:

  1. What comes in, and what goes out? Name the artifact on each side.
  2. Who owns it? One name.
  3. How many units per week or month?
  4. What does it cost today, in hours or dollars, and how confident are you in that number?
  5. Where does it wait? Not where is it hard, where does it sit doing nothing.
  6. What does a bad outcome look like, and what happens when one occurs?

Question five earns its place. People describe the hard parts, because those are what they think about. But the time is almost never in the hard parts. It is in queues: work sitting in an inbox until someone gets to it. Queues are where redesign pays, and they are invisible to the people inside the process, because waiting does not feel like work.

After the first one

Pick the second process because the first one taught you something, not because it was next on a list. In practice the second is usually adjacent: the step immediately upstream or downstream, where you now control both sides of a handoff. That is where the compounding starts, and it is the honest reason to do one at a time rather than five in parallel.