"Which part do we agentize first?" is the wrong first question

The conversation about AI agents changed this year. A year ago everyone asked "are these real?" Now the question across boardrooms is "which part of my company gets agentized first?" — IT support, sales, payroll exceptions, customer service, document processing. It sounds like the smart, strategic question. It's the wrong first question, and asking it first is how good companies start bad agent programs.

"Which part" is a targeting question — and targeting isn't the hard part

Asking where to point the agent assumes the difficulty is choosing the target. It isn't. The hard part — the part that decides whether anything you build works — is what the agent stands on. Two questions come before "which process," and almost nobody asks them first:

1. What is your source of truth? An agent is only as good as the data underneath it. If two of your systems disagree about a customer's balance, an agent will confidently act on whichever wrong number it happened to read — and now the mistake is automated and fast. The industry's own honest finding is that most companies pour money into model choice and orchestration while treating their data as an afterthought, running on data that's fragmented, outdated, and impossible to govern at the speed AI demands. You cannot ground an agent on a swamp. If the facts the agent needs don't live in one trustworthy place, picking a juicy process to automate just means automating your mess.

2. Who is accountable? Before "which process," answer "when the agent acts here, whose name is on it?" If the honest answer is "nobody in particular," you're not ready to agentize anything — that's the org-chart problem, and no clever choice of target fixes it.

Why the wrong question is the popular one

"Which part first" is seductive precisely because it's the fun question. It sounds like opportunity. It demos well. It gets executive nods. The two real questions are boring, and worse, they expose that you might not be ready — your data is a mess, nobody wants to own the failures. So everyone asks the exciting question and quietly skips the uncomfortable ones. That is, almost exactly, why most agent pilots stall: not the model, the foundation. It's the production wall again, entered from the strategy side.

The order is the strategy

The fix isn't "pick a process, then bolt governance on afterward." It's the reverse. Establish the source of truth and the accountability layer first — one governed, current set of facts, observability so you can see what each agent did, clear ownership and scoped permissions — and then "which part first" becomes easy. Because once an agent stands on solid ground and someone owns it, almost any well-scoped process works, and several do. The order is the whole strategy. Leading enterprises figured this out and now build the governance and data foundation before the rollout, not after.

The test

Before you pick what to agentize, run any candidate process through two questions:

• Is there one place that's authoritative for the facts this agent needs — and do you trust it?
• Is there one person who owns what the agent does here?

Two clean answers, and the process is a fine first pick — and so are several others. Either answer murky, and the process was never the problem; your foundation is, and choosing a different target won't save you.

"Where do we start agentizing" is the question that makes you feel like you're moving. The questions that actually move you are quieter: what does the agent stand on, and who answers for it. Get those two right and the "which part first" answer is nearly arbitrary — take the highest-value process with a clean source of truth and a clear owner, and go. Get them wrong and it doesn't matter what you pick; you've just automated your confusion at machine speed. The first question was never "which part." It was "are we grounded, and are we accountable." Answer those, and the rest is the easy part.