Don't Automate Out the Governance: What a COO and Chief of Staff Keep Human When AI Takes Over a Process
In an AI rollout, one of the most important calls a COO or Chief of Staff makes is which steps stay human. You don't automate out the governance in a process. You automate to keep it.
I build automations. I like them. A few weeks ago I wrote about a Slack workflow I built that takes contractor hours, builds the invoice, routes it for review, and sends it to accounting with one click. It saves real time every month.
It also has three approval points in it, on purpose. I could have removed every one of them. The tool would have let me. I didn't, and the reason is the whole point of this article.
Governance is not the slow part
Every process that moves money, changes a record about a person, or commits the company to something has steps that look like friction. Someone checks the numbers. Someone compares this month to last month. Someone signs off before anything is released. Someone is named as the person who answers for it when it's wrong.
When a team sets out to automate a process, those steps are the first ones that look like waste. They're manual, they wait on a human, and they rarely find anything. So the temptation is to automate straight through them.
But those steps aren't the inefficiency in the process. They're the governance. They're how the company knows the output is right, and how it knows who to ask when it isn't.
When you automate end to end, you don't remove the approval. You move it to nobody.
What the human steps protect
It's easy to say "governance" and lose people, so here's what those steps are actually doing for you in any money or people process.
Accountability. When something goes wrong, someone has to own it. "The system did it" is not an answer an employee, a vendor, or an auditor accepts. A named approver means a named owner.
Separation of duties. The basic control in any finance process is that the person who prepares something isn't the person who approves and releases it. When one automated flow prepares, approves, and releases, that separation collapses into a single point of failure.
Error containment. A person who makes a mistake usually makes it once. An automated process that's wrong is wrong every time it runs, quietly, at scale. The human check is often the place an error gets caught before it reaches someone's bank account.
Judgment on exceptions. Rules handle the normal case well. Processes break on the cases nobody wrote a rule for: the unusual adjustment, the change that came in late, the situation that's technically allowed and obviously wrong. Those need a person who understands what the rule was for.
The audit trail. Governance isn't only the check. It's the record that the check happened. "Approved by a named person on this date" is evidence. "Ran automatically" is not.
Three kinds of steps
On Monday I wrote about why someone has to learn the workflow before AI touches it. Once you know the workflow, the next question is what to hand over. I sort every step in a process into one of three kinds.
1. Assembly steps. Automate them fully.
Gathering the data, pulling numbers from different systems, calculating, formatting, reconciling, routing to the next person. This is where AI and automation are genuinely excellent, and where people have been spending hours for years. Hand all of it over.
2. Check steps. Automate the detection, keep a person on the result.
Comparing against last period, catching duplicates, spotting anything outside the normal range. Let the tool do the comparing, and do it more thoroughly than a tired person at 4:00 on a Thursday ever could. The flags go to a human, and a human decides what they mean.
3. Decision steps. Keep them human, and make them better.
Approving, releasing money, signing off, deciding an exception, anything a person would have to explain later. These stay with a named owner. The automation's job here is to make that person faster and better informed: a clean summary, the flags already surfaced, everything they need on one screen. The approval should take five minutes instead of an afternoon. It should not take zero minutes because it no longer exists.
That's what I mean by automating to keep the governance. The assembly gets cheap, so the people involved spend their time on the part that actually requires them.
Five questions a COO asks before automating anything end to end
If you're looking at a process (vendor payments, customer refunds, anything where money or people are on the other end), ask these first.
1. Where does money move, a commitment get made, or a record about a person change?
Those are your decision steps. Each one keeps a human sign-off.
2. Who would put their name on this output?
Name the owner first. That person is who the process runs for.
3. When this is wrong, who finds out, and when?
The goal is a person catching it at a check, before it reaches anyone it affects.
4. Is the person who prepares it still different from the person who approves it?
If the automation does both, put a person back on one side.
5. Could you show an auditor who approved this, and when?
If yes, your governance is still in the process, even if most of it now runs on its own.
Side by side
| Automating out the governance | Automating to keep it | |
|---|---|---|
| Goal | Remove every human step | Remove the busywork around the human steps |
| Approvals | Gone, or rubber-stamped by the system | Kept, faster, with better information |
| Accountability | "The system did it" | A named owner for every decision |
| Separation of duties | One flow prepares, approves and releases | Preparing and approving stay separate |
| Errors | Repeat silently at scale | Flagged and caught before they land |
| Exceptions | Forced through the rules | Routed to someone with judgment |
| Audit trail | "Ran automatically" | Who approved what, and when |
| How it fails | Quietly, then all at once | Visibly, at a check, while it's still small |
The real efficiency gain
This is an argument for automating, and for automating the right part.
The hours that matter in most processes aren't spent on the approval. They're spent on everything before it: chasing data, fixing formats, reconciling spreadsheets, finding out what changed. Automate that and the process gets dramatically faster, while the five minutes where a person looks at it and says "yes, this is right" stays exactly where it is.
That five minutes is the cheapest insurance in the company. It's where trust in the process comes from. And it's the one part no tool can own for you, because owning it is the job.
There's one more kind of efficiency that gets missed, and it's information. A good automation gives you what you need, when you need it. Not more, not less.
I've used AI tools that answer a yes or no question with a history lesson. Ask whether the sky is blue and you get the history of the color blue, the history of the sky, and then, finally, "yes." All I wanted was the answer. If it takes five more prompts to get it to shorten up, the tool has cost me time, not saved it.
The tools I've set up well work the other way. I ask a yes or no question, I get yes or no, and maybe an offer: "Want more detail?" Usually I don't. When I do, it's there.
This matters most at the decision steps. The person approving needs the summary, the flags, and the answer, not a report to dig through. Overload is its own kind of inefficiency. And it's a useful test for any automation: if getting the right output takes more work than doing the task yourself, that step isn't ready to be automated yet. A good tool gives you what you asked for. It can offer more. It shouldn't bury the answer.
If you're working out which parts of your processes to hand to AI and which to keep, that's the work I do with founders every day. Let's talk.
