The first time I asked a client to walk me through their "approvals workflow" before we automated it, three different people gave me three different versions. None of them were wrong. There simply wasn't one workflow. There were three, running in parallel, depending on who initiated, who reviewed, and what the dollar amount was. The org had been running like this for years, and most days it worked.
Then we tried to give the workflow to an agent. An agent has to pick one. And the conversation that followed — about which version was actually the workflow — was the most valuable thing we did that quarter. Not because we automated anything yet. Because for the first time, the team agreed on what they were doing.
"You can't deploy what you can't describe. Most rollouts fail because the description happens after the build, not before it."
The pattern under the chaos
Every team has workflows that exist as muscle memory. The senior person knows when to skip the form. The finance lead knows which approvals are real and which are theatre. The implicit rules are why the system runs at all — and they're exactly the rules an agent has no access to.
So before you build, you document. Not the org chart. Not the policy. The actual sequence of things that happens when this work gets done well. Who initiates, who's notified, who decides, who's blocked, who gets cc'd, who never finds out. Six fields. Most teams have never written it down.
The pre-AI workflow audit, in five steps
Pick three workflows. The top three by hours-eaten or by "what would break if our best operator left." Not the easy ones. The expensive ones.
Walk through one in real time with the person who owns it. Not in a meeting. Not from memory. The next time it actually happens, sit next to them and write down every step — including the ones they forgot to mention because they're now reflexive. The gap between the documented version and the reflexive version is where the agent will fail.
Capture exceptions explicitly. "Usually we send to the client manager, but if it's over $50K it goes to legal first; if the client is in EMEA we cc the regional partner; if it's a renewal we skip the second review." These aren't edge cases. They're the workflow. The agent needs all of them.
If your workflow document doesn't have at least three exceptions, you didn't document the workflow — you documented the sanitised version of it. Real workflows have edge cases. Pretending they don't is how rollouts fail in week six.
Map the data each step touches. Where does the input come from? Where does the output land? What systems does it cross? Most workflows look simple in a slide and turn out to involve six tools, two of which weren't sanctioned by IT. The agent has to read from and write to all of them — or know not to.
Identify the human checkpoints before you remove them. Some are real (irreversible action, asymmetric downside, expensive judgment). Some are habit (the manager wants to feel involved). Mark which is which. The first set stays. The second set is where the productivity gain lives.
What you find before you build
On the project management restructure I just finished — a 40-person professional services team — the documentation pass alone surfaced things the leadership had been arguing about for months. Two teams were running parallel intake processes that didn't sync. A weekly status meeting was ingesting the same data from three places. The "approval gate" everyone referred to in the abstract turned out to be a Slack DM the COO sent on Tuesdays, sometimes.
None of that is an AI problem. All of it is a workflow problem the AI would have inherited and amplified. Documenting first means you fix the workflow once — then automate the fixed version, not the broken one.
The artefact that survives
The output of a pre-AI workflow audit isn't a flowchart in a tool nobody opens. It's a one-page document, written in plain language, that anyone joining the team in six months could read and follow. Title, owner, trigger, steps, exceptions, systems touched, checkpoints, output. Eight fields. Two pages maximum.
That document does three things. It lets the agent be configured against a real specification. It gives the team a shared definition of "done well" so the human reviewers know what they're checking. And the third — most important — is that the document outlives the rollout. Six months later, when the team wants to extend the workflow, the document is already there. You don't re-document. You amend.
The cheapest hour in any AI rollout is the one you spend before the build. The most expensive month is the one you spend trying to retrofit documentation onto an agent that's already been deployed. Document first. Deploy second. Don't reverse the order.
An hour of mapping saves a week of debugging. The pre-AI documentation pass is the cheapest insurance policy in any rollout.
Start the conversation →