Before You Aim for 80%, Design the Repeatable-Work Audit
Before You Aim for 80%, Design the Repeatable-Work Audit
Say you run a design studio. Every time a client signs, your operations lead sets them up. She pulls the details from the signed agreement, checks the scope against the proposal, creates the client in your billing system, sends the first invoice, makes the shared folder, sets up the project and writes the welcome email with the kickoff link.
None of it is written down. When someone new needs to learn it, she shares her screen and talks them through it. "This is the field I copy from. This is where I slow down. If the scope changed, I ask the account lead before I go any further."
Then the call ends, and the next person rebuilds the process from memory.
Now say you want to hand off 80% of client setup. That's a reasonable thing to want. It's also a target you've set, not something anyone has measured. Before the number means anything, you need to know what it's counting.
The percentage comes last. First you have to make the work visible.
A recording is something to review
The easy version goes like this: record her doing it once and call the process handled.
Real work is less tidy. Michael Polanyi put it plainly in The Tacit Dimension: "we can know more than we can tell." Your operations lead makes small judgments she never says out loud. She notices when the client's legal name on the agreement doesn't match the proposal. She knows the billing system will text a sign-in code to her phone. She recognizes when a client who pays through a procurement portal needs a different first invoice.
So ask her to set up a real client while explaining what she's trying to get done at each step, and record that. Doing it live, with the agreement open in front of her, she'll mention checks she would leave out if she described the job from memory.
Watch it back with her. Can you name each step, see where its information came from and point to the places where she stops to check or ask? If you can't, all you have is a video.
Put the workflow together in order
Start with what she actually did and what she said she was trying to accomplish. Then review it together. Do the steps hang together? Does each one trace back to something she did on screen? Did the recording miss the thing she always watches for, like the scope check?
Next, write it down. Say what client setup is for and list the steps, each one pointing back to the moment in the recording it came from. Then decide which steps wait for a person. The first invoice bills the client and the welcome email lands in their inbox, so someone approves both before they go.
Only then, try it. Run it once as a test where nothing real changes: no invoice goes out, no email is sent, no folder is shared with anyone. Until the first real run, no client has actually been set up.
NIST's AI Risk Management Framework, which is voluntary, includes two of these steps: define the specific tasks an AI system will support, and document how people oversee it. For the studio, that's the list of setup steps and the name of whoever approves the invoice and the email, written down before anyone counts anything.
Decide what the 80% is counting
An 80% target can sound exact while hiding every decision that matters. Eighty percent of what, over which weeks? Which steps were eligible, and who decided they were low risk?
Write the count as a sentence a colleague could check:
For [this named workflow], during [this period], we will count the eligible, low-risk steps that went through the reviewed workflow.
For the studio, that might read: "For new-client setup, during March, we will count the eligible steps that went through the reviewed workflow." Nobody reading that will think you mean 80% of her job. It covers one workflow for one month, and only the steps you've agreed are eligible.
Then decide what stays out of the count:
- High-risk actions, like sending the first invoice or the welcome email.
- Sign-ins that need a person, like the code the billing system texts to her phone.
- Exceptions that don't match the reviewed workflow, like the client who pays through a procurement portal.
- Human decisions, like what to do when the signed scope differs from the proposal.
That leaves three of the seven setup steps: pulling the details from the agreement, making the folder and setting up the project. The scope check, the billing setup, the invoice and the welcome email stay with her. If six clients sign in March, you're counting 18 step-runs, not "client setup." When one of them hits a sign-in prompt or a question nobody planned for, don't squeeze it into the percentage. Record it as an exception, then decide whether the workflow needs a change, a tighter boundary or a handoff to a person.
Keep a record you can read
You can't judge a target like this by a feeling that setup went faster in March. You need a record that answers plain questions about the month:
- Which runs started, finished, stopped or were left unresolved?
- Which eligible steps went through the reviewed workflow?
- Where did someone have to approve something before the work could continue?
- What failed, and why?
- What fell outside the workflow altogether?
- Should you keep the workflow as it is, revise it, narrow it or stop using it for this work?
A finished run isn't the same as a correct one. If the workflow made the shared folder from last year's template, the step finished and the work was still wrong. Someone still has to open the folder and check it before the client sees it.
When April comes, you decide what to change from the record instead of from anyone's memory.
Exceptions belong in the record
Teams often treat exceptions as proof the work wasn't really repeatable. Sometimes they point to a bad boundary. Often they're just what the work is like.
Some of them repeat: the billing system will text her that code every time. Others turn up once, like a purchase order the recording never saw, or a client asking, "Can we split the deposit across two invoices?" No step in the workflow can answer that one, so it comes to her.
Each of those should land in the record where you can see it, not in someone's inbox or a late-night message to your operations lead.
They also show you where to start. Don't begin with the whole of client setup. Begin with the part that's stable enough to describe and low risk enough to put through a reviewed workflow. For the studio, that's the three steps in March's count.
Running it on Supanova
Say you describe client setup to Supanova in plain words, using the steps you wrote down. Mission Capture compiles that into a workflow, lets you test-run it without touching anything real, then keeps making it better. So the first run is the test, and nothing reaches a client. After that, finished work comes back for your review, which means your operations lead still checks the folder, the project and the drafted welcome email.
The audit stays yours: which steps counted, what stayed out and what you'll change for April. We want the person who knows where client setup gets tricky to be the one who decides what's fair to count.
Start with one workflow you could show tomorrow
March's record might show that only the folder and the project belong in the workflow for now, because mismatched legal names keep tripping up the agreement step. That's two of seven steps, about 30% of client setup. Eighty percent may be fair after a few revisions. Or you'll decide the invoice never leaves your operations lead's hands.
Pick one process you could show someone tomorrow, and ask the person who knows its edges to do it for real while you record.
Every part of the audit builds on that recording. That's how you earn the right to say 80% later and mean it. When you're ready to try it with an autonomous workforce, start at supanova.team.