· Saul Mateos · AI-Native CFO · 5 min read
The 70% That Ships: Perfectionism in Finance Teams
Our check-matching tool shipped at 70%, not 100%. Why perfectionism in finance teams is the slowest goal, and three questions to ask before you ship.
Two weeks ago I told you I sent my best FP&A manager to deposit checks. Next week, the thing he built ships to production. It’s my best argument against perfectionism in finance teams.
It automates about 70% of the work. Not 100%. Not 95%. Seventy.
If you close this tab right now, that’s the lesson: stop waiting for the perfect version. But since you’re still here, let me tell you why the number matters.
What he built
Quick recap for the people who missed it. He spent a week shadowing cash receipts. Downloaded checks from the bank portal. Uploaded the PDFs. Matched each check to the right Salesforce opportunity by plaintiff, amount, property address, and law firm. Created the cash receipt. Posted it. Moved on.
It used to take two or three people several hours a day.
He didn’t write a memo about it. He built a Claude skill that does it. 25 checks per batch. Matches the opportunity automatically. Creates the cash receipt on the right account. We ran it against two weeks of real data before I signed off.
It works on about 70% of checks cleanly.
The other 30% have ambiguity. Same plaintiff, two law firms. Weird property addresses. An amount that doesn’t quite tie. A check that references a case we’ve already closed. The kind of edge cases where a human needs to look and decide.
The version that didn’t ship
We had a conversation two weeks ago about whether to keep building until it handled everything.
If we pushed for 90%, we’d ship in July. If we pushed for 95%, maybe September. If we pushed for 100%, we’d never ship. There’s always another edge case. There’s always a reason the model isn’t confident. There’s always a human who swears the system will “definitely get this one wrong.”
Here’s what I told the team. We ship 70%. The other 30% route to a human. Same person who used to do 100% of the work now does 30% and supervises the rest.
Their day just got shorter. Our accuracy just got higher (because when the model is confident, it’s more reliable than the tired human at 4:30 PM). And we have six months of real production data to feed back in before we push to 90%.
70% ships. 70% compounds. 70% buys you the data to go to 90 next quarter.
The same pattern somewhere else
Our risk management agent is the same shape.
Every open case in our servicing book gets reviewed. The agent pulls the account, reads the history, drafts a next action and an email to the attorney. A human clicks send on every one.
I’ve said it out loud to the team. I will not trust full automation on anything that touches a customer for at least another year or two. Not because the agent can’t write a good email. It can. But because the cost of one wrong email to the wrong law firm is not a rounding error. And the cost of a human clicking send on 80 drafted emails a day instead of writing 80 from scratch is massive.
Same shape. AI does the draft. Human does the decision. 70% automation. 100% oversight.
Perfectionism in finance teams: why 100% is the slowest goal
Every team I’ve run has a version of this debate. Same conversation.
“Can we ship it if it’s not perfect?”
The answer is always yes. The answer has to be yes. Because “perfect” is how you get budget clawed back in Q3. “Perfect” is how a competitor ships something worse but ships it. “Perfect” is how your best builder burns out trying to close the last 5% while the first 70% sits on a laptop.
100% is the slowest goal in the building. It’s also the one that gets you fired if you’re wrong, because by the time you realize the edge cases weren’t really edge cases, you’ve missed the window.
70% is the fastest real-world goal. It forces you to think about what a human handles, what escalation looks like, what the fallback is. It produces a system that works on day one instead of a demo that works in a deck.
How to tell if you’re ready to ship at 70%
Three questions I ask before we push anything live.
- What happens when the model is wrong? If the answer is “a human catches it and redirects,” you’re ready. If the answer is “a customer finds out,” you’re not.
- Who owns the 30% that didn’t automate? If nobody’s name is on it, don’t ship. Something unowned degrades fast.
- Can we tell the difference between “the model got it right” and “the model was confident”? If you can’t see the confidence score, you’re flying blind. Build the instrumentation first.
If you can answer all three, ship it. Don’t wait for the steering committee. Don’t wait for the next sprint. Don’t wait for the 3% edge case you can barely reproduce.
Ship the 70%. Watch what breaks. Fix the thing that actually broke, not the thing you were afraid of.
What I’d tell you if we were grabbing coffee
I’ve shipped a lot of half-finished things this year. Not careless. Deliberate. A check-matching tool at 70%. A risk management drafter at 0% send (human always clicks). A variance commentary template that handles the five most common accounts and punts on the rest.
None of them are the version I pitched in the deck. All of them are in production.
The version in the deck doesn’t save anyone an hour. The version that ships at 70% saves someone three hours a day starting Monday.
Pick the version that ships.