Exception Math - The 1% That Decides Whether I Ever Hire Again

July 2026

Last week I was arguing with Claude about our org chart. Yes, I argue with Claude. It's the only teammate that doesn't whine to The Real Boss™ I'm being too hard on them.

Quick recap before we dig in: a few weeks ago, I wrote about onboarding a 90 loan customer in a single day. That process is approaching self-service. Docs come in, get organized, abstracted, QA'd, and loaded with almost no human touch. Call it 99% hands-free before our trained experts review everything. A week before that I wrote about my own layoffs directly due to AI.

I believe that our success implementing AI has largely eliminated entry level analyst hiring. We could double the number of clients without adding a single person. Claude mostly agreed, but with one exception.

99% automated feels like the finish line. It isn't. Because the 1% that still needs a human isn't a fixed number, it's a percentage. Grow the volume and the exceptions grow right along with it.

• 100 loans at 99% = 1 exception. A rounding error.
• 10,000 loans at 99% = 100 exceptions. That's an entire job.

The better the automation works, the more we use it. The more we use it, the more that leftover 1% quietly becomes a department. Success is what creates the workload.

Which brings me to the point Claude was trying to make:

Headcount was never really a function of customers. It's a function of exceptions.

For our twenty years in business, customers and exceptions grew together, so we couldn't tell them apart. Every new customer generated some bumps, so we staffed to customers. AI just severed that link. That means we can finally staff to the real driver:

exceptions per week x time to resolve each one = the humans we need

Headcount stops being a reactive hire and becomes a number we can actually calculate. Set a service standard on the 1% and staff to it.

I thought I had stumbled across some unique discovery, but Claude put me in my place again. In 1983, a researcher named Lisanne Bainbridge published a five page paper called "Ironies of Automation." Her conclusion, watching factory control rooms: the more advanced the automation, the more crucial the human handling whatever is left over. It's one of the most cited papers in the history of human factors engineering, and the citation rate is accelerating. She figured this out 40 years before my first ChatGPT prompt.

But whether we ever hire again comes down to one question about those exceptions: Does each one make the system smarter? Or get handled by a teammate because it was unpredictable? The flavor of the exception dictates that answer.

Repeatable exceptions - the same weird thing, over and over. Every one we solve should become a permanent subtraction. Resolve it once, feed it back into the system, and that exact case never touches a human again. These aren't really exceptions at all. They're bugs waiting to be uncovered and fixed permanently. Predictable.

Long-tail exceptions - genuinely novel, every single time. CRE is full of these. This is why softare in CRE is so darn hard and why so many of us have nightmare stories about tech built by non-CRE pros. Bespoke covenant language. A lender's one-off quirk. A doc structure nobody has ever seen. These never automate away. This is the irreducible human floor, and it's exactly where the judgment and wisdom I wrote about last month lives.

Here's why the distinction dictates my hiring plans. If every repeatable exception gets fed back into the system, my exception rate falls even while my volume climbs. Learning outruns growth. That's how we double clients with zero hires.

If exceptions just get handled but not built into our AI processes, the rate stays flat, volume wins, and I’m hiring within a year.

Same 99%. Same growth. Opposite outcomes. The only difference is whether there's a feedback loop.

Volume is pushing the exception count up. Learning is pushing the exception rate down. Whichever one wins determines whether the next hire happens.

AI didn't eliminate new headcount. It changed what headcount measures and what the team works on. Next week I’ll write about how I am shifting the company composition to account for this new dynamic while still obsessing about customer service.