Most practices didn’t choose to have their physicians typing notes field by field into an EHR, coding their own visits, or chasing down a denial between patients. It happened gradually: the automated record needed structured data, the billing needed codes, and both took judgment — so both landed on the most expensive, most credentialed person in the building. The clerk who used to take dictation is gone. Nobody replaced her; the typing just moved up.
AI is repeating the pattern. A model drafts a note or suggests a code, a human has to review it, and “review” quietly becomes the provider’s job again — because reviewing feels like it requires the same judgment as doing it from scratch. Automation didn’t remove the work. It just relocated it to whoever was left standing when the mechanical part disappeared.
That’s not how automation has to behave. It’s what happens when nobody decided otherwise.
The check we run before automating anything
Before we build automation into a practice’s workflow, we ask four questions — in this order, every time:
- What does this task actually require? Separate the mechanical part (data entry, lookups, formatting, routing) from the judgment part (a decision that changes the outcome if it’s wrong). Most tasks billed as “requires clinical judgment” turn out to be mostly the former.
- Who is the cheapest role actually capable of supplying that judgment? Not who currently does the task — who should, if the current owner is just whoever the mechanical work fell on by default. It’s almost never the provider.
- Automate the mechanical part. Route the judgment part to that role, not to whoever’s most senior. The automation’s job is to move data, draft the format, and queue the review — not to hand a rough draft to the most expensive person in the building and call that “AI-assisted.”
- Wire it in two stages, source system to structured intermediate to consumer, so the pattern survives the practice changing EHRs or billing systems later.
If a new tool ships and the review step still lands on the same person who used to do the whole task by hand, question 2 got skipped.
What this looks like in practice
We run full billing operations — insurance verification, charge capture, claims submission, denial management, AR — for the practices we work with, against contracted turnaround targets. That’s the same instinct as the check above, applied to an entire function: billing was never meant to be “whoever has five spare minutes,” and in a lot of practices, that quietly ends up being the physician or the front desk. We built it as a dedicated service instead, so it has an owner who isn’t the provider by default.
The same question applies smaller-scale, inside a single workflow:
- Payment posting. EOB-to-ERA conversion and posting is pure mechanical translation — no clinical judgment involved at all — so it should never have been eating staff hours in the first place. See how we handle it.
- Denial and underpayment tracking. Spotting a systematic underpayment is a data-matching problem, not a judgment call. The judgment only shows up when you decide whether to appeal — that’s where a person should be spending time, not on the detection. More on our billing analytics.
- Records and fax routing. Sorting incoming documents to the right chart or the right person is mechanical triage that easily defaults to “whoever’s near the fax machine” — usually front desk, sometimes clinical staff. Routing it automatically keeps it off both.
None of this is about removing people from the loop. It’s about making sure the person in the loop is there because their judgment is actually the scarce resource — not because they were the last stop after the mechanical work got automated out from under everyone else.
The question worth asking about your own practice
Look at whoever on your team is currently doing a task that has “review” or “verify” in the description. Ask whether that’s their job because their judgment is genuinely needed there, or because a tool upstream drafted something and the review just landed on whoever was already using the software. If it’s the second one, that’s not a training problem or a staffing problem. It’s a routing problem, and it’s fixable before the next tool gets bolted on.