Healthcare billers still supply the keystrokes AI hands back to them

Electronic-transaction rates measure whether a standardized rail exists for an exchange and whether traffic used it. They do not measure whether a person had to decide something before the rail could carry anything. That is why CAQH can put eligibility at 96% electronic while prior authorization sits at 40% ^[2][3] without either figure being wrong: eligibility is a lookup, and prior authorization is a judgment with keystrokes attached. Attachments make the point harder, at 24% in medical spaces and declining ^[4], which leaves roughly three-quarters of that traffic with no standard transaction to ride ^[2]. Oleg Nesterov's summary of the pattern is that everything with a standard transaction gets automated and everything that requires deciding and then acting gets a work queue and a person to sit in it ^[5].

The distinction he draws is between coordination and execution: a coordination system tells a person what to do, an execution system does it, and by his account revenue cycle management has run the first for two decades while calling it the second ^[15]. Most RCM AI scores a denial by its likelihood of overturn and drafts the appeal, then routes the task back to a person to log in and finish ^[6]. Staff still open as many as 11 portals a day ^[7], and a single fully electronic prior authorization can mean eligibility in one portal, submission in a second, status in a third and documentation in an EHR connected to none of them, which he estimates at 10 minutes or more of manual work ^[8].

The two MGMA readings are the closest thing here to a labor measurement, and they need handling. Sixty-eight percent of medical groups added or expanded AI tools in 2025 ^[9]; in a different poll with a different set of 260 respondents, 68% said AI had not led them to redesign a single role or change staffing ^[10]. Those are separate populations, so the pairing is directional rather than a before-and-after. What makes it credible anyway is the mechanism: a tool that ends by handing work to a biller cannot remove the biller, and the 60% of prior authorizations that are not fully electronic ^[1] are precisely the work that ends that way.

The diagnosis is convenient for its source: Nesterov's firm builds AI agents and custom software for US revenue cycle management ^[1], and a market that bought advice and still needs action is a market for what he sells. Two points hold up regardless. His load-bearing figures come from CAQH, MGMA and HFMA rather than his own book ^[2][9][12], and his criteria are concrete enough for a buyer who does not trust him to test: holding credentials through two-factor prompts and session timeouts, which he calls the unglamorous 80% of the job; reading remittances, denial letters and payer policy PDFs that are often scanned; and authority to correct, resubmit or write off under rules that change monthly ^[13]. The material contains no case of a system meeting those criteria reducing billing headcount, so on labor cost the honest answer is that we do not know yet.

So the choice this quarter is narrower than the category pitch. Most groups have already bought something ^[9]; what remains open is whether the next renewal is scored on completed actions rather than accurate recommendations, because a tool that terminates in a work queue commits the practice to staffing that queue for as long as it owns the tool. On the other side, a payer engine returns a decision in seconds and the provider answers with a person opening a portal, reading the reason and assembling the appeal by hand ^[14]. That asymmetry compounds on a monthly cycle, not a decade-long one.