A fractional AI ops department for healthcare, claims paid, prior auth moving, PHI protected.

Every funded healthcare provider, digital health platform, or specialty clinic group between Series A and Series B shows up to the audit with the same problem shape. Claims processing runs against ten to twelve payor schemas, each with its own CPT code rules, modifier requirements, and denial reason codes. Prior authorization sits in front of the most expensive treatments and the average prior-auth turnaround is two weeks per payor per request. The denial follow-up queue grows faster than the team can work it, so claims hit timely-filing limits and write off. Billing reconciliation against the EHR, the practice management system, and the clearing house is a Monday morning panic that bleeds into Tuesday. PHI lives in every workflow, which means the architecture conversation starts before the first agent goes live.

The team running this is usually one billing manager, one prior-auth coordinator, one or two denial-follow-up associates, and an external billing service for overflow. Year-one loaded cost between $250K and $400K depending on geography and specialty mix. The billing service charges 4 to 7 percent of collections on the work it touches, which on $5M of annual collections is another $200K to $350K. The denial rate sits between 8 and 15 percent. The clean-claim rate sits between 80 and 92 percent. The collections curve gets shaped by which claims get followed up before timely-filing, which gets shaped by how much the team has bandwidth for that month. Nobody is paid to track the claims that quietly write off.

The default fix in the leadership meeting is the four-hire ladder. A senior billing specialist, a denial analyst, a prior-auth specialist, and a revenue cycle manager by next year. Year-one cost north of $350K loaded, recruiter fees stacked on top. Most of them spend their first quarter learning the payor mix and the CPT code rules. The denial rate barely moves because the underlying problem is throughput, not skill. The other path is to take the healthcare ops function and run it as a fractional AI Ops Department tuned for healthcare data shapes from day one, with on-device handling for every workflow that touches PHI.

Healthcare ops is structurally harder than SaaS ops because PHI lives everywhere and the regulatory floor is higher. HIPAA, the BAA with every vendor, state-level rules on top of federal floor, payor-specific data handling clauses. The architecture is not optional. The sanitized cloud lane handles admin work that touches no PHI: vendor invoice processing, internal copilot for the policy wiki, expense reconciliation, dashboard rendering for aggregate metrics. The on-device lane handles every workflow that touches a claim, a prior-auth request, a denial appeal, or a patient record. Same agent operator supervising both lanes, audit trail on every step.

The reason healthcare claims processing is harder than it looks is that every payor models a clean claim differently and the rules shift under the surface. UnitedHealthcare requires a specific modifier combination for telehealth visits in one state and a different combination in the next. Aetna decides a CPT code needs a corresponding diagnosis at the modifier level that the billing manager was not warned about until a denial batch came back. Medicare has its own LCD policies by MAC region. Medicaid varies by state. The commercial payors have their own clinical edits that get updated quarterly with three weeks of notice and a 200-page changelog the team does not have time to read.

A billing specialist who is excellent at the work catches about 90 percent of the modifier and code edits before submission. That gets the clean-claim rate to the high 80s or low 90s. The rest go out the door, come back as denials, and land in the follow-up queue. The follow-up queue has its own throughput limit. A denial analyst can rework 25 to 40 denials a day depending on complexity. At a $5M-collections practice generating 800 to 1,500 claims a month with a 12 percent denial rate, the math is roughly 100 to 180 denials a month. One analyst can keep up if the queue is steady. Two analysts can keep up through a spike. A spike that lasts three months (a payor changes its edit policy, a new service line launches, a coder leaves) and the queue blows past timely-filing windows. Claims write off. Collections drop. The CFO finds out in week ten of the slip.

A fractional AI Ops Department for healthcare reads the payor edit policies as they update, applies them against the claim at submission time, and pre-empts the modifier and code mismatches that would otherwise come back as denials. The clean-claim rate moves from the high 80s into the mid 90s within the first ninety days. The denial queue shrinks. The follow-up team finally has bandwidth to work the denials that actually need a human appeal letter, which is the high-dollar, complex denials where the appeal recovery rate is highest. The collections curve straightens.

The on-device lane is where the claim content lives. PHI never leaves the perimeter. The local model reads the chart note, the CPT code, the diagnosis, the payor card, the prior-auth status, and the modifier requirements. It assembles the clean claim, runs the edit checks, and submits through the clearing house. The audit trail captures every step. HIPAA exposure is contained because the model that touches PHI runs inside your perimeter and the cloud lane never sees a patient record. The sanitized cloud lane handles aggregate metrics, vendor invoices, and the policy copilot.

Prior authorization is the single most user-hostile workflow in American healthcare and the single largest holdup on patient care for funded providers. A clinician orders an MRI on Monday, the prior-auth coordinator submits the request on Tuesday, the payor SLA says fifteen business days, the patient calls on Friday asking why nothing has been scheduled, and the coordinator does not have an answer because the payor portal does not return a status update until the decision lands. Three weeks later the auth comes back denied because the chart note did not document the conservative treatment trial the payor required, the coordinator submits the appeal, another two weeks pass, the patient finally gets the MRI six weeks after the order was placed.

The patient experience is bad. The clinical staff frustration is high. The operational cost is hidden because nobody puts a dollar figure against the coordinator hours, the appeals work, the scheduling rework, the patient retention impact. A funded specialty practice loses a non-trivial fraction of patients between order date and treatment date specifically because the prior-auth wall keeps the patient waiting until they go elsewhere or give up.

A fractional AI Ops Department tuned for healthcare pulls the prior-auth assembly into the on-device lane. The local model reads the chart note, identifies the required documentation per the payor policy, formats the request to the payor template, and submits through the portal or the fax queue at the moment the order is placed. The assembly step that took a coordinator forty minutes per request takes the agents ninety seconds. The chart note review step that used to surface missing documentation three weeks late (when the denial came back) catches it before submission. Conservative-treatment trial documentation, imaging history, prior failed therapies, clinical guideline citations, all assembled into the request at first submission.

The compounding effect is enormous. Patients get to treatment two to four weeks faster on average. Clinical staff stop fielding the "why is this taking so long" calls. The coordinator team has bandwidth to work the high-complexity prior-auth appeals where a clinician peer-to-peer is required. The patient-retention rate in the order-to-treatment window improves. The reputation conversation with referring providers stops being defensive. None of those wins fit on a spreadsheet but every one of them shows up in the next quarterly leadership review.

The billing manager at a funded specialty practice is one of the most under-supported roles in healthcare. They own the payor relationships, the contract negotiation conversations, the credentialing renewals, the fee schedule reviews, and the denial appeal strategy. They are also the person rebuilding the spreadsheet on Monday morning when the denial batch from UnitedHealthcare lands and the modifier rules changed two weeks ago without a notice. In most practices they are also the person fielding the call from the clinical team about why a patient cannot get scheduled because the prior auth is sitting in someone's queue.

Get the spreadsheet and the prior-auth queue out of their life and the same person can operate a practice twice the size. The billing manager who was holding a $5M-collections specialty practice together by hand can run the revenue cycle for a $15M-collections multi-site group without breaking sweat, because the clean-claim assembly, the prior-auth submission, the denial triage, the appeal drafting, and the billing reconciliation are running in the background. Their time goes to the work the practice actually needs from them. Payor contract renegotiation, fee schedule reviews, credentialing maintenance, the conversation with the leadership team about the next service line.

The savings show up in three places. Billing manager retention, which is the single most under-budgeted line item in healthcare ops. Higher clean-claim rate, which means faster cash flow and lower write-off. And the operational headroom to add a service line or a new site without hiring four people first, which is what funds the next stage of growth.

There is one more line that does not fit on a spreadsheet. The clinical staff and the patients notice when the practice operations work. Faster prior auths, fewer billing disputes, cleaner statements, faster appeal turnarounds on legitimate denials. Patient retention improves. Referring provider relationships improve. The reputation conversation stops being defensive. The fractional fix is what funded practices are supposed to do at this stage and almost none of them do at the right time.