July 11, 2026 8 min read

Medical billing automation: what production deployments actually look like

Joe Hewett

Joe Hewett

CTO & Co-founder

healthcare billing rcm automation browser-agents payer-portals

TL;DR

  • Medical billing automation means removing the manual portal work inside the revenue cycle: eligibility checks, claim status follow-up, payment posting, enrollment, and denial work that staff run in payer portals and EHRs all day.
  • Most of that work stays manual because the systems involved expose no usable API. The status quo for billing companies at scale is offshore delivery teams doing the login, lookup, fill, submit, and log loop by hand.
  • Browser agents automate that loop through the same portal screens a person uses. One production fleet runs roughly 26,500 automated executions a month across an enrollment and CRM workflow, with about 13% of volume handled by a dedicated recovery workflow.

Medical billing automation, concretely, is removing the manual steps between “the claim data exists” and “the money is posted”: eligibility checks, claim submission and status follow-up, enrollment applications, payment posting, and denial work, the work your delivery team runs in payer portals, clearinghouses, and EHR screens all day. It sits within our broader guide to healthcare workflow automation, alongside the claims and denials workflows this same delivery team runs.

The revenue cycle, mapped to what’s actually automatable today

StageWhat it involvesAutomatable today
Eligibility & benefitsVerify coverage before the visitYes: portal check, structured write-back
Charge capture & codingTranslate the encounter into billable codesNo: clinical and coding judgment
Claim submissionSend the claim to the payer or clearinghouseYes: EDI where it exists, portal where it does not
Claim status & follow-upCheck pending claims, chase stalled onesYes: status checks plus a recovery workflow
Payment postingPost remittances to the right claim or accountYes: structured write-back from the ERA or portal
Denial managementInvestigate and argue denied claimsPartial: triage yes, the appeal argument stays human

The pattern: portal logins for a deterministic lookup, fill, or submit are automatable now; work that requires clinical or coding judgment stays with a person.

26,500 portal executions a month, and what the numbers teach

One US health-insurance call-center operation runs its entire back-office enrollment layer on browser agents. Phone agents capture the deal; from there, agents log into the marketplace platform, reconcile consumer data, complete the application, select a compliant plan, attest and sign, and write the confirmation ID back to the CRM. Five production agents carry roughly 26,500 executions per 30 days, about 870 a day.

Three numbers from that fleet matter more than any feature list.

First, the enrollment workflow is 18 nodes and 34 transitions: duplicate detection, identity verification failures, no-acceptable-plan cases, CRM write-back on every failure path. Anyone demoing a five-step happy path hasn’t built the workflow yet.

Second, recovery is its own workflow. About 3,500 executions a month, roughly 13% of fleet volume, is a dedicated agent that finds deals stuck mid-flow and either finalizes them or routes them to a human with a structured note. Stuck work is a recurring production fact, and a deployment with no recovery lane hasn’t been tested at volume.

Third, correctness beats completion. Individual agents in production have run over 10,000 executions without a false positive, built to fail cleanly (member not found, identity ambiguous, no valid plan) rather than force a bad submission through. In billing, a false positive becomes a denial four weeks later.

Figure: the production enrollment pipeline, live. One request leaves the call + CRM intake, the agents work the marketplace platform and write the confirmation ID back to the CRM, every step ticked off and recorded — with a dedicated recovery workflow carrying about 13% of volume.

Every manual touch in that pipeline used to be headcount that scaled with deal volume. If your delivery team grows every time you sign a client, book a demo and watch the automated version run on a workflow you actually deliver.

Why billing work is still manual: the API gap

The revenue cycle has more standardized rails than most of healthcare. Clearinghouses carry the 837 claim and 835 remittance transactions, and eligibility has the 270/271 pair. If your payer mix speaks EDI, meaningful automation already exists.

The work that lands on a delivery team is everything the rails do not reach. Denial follow-up and appeal submission live inside per-payer portals. Benefit detail beyond “plan is active” is frequently on-screen only, which is why teams still run checks in multi-payer portals by hand. Credentialing and payer enrollment run through systems with no write API at all, and the long tail of regional payers and Medicaid managed-care plans each bring a portal of their own.

For a billing company, the long tail is the product itself. A new client brings a new payer mix, a new EHR, and new portal credentials; the standard answer has been to hire for it. Hospice, home health, infusion, and DME billing run the same pattern, just with a different specialty EHR at the center.

Robotic process automation for medical billing: where it fits and where it breaks

Selector-based RPA records a script against a specific page layout, then replays it: click, type, submit. That works on a payer portal that never changes, and it’s cheaper to stand up than an agent fleet. But payer portals redesign form fields, add interstitial pages, and branch by plan or procedure without warning, and a recorded script cannot recognize a page it wasn’t trained on. Most RPA billing deployments stall after two or three payers: each new payer is a new recording, and every layout change is a maintenance ticket. Browser agents read the page they’re actually on rather than replaying coordinates, which is what lets the same approach scale across payers instead of breaking on the next one.

The offshore status quo, priced honestly

The dominant way billing and RCM services firms run this work today is offshore delivery teams: AR callers, eligibility verification specialists, charge entry and payment posting staff, credentialing specialists. Calling offshore delivery a dirty secret gets one word wrong: everyone in the industry knows, because it is simply the business model, and it persists because it works. Any automation pitch that pretends otherwise is not serious.

But the model has a cost structure the operators running it know intimately. Headcount scales linearly with the client roster, so margin compresses as you grow. Churn forces a standing training machine: new-hire classes, nesting, QA, recertification, run continuously, and payer knowledge walks out the door with every departure. And HIPAA exposure grows with every person who touches PHI.

The uncomfortable observation underneath all of it: the work itself is deterministic. Log in, look up, fill the form, submit, poll the status, log the result. That loop is exactly what the seats exist to run, and exactly what software can now run instead.

Medical billing automation software: the real options

An operations leader deciding how to take that loop off human hands has five realistic paths.

OptionWhere it worksWhere it breaksCost shape
API / clearinghouse integrationStandardized EDI transactions on the largest payersThe portal-only work that created the delivery teamIntegration project per payer
Offshore ops teamsThe status quo above: anything a trained specialist can do in a portalTurnover, the standing training machine, payer knowledge that leaves with each departureLinear in headcount; per-transaction cost is the benchmark to beat
Selector-based RPAStable portals that never changeLayout changes break the recording (see above)Cheap to record, expensive to keep alive
Internal scriptsThe top one or two payers your engineers script firstThe long tail every new client brings; maintenance burden stalls it thereEngineering time, forever
Browser agentsAny portal task a person could do in Chrome; the agent reads the page rather than matching selectors, and escalates when a run cannot completeWork that genuinely needs judgment, which routes to a human by designScales with credentials and runs, not headcount

The last row has the property that matters when your next client brings a new portal: the coverage question stops being “which payers have we integrated” and becomes “can a person do it in Chrome.”

Asteroid’s role here, stated plainly: Asteroid ships prebuilt, HIPAA-compliant browser and desktop agents for billing and payer-portal workflow categories, run as managed fleets with concurrency controls, structured output, audit trails, and human escalation built in. That is a category-level claim about the work, not a partnership or integration claim about any named portal or payer. The 26,500-execution fleet above is this option in production.

The real economics of medical billing automation

Two operational facts from production fleets should anchor any business case.

Per-run speed is not the value: a branching payer-portal run can take 10 to 15 minutes, not faster than a specialist on a single check. The value is parallelism and off-hours batching: dozens of runs overnight, unattended, staff arriving to a completed queue. One healthcare automation team cleared a backlog of 2,000 unprocessed referrals in about two weeks this way.

Throughput is governed by credentials, not compute. Payer portals flag and lock accounts that open many concurrent sessions on one login, so production fleets run per-credential concurrency caps, around five sessions per credential at one deployment. Your ceiling is credentials times cap; rotation becomes part of your SLA.

Agents configured once per client-and-portal pair keep the payer-quirk knowledge that otherwise evaporates with staff turnover.

What escalates to a human, and why

No production billing automation is fully automated, and the exception queue is where credibility lives. From production fleets, the recurring escalations:

  • Identity and data mismatches. A member ID that doesn’t resolve fails cleanly with the searched values attached, rather than guessing at a near-match under the wrong member.
  • No valid option. No plan meets the constraint, or a denial needs an argument, so it routes to a person, not a retry.
  • Duplicates. An existing application or claim gets flagged and annotated in the CRM, never silently resubmitted.
  • Portal downtime and maintenance windows. Batch runs reschedule; anything still failing at the SLA deadline surfaces to staff.
  • MFA and credential rotation. OTP prompts and forced resets escalate immediately to a designated person, because lockouts are an SLA event.

The design principle: refuse to succeed incorrectly. The exception queue is not the automation failing. It is your staff working only the items that genuinely need a human, with the run context attached. Prior authorization work, one step upstream of the claim, has the same shape; the prior authorization case study shows the pattern applied there.

The cost of staying manual

Every eligibility check skipped for capacity reasons is a coin flip on a downstream denial, and every denial worked by hand costs more than the claim’s margin. If delivery headcount decides how many clients you can take, that is the real cost of the status quo, and it compounds with every contract you sign. Book a demo and watch a billing agent run a real branching portal workflow, structured write-back and exception queue included.


Asteroid is not affiliated with, endorsed by, or partnered with any payer or portal vendor referenced. This content is educational; product names are used for identification only.

Frequently asked questions

The deterministic portal work: eligibility verification, claim status checks, enrollment applications, payment posting, credentialing lookups, and structured write-back into a CRM or practice management system. Coding judgment, payer phone calls, and patient conversations stay human. Production deployments automate the repetitive loop and route exceptions to staff.

Joe Hewett

Joe Hewett

CTO & Co-founder

Joe Hewett is co-founder and CTO of Asteroid, focused on the agent runtime and the platform that builds, runs, and monitors browser agents at scale.

See a billing agent run a real portal workflow.

Watch the automated version run on a workflow you actually deliver: branching payer-portal work, structured write-back, and the exception queue included.