Skip to main content
Viithiisys
Back to blog
Strategy7 min readWed, Nov 18, 2026

Robotic process automation in healthcare: three workflows

Robotic process automation healthcare playbook: prior authorization, denied claims and faxed referrals, with the systems involved and what stays with people.

Jatin Chhabra

AI Engineer, Viithiisys

Robotic process automation in healthcare: three workflows

Robotic process automation in healthcare: where the bot fits and where it stops

Robotic process automation in healthcare pays off on rule-bound work that crosses screens with no API behind them: payer portals, billing queues and faxed referrals. It does badly where a clinician has to decide something.

This post designs three such workflows. Each one names the systems, where the work jams, what the bot should do and what a person keeps. None of these is a delivery record. They are reference designs written from public documentation, and any number described as an illustration is hypothetical.

Provenance: Viithiisys sells software engineering, so read this as a vendor view. It is written from public payer and standards documentation, and we have no confirmed healthcare delivery record to cite.

Our engineering base is in Mohali, with an office in Markham, Ontario. Treat the three designs below as designs to test against your own queue, not as a record of work delivered. The approach they assume is the one we would build to: every step logged, a hard stop when a screen looks wrong, no real patient data in test environments, and a named person on every exception.

A claims clerk's screen showing three queues feeding one billing system, with a bot handling the repetitive steps

Why do bots break at the first portal change?

In our view, healthcare RPA most often fails at maintenance. A bot that clicks through screens depends on the page layout, so a payer portal update can stop it overnight and nobody notices until the queue grows.

Microsoft's documentation on UI elements in Power Automate desktop flows shows how a screen bot finds a button: by selectors tied to the page. Change the page and the selector can stop matching.

A generic SaaS tool has its own problem. It does not know your payer mix, your denial codes or how your scheduling team names a referral type.

So the build we would recommend for any of the three workflows below has the same bones. The bot logs every step, stops when a screen looks wrong, and hands the item to a named person. A silent failure is worse than a loud one.

Prior authorization: automate the chasing, keep the judgement

Automate the status chasing and the re-keying, and leave the medical necessity argument to a person. The jam sits at the payer portal. A coordinator reads the order in the EHR (Epic, Oracle Health or athenahealth), checks the payer's rules, logs into Availity or a payer-specific portal, re-keys the clinical details, attaches notes, and then returns every day to check status.

Each payer portal looks different. Status checks are the quiet cost, because they repeat until the answer changes.

The CMS interoperability and prior authorization final rule (CMS-0057-F) sets decision timeframes for covered payers of 72 hours for expedited and seven calendar days for standard requests, starting in 2026. It also requires payer APIs from 2027.

That changes the design. Build the bot behind a thin interface so an API can replace the portal step later.

Where the bot stops and a person starts

  • Bot does: checks eligibility, assembles the packet from the chart, submits it, polls for status and writes the result back to the EHR. A rules engine decides which orders need authorization at all.
  • Person keeps: the medical necessity argument, the peer-to-peer call and every appeal. If a payer asks for something the bot has not seen before, the item goes to a named coordinator with the page attached.

Denied claims: sort by fixability before anyone opens one

Sort denied claims by how fixable they are, not by age, and automate only the fixes a rule can state in one line. A denied-claim queue usually gets worked oldest first, so easy fixes sit behind hard appeals. The remittance arrives as an X12 835 file through a clearinghouse such as Availity or Optum, and a biller opens each claim in the billing module of Epic or athenahealth to find out why.

The reason sits in the claim adjustment reason codes, which are a published code set. Many are mechanical: a missing modifier, a stale insurance ID, a duplicate submission.

As an illustration, imagine a clinic whose denial queue holds mostly old items that are simple eligibility mismatches a person could fix in two minutes. Nobody finds them because the queue is sorted by date.

The cost is cash that arrives late or never, plus billers spending the day on lookups.

What the bot does with each denial

  • Bot does: reads the 835, groups denials by code, and sorts them into three bins: fix and resubmit automatically, route to a specialist, or hold for appeal. Only the first bin is automated.
  • Person keeps: sign-off on every appeal and the odd cases. An AI model earns its place in the second bin, reading the payer's explanation letter and summarising what is missing, but a biller still signs off.

Faxed referrals: read the page, let a person confirm the match

Referrals still arrive as faxes. A fax server drops a PDF in a shared inbox, a staff member reads it, finds or creates the patient in the EHR, keys the referral order and starts scheduling.

The jam is the reading. Handwriting, a missing date of birth, or a patient with a common name all stall the item, and a wrong match puts one person's information in another person's chart.

That is a privacy incident, so this workflow needs the most care of the three.

A faxed referral page being read into structured fields, with a low-confidence patient match flagged for a staff member

Matching the patient is the risky step

Document extraction, covered on our document AI page, reads the fax into fields. A rule then searches the master patient index, which is the EHR's list of known patients, and scores the match.

  • Bot does: reads the page, scores the match and drafts the referral order when the score is high.
  • Person keeps: every low-confidence match, with the page and the candidate matches side by side, plus triage.

The bot never creates a new patient record on its own. Any vendor that processes the fax needs a business associate agreement, which HHS explains in its business associate guidance.

What should the bot do, and what should a person keep?

The three workflows share one shape: the bot prepares and tracks, the person decides. The differences are in the systems and the failure risk.

  • Prior authorization
  • Main systems: EHR, Availity or payer portals.
  • Bot does: eligibility, packet, submission, status polling.
  • Person keeps: necessity argument, peer-to-peer, appeals.
  • Biggest risk: portal changes break the script.
  • Denied claims
  • Main systems: billing module, clearinghouse, X12 835.
  • Bot does: group by code, resubmit simple fixes.
  • Person keeps: appeals, sign-off, odd cases.
  • Biggest risk: a wrong auto-fix resubmits bad claims.
  • Faxed referrals
  • Main systems: fax server, EHR, patient index.
  • Bot does: read page, score match, draft order.
  • Person keeps: low-confidence matches, triage.
  • Biggest risk: wrong-patient match.

Rules decide, AI reads, and neither one signs

Plain rules should make every decision that can be written as a sentence. AI should read what has no fixed layout, such as a fax or a payer letter.

AI output is a suggestion with a confidence score, never an action on a chart. For structured data exchange, the HL7 FHIR standard gives a common format, so wherever an EHR exposes FHIR, prefer it over screen scraping.

A bot that cannot say "I am not sure, here is the page" will eventually file the wrong thing quickly.

Our AI workflow automation work follows the same split, and AI integration covers connecting these tools to an EHR or billing system.

Cost drivers nobody puts in the quote

Build effort is the visible cost. The hidden ones are upkeep when portals change, service accounts and access reviews, audit logging of every touch on patient data, and test data that contains no real patient information.

Exception handling takes more design time than the happy path, because every unreadable fax or odd denial needs somewhere to go.

On commercial shape, we scope this per project after a discovery call, as a fixed scope for one workflow, then support on a fractional basis once it runs. We do not publish a figure, because it would be invented before we have seen your systems and volumes.

Start with one queue and keep the first build small. A bounded first build tells you whether the pattern holds before you commit to more. Regulated, multi-system work like this takes longer to prove than a simple app, so treat a pilot on a single queue as a test of the design, not a promise of a fixed delivery date. Our Moonship 30-day MVP shows how we scope a fixed-scope first build, though EHR and payer-portal work needs its own timeline.

Send us the workflow that gets stuck

Pick the queue where staff repeat the same lookup all day, then write down who touches each step and where an item waits longest. That sheet is most of a design.

If you would rather have a second pair of eyes on it, our broken workflow assessment takes that description, the systems involved and your rough volumes. You receive a written recommendation naming which step to automate first and which to leave with people, with the reasoning for each.

Send us the workflow that gets stuck, whether it is a prior authorization portal, a denial queue or a stack of faxes.

Sources

FAQ

What is robotic process automation in healthcare?
It is software that operates other software the way a staff member would: logging into a payer portal, copying data between an EHR and a billing system, or filing a document. It suits repetitive, rule-bound tasks and is usually paired with human review for anything clinical.
Is healthcare RPA compliant with HIPAA?
Compliance depends on the design. The bot handles protected health information, so it needs its own credentials, access logging, minimum necessary access and a business associate agreement with any vendor that touches the data. HHS publishes the rules and guidance on business associates.
Should we use RPA or wait for payer APIs?
Use both, in that order. The CMS interoperability and prior authorization rule pushes payers toward standard APIs from 2027, but portals will stay in use for years. Build the bot behind a thin interface so an API can replace the portal step later.
Which healthcare process should we automate first?
Pick the queue where staff repeat the same lookup and re-key the same fields all day, and where a wrong action is easy to reverse. Denied-claim triage usually fits better than anything that writes to a clinical record.