My conversation with Hank Capps, MD, FAAFP, reminded me of a question our team often starts with when automating healthcare claims-processing workflows: Why does this process exist?

Hank's experience as a physician and healthcare executive has given him a front-row seat to how AI is being implemented across healthcare, both on the provider and payer sides. During our conversation, he explained how many workflows exist because of bad data, poor incentives, and disconnected legacy infrastructure.

Over time, teams build processes to work around these problems: manual review steps, emails followed by phone calls to providers, and other administrative tasks that become permanent parts of the workflow.

The problem with applying AI to these existing processes is that we risk investing resources in automating work that shouldn't need to exist in the first place.

Garbage in, garbage out, but faster, is not the change we're looking to make in the world!

Example One: Claim Rejection Reasons

Claims are often rejected because information required to process them is missing or invalid.

When a claim is rejected back to the provider, the rejection typically includes a reason, but that reason doesn't always make it clear what to fix.

For example, a rejection might say that a procedure code is invalid without identifying which code failed on a claim with 20 service lines. Someone on the provider's billing team then has to review all 20 lines, find the error, correct it, and resubmit the claim. If the correction doesn't work, the cycle starts again.

But if the payer knows which validation failed, why not return that information with the rejection? Instead of "Invalid procedure code," we could say: "Service line 12: Procedure code XXXXX was not valid for the submitted date of service."

Having payer and provider billing teams go back and forth over these types of errors is expensive and inefficient for everyone.

AI can help process and reject claims faster. Providing more specific rejection reasons would also make errors easier to identify and correct the first time, reducing the work required on both sides.

Example Two: Exception handling review

Moving a claim from receipt to adjudication has historically required significant human review. When administrative teams consider AI, a natural question is, “How can I offload some of this review to an AI agent?”

At Amera, we try to start with a different question: Why does this review need to happen in the first place?

For example, incorrect eligibility information can be fuzzy-matched against existing records. Provider information can be validated against trusted sources. Claims and supporting documentation can be automatically classified and routed based on defined rules.

These are all examples of work that has historically required human intervention because of imperfect data or poorly integrated systems.

Some claims are genuinely complex and require human judgment. Those should be routed to a human reviewer. Other claims end up in manual review queues because of issues like mismatched member information or missing provider data that can be resolved automatically.

Our engineering team focuses on addressing the underlying data and workflow problems so that fewer claims require human intervention in the first place.

Example Three: Stop-loss reporting 

Traditional stop-loss workflows involve staff wrangling information across multiple parties and systems to support claim submissions. Staff need to pull claims reports, identify members approaching specific deductibles, collect eligibility and pharmacy data, assemble reports for carriers, respond to requests for additional information, and track reimbursements. 

All of this data collection represents significant manual labor for the TPA. 

AI offers plenty of opportunities to make each of those tasks faster. But again, we can also ask why each task exists. If the underlying claims, eligibility, pharmacy, and reinsurance data are already available, why does someone need to pull and combine those files every reporting period? 

For our stop-loss product, we connect to underlying claims and other data sources, use that information to identify reporting triggers, and automatically generate aggregate, specific, and trigger reports. This eliminates much of the manual work traditionally required to assemble reports and monitor claims approaching reporting thresholds.

Taking this one step further, we should connect directly with stop-loss carriers and exchange information programmatically. If both parties already have the data they need, why should anyone have to email reports and Excel files back and forth?

The opportunity is to eliminate much of the work involved in preparing and exchanging reports.

AI allows Us to Rethink Which Workflows We Actually Need

As Hank put it during our conversation:

"There are so many things that when you look at how healthcare has been constructed, that if you were starting with a blank sheet of paper, you just wouldn't draw it out that way."

Improving data quality, connecting disconnected systems, and making information more accessible can eliminate entire categories of administrative work. AI makes it possible to tackle these problems in ways that weren't practical before.

Some manual processes serve important purposes, including clinical judgment, regulatory requirements, and oversight. Others have become permanent workarounds for problems elsewhere in the system.

For new entrants, replacing a legacy point solution with an AI-native alternative isn't necessarily progress. We should ask why the workflow exists, preserve the parts that serve a purpose, and eliminate the rest.

Otherwise, we've just found a much faster way to produce the same garbage.