Vice President, Sales
AI accuracy gets the headline. How the system handles uncertainty decides whether the dollars arrive.
After 20 years working with healthcare organizations on revenue cycle solutions, I have learned to value one trait in people and technology alike: knowing when to raise your hand.
People I respect the most are not the ones who answer every question instantly. They recognize something unusual, stop, and ask questions. The people who most concern me are the opposite. They are confident about everything, including the things they have wrong. That confidence masquerades as competence right up until the consequences show up downstream.
I hold AI to the same standard.
When I talk to healthcare leaders, they often evaluate automation and are quick to quote accuracy rates. That’s understandable, but an equally important question gets far less scrutiny: “What does the system do when it isn't sure?”
Every automated system encounters ambiguity. Documentation is incomplete. Payer requirements change. A claim falls outside the patterns the system has seen. The question is whether the system recognizes uncertainty before it acts, and what it does next.
Everyone talks about accuracy. Ask about the remainder.
An impressive accuracy rate makes a good headline. For a revenue cycle leader, the more interesting question is “What kinds of cases are being missed?”
Does the model recognize when its confidence drops? Does it escalate those cases? If so, how? Or does it decide with the same apparent certainty it brings to straightforward work?
Aggregate accuracy alone cannot answer those questions.
The National Institute of Standards and Technology's AI Risk Management Framework recommends evaluating AI performance under conditions like those in which the technology will operate. NIST also calls for organizations to understand knowledge limits, the potential costs of errors, and the appropriate role for human oversight.
That distinction matters in the revenue cycle because not every error carries the same consequence.
An incorrect action repeated at machine scale can create considerable downstream rework, delay reimbursement, contribute to denials, or affect yield. The objective cannot simply be to automate as much work as possible. Organizations need to understand where automation can act reliably and where the consequences of being wrong warrant human-in-the-loop.
A well-designed system knows when to hesitate
Recognizing uncertainty is a capability in itself.
NIST notes that human judgment should inform the metrics and threshold values used to evaluate AI systems. It also recognizes that acceptable risk varies by organization, application, and use case.
For revenue cycle leaders, that leads to an important point: a confidence threshold is not simply a technical setting. It is a business decision.
A well-designed approach to uncertainty does three things.
1. It recognizes when confidence drops
The system needs to distinguish routine work from cases where the available information does not support a sufficiently reliable decision.
And that line does not necessarily belong in the same place for every transaction. Specialty, payer requirements, financial exposure, documentation complexity, and the consequences of an incorrect action can all influence how much certainty an organization should require.
2. It knows what happens next
Low confidence should trigger a defined path, not create a dead end.
Depending on the work, the next step could be another automated action, another source of information, or escalation to someone with the expertise to resolve the exception.
What matters is that the path has been designed intentionally. An organization should know what happens below the confidence threshold before thousands of transactions reach it.
3. It hands off the work with context
This is where I think many conversations about human-in-the-loop AI stop too early.
Sending something to a person is not enough. If that person must reconstruct everything the system already did, you have created a handoff, but you have not necessarily created an efficient one.
A useful escalation should:
Preserve what the system evaluated
Show what actions it already attempted
Explain why the case requires additional judgment
The person receiving it should be resolving the exception, not starting the work over.
The threshold is a financial dial
This is where the technical conversation becomes a financial one. Set the threshold too loose and errors can scale at machine speed. Set it too tight and work that could be automated keeps flowing to people, weakening the economics of automation.
Neither extreme is the goal.
The question shouldn’t be, “How much can we automate?” The better question is, “Which work can we automate reliably, and where does another layer of judgment optimize the outcome?”
That is why revenue cycle leaders should look beyond the percentage of work a system can automate. They also need to understand which work it automates, under what conditions, and with what tolerance for error.
The Peterson Health Technology Institute made a related point in its 2026 analysis of administrative AI in healthcare and found that applying AI on top of flawed administrative workflows can exacerbate underlying problems. Realizing AI's potential, the organization concluded, will require redesigning the processes on which the technology is deployed.
Before an organization determines where AI should act independently, it needs to understand the process surrounding that decision. Where does the work originate? What happens next? Where do exceptions occur? What do those exceptions cost? Where does expert judgment add value?
Those are operational questions before they are technology questions.
Six questions I would ask any AI vendor
When someone shows me an impressive AI accuracy number, I want to understand what sits behind it. I would ask:
How does your system measure its own confidence?
Who determines the confidence threshold, and what factors influence it?
What happens to work below that threshold, step by step?
What information does a reviewer receive when work is escalated?
How are escalations and reviewer decisions captured for future improvement?
How do you identify errors when the system was confident enough not to escalate?
That last question matters.
A system that recognizes uncertainty can route an exception. A system that is confidently wrong creates a different problem.
Knowing when to stop is part of knowing what to automate
AI will continue to take on more revenue cycle work. I don't think the measure of a mature operating model will be how rarely a person touches that work.
It will be how deliberately organizations decide when someone should.
Expert judgment belongs where ambiguity, exceptions, and consequential decisions require it. Automation should handle work that can be executed reliably at scale and route the remainder without losing context or momentum.
After 20 years around revenue cycle technology, what I want to know about an AI system isn't simply how often it gets the answer right.
I want to know what happens when it doesn't know the answer.
Because recognizing uncertainty is only the first step. Once the machine raises its hand, the next question is whether the right expertise is there to answer.
Next week, we’ll look at where human expertise adds value, what an effective handoff requires, and how those decisions can improve both the technology and the processes around it.
About Access Healthcare
Access Healthcare stands as one of India's largest and fastest-growing providers of healthcare business processes and technology solutions. Our team of over 33,000 professionals operates from 23 service delivery centers across three countries, emphasizing global delivery, workflow optimization, and our award-winning AI-enabled technology platform.
Since 2011, Access Healthcare has been a trusted partner to the US healthcare sector, leveraging domain expertise, technology, automation, and analytics to enhance clinical outcomes, financial performance, and operations for healthcare providers and payers.