Data Report · RedactSure Research
Chris Sowa · Founder, RedactSure
Published
Last updated
Where Does PHI Hide When You Automate Medical Coding?
Lessons from revenue cycle leaders on AI, automation and protected health information
In brief
-
In revenue cycle automation, the hardest part is mapping every manual step and the PHI attached to it.
-
Many organizations keep frontline staff away from AI entirely because the risk of PHI leaving is too high.
-
The main risk is ignorance, not malice: people paste charts into public chatbots without realizing what they send.
-
Many useful AI questions can be answered without any PHI, such as whether a code has been used at a site in the last 90 days.
-
Screen recordings made to train automation can become a second copy of every patient detail on the screen.
-
The minimum necessary standard is a good design rule for AI: give the model what the task requires and nothing more.
Medical coding is one of the most obvious places to use AI in healthcare. The work is high volume, rules based and expensive to staff. It is also full of protected health information on every screen. Over the past few months I have worked closely with leaders at a healthcare services company that runs revenue cycle and coding work for providers, and talked with a strategy executive at a health system about where the risk sits. Their experience shows where PHI hides when you automate, and how to design around it.
Why coding teams build their own AI
A technology executive at the company described how his team got into AI:
“We started identifying areas within the company that we wanted to first try to automate, and then we decided that we're doing some things differently from some of the other companies in RCM. So we started building some of our own solutions, just because we couldn't find solutions that met all of our requirements off the shelf or in the market.”
Technology executive, healthcare services company
An AI engineer on his team said medical coding is the main focus today, and that the hardest problem is interoperability, with security and compliance on top of it.
The hardest part is finding the PHI
When we talked about where automation gets difficult, the technology executive did not point to the AI model. He pointed to the process.
“That's the most difficult part of the whole RPA automation piece: identifying all the various steps in the manual process and all the PHI that's associated with it.”
Technology executive, healthcare services company
Every step a coder takes, every screen they open and every field they read is a place where PHI can enter an automation. A script or an agent that replays the process inherits all of it unless someone decides, step by step, what it needs.
Who gets AI, and who does not
The same company keeps general operations staff away from AI and concentrates its AI work with a senior team on defined problems: coding, accounts receivable collections and large-scale reporting such as net collections rate. The reason is the risk of data leaving, not doubt about the value.
“Would that be able to help them in their day-to-day reporting and summarizing and making them more efficient? Most likely it would. But the risks are too high for us in that area.”
Technology executive, healthcare services company
Ignorance is the bigger risk
Asked where he sees the real danger across the industry, he did not describe attackers.
“It seems to me that ignorance is the biggest problem in this space, and especially in healthcare. There are professional organizations that do their own medical coding, and I'm sure people are just putting […] patient charts into Claude and just trying to get them coded and telling them what codes they are, not realizing that they're submitting all kinds of private information to any one of these AI out there.”
Technology executive, healthcare services company
Well-meaning staff trying to get through a backlog are the most likely source of a PHI disclosure to an AI service. Training helps. A safe route that does the same job helps more.
Many useful questions need no PHI at all
The AI engineer described a use that shows how much value is available without patient identifiers. He wants coders and managers to be able to ask the coding data direct questions.
“If we give the users a bit more of a chatbot scenario where they could just query the DB and be like, hey, for this center, have we coded Z09 on something, on any charts within the last 90 days? And then they could pull that up without any sort of PHI.”
AI engineer, healthcare services company
The technology executive added a second idea: if identifiers can be removed from clinical notes, the team could test how different AI models code the same chart.
“If they're able to redact sensitive information in a surgical note, we could easily do comparisons on coding results, or at least create maybe baselines and compare to what ours is doing by our training versus what is available.”
Technology executive, healthcare services company
Recordings can become a second copy
Some outsourcing firms record what frontline staff do on screen, partly to measure work and partly to automate it later. That is a reasonable goal, but every recording also captures the patient names, member IDs and diagnoses that were on the screen. If the recordings are kept, shared with a vendor or used to train a model, the organization has created a second store of PHI. Recordings made in a tokenized environment avoid that problem, because the sensitive values were never on the captured screen in the first place.
The weak point is often the handoff
A strategy executive at a health system offered a useful outside view. Large health systems, he said, tend to secure their own environments well. The exposure grows at the handoffs to vendors and subcontractors, such as revenue cycle firms and the AI tools now arriving in areas like radiology. He also pointed out that smaller independent providers, such as freestanding emergency departments and urgent care groups, usually have the thinnest security teams and the most manual billing work.
Minimum necessary as an AI design rule
HIPAA's minimum necessary standard asks covered entities to make reasonable efforts to limit the use and disclosure of PHI to the minimum necessary for the intended purpose (U.S. Department of Health and Human Services, n.d.). It translates directly into AI design. For each step in a coding or billing workflow, decide what the model actually needs to see. In most steps, the model can work on placeholders for names, dates of birth and member numbers, and the real values are needed only at the moment a person or system writes them back.
Where RedactSure fits
RedactSure applies that rule at the screen. Patient identifiers become tokens such as PATIENT_001 before the model reads the page, so an agent can work through a coding or billing workflow without the real values entering the model. The design is aligned with PHI data-minimization requirements. A named person approves consequential actions, and every step is recorded. RedactSure is in pilot with organizations in regulated industries.
Frequently asked questions
Can AI do medical coding without seeing PHI?
Much of it, yes. Many coding steps can work on documentation with identifiers replaced by tokens, and many analytical questions, such as code frequency by site, need no identifiers at all. Real values are needed only where a record is written back.
What is the biggest PHI risk with AI in revenue cycle work?
In the experience of the leaders quoted here, it is ignorance: staff pasting charts into public AI tools to speed up coding, without realizing they are sending private information to a third party.
How does the minimum necessary standard apply to AI?
Treat the AI like any other user of PHI. For each task, give it the minimum information it needs, and keep identifiers out of the model wherever the task can be done with placeholders.