AI in the Office: Putting AI to Work as Your Hardest‑Working Clerk, Not Your Junior Doctor

By Mikel D. Daniels, DPM, MBA, President and Chief Medical Officer, WeTreatFeet Podiatry

I’m often asked about the operational benefits of using AI in clinical practice, and how to use it without feeling like a robot, has quietly moved into the exam room. Honestly, I wrestle with the same concern. Sometimes it feels like I may have created a monster. But as I learn to use this technology, I’ve realized that I don’t need (or want) AI acting like a junior doctor inside my electronic health record (EHR). What I want is an incredibly fast and efficient clerk that helps me keep my documentation accurate and my claims clean. My goal is simple: spend more of my time and mental energy practicing medicine, and less of it chasing modifiers, correcting paperwork, and fighting denial codes.

In my own practice, the EHR already feels like a third person in the exam room. As I’ve written before, the chart is no longer a quiet prop. It has become an active character, constantly demanding my attention with alerts, required fields, and billing prompts. The result is that many of us spend more time looking at the computer than engaging with the patient. Now AI is sliding into the room alongside the EHR, raising a simple but important question: Will it become a fourth wheel competing for my attention, or a quiet ally that handles the tedious work. When I’ve used ambient AI tools in my practice, I’ve only felt comfortable when the technology stays firmly in the “clerk” lane. It can listen, capture, organize, and surface important details, but it must hand everything back to the clinician, coder, or biller who decides what is clinically meaningful and what ultimately gets signed. Once we begin viewing AI as a tireless, highly efficient clerk rather than a virtual physician assistant, I think much of the anxiety surrounding it will fade.

Where AI belongs in the practice, but not at the bedside

In most practices, the safest and most useful place for AI right now is for clinical-adjacent work. This includes billing, coding, documentation quality, revenue cycle, and administration activities. These processes live on top of data we already have collected. EHR notes, audio from visits, claims history, and denial patterns are created just by the nature of having a visit. What should AI not do? While opinions vary, I think of things like not diagnosing heart failure. AI should work on catching up to the paperwork the practice is already drowning in. Think of it this way, if AI touches data and workflows, that is ok. The humans still own diagnosis, treatment decisions, and the final record. Everyone is happy.

I look at it in tiers. At the front door, AI can summarize patient intake forms into something a human can read in under a minute (instead of seven). In the back office, it can comb through years of claims and denials, surface patterns, and help a revenue cycle manager (RCM) see what they already suspect, but have not been able to quantify. In between, during, or after the visit, ambient tools can capture the full conversation and build a draft note that a physician edits, signs, and owns. At no point does the AI decide the diagnosis, or whether the patient goes to surgery. That distinction matters. It keeps AI as a virtual physician’s clerk in the administrative sense, not a junior doctor practicing medicine without a license. There are absolutely situations in which AI has been providing the care and making the medical decision. Think I am exaggerating? Nope, check out https://www.healthlawadvisor.com/emilie-is-not-a-psychiatrist-pennsylvania-board-of-medicine-alleges-unlawful-practice-of-medicine-by-an-ai-chatbot

Billing and coding: AI as a second coder with a good memory

If you look at where most practices bleed revenue, it’s not some mysterious dark corner. It’s undercoded visits, missed procedures, sloppy modifier use, and documentation that doesn’t quite support the level of service we actually delivered. Coders and billers know this. They live it every day. What AI is good at in this lane is pattern recognition on structured text. That is very different from making a diagnosis in the room.

Here is how it looks in my practice. I finish a note the way I normally would. The AI reads it and proposes ICD-10 and CPT codes that fit the documented problems and the work I actually did. It might say, “You documented diabetic neuropathy and loss of protective sensation; do you want to pair this EM code with the appropriate G-code that supports coverage and risk?” It will also flag where the documentation doesn’t support the level of service I’m trying to bill. If I clicked a 99214 but never clearly showed moderate complexity in the note, the AI will point that out. This is not the machine telling me I can’t bill what I did. It’s the machine reminding me that, under Medicare’s rules, if it isn’t documented, it didn’t happen, and you can’t bill for it.

Over time, AI can analyze years of claims data to identify patterns of consistent undercoding. This will reveal how much revenue may have been lost because of documentation habits or an incomplete understanding of coding complexity. For example, I’ve seen practices discover that they almost never attach the correct diabetes complication codes to foot procedures. The result is that the payers treat everything as routine care and deny it. AI can surface that pattern as soon as you feed it the data. A human would take months to piece it together. The key here is what you do with it. In my conversations with colleagues, I always recommend starting with AI as an auditor. Run it parallel with your existing coding. Let it catch the undercoding and documentation gaps. Then decide where to change your templates or workflows. That feels a lot safer than “starting Monday, the AI will code everything.”

Clean claims, not pre‑claim gatekeeping

One thing I’ve learned the hard way is that pre-claim AI gatekeeping can feel punitive. If the system starts blocking claims until the AI is satisfied, you might see a physicians’ revolt, and frankly I don’t blame them. I personally believe in constant improvement, but not everyone does. This model works better when using AI to improve claim quality over time, not as a gate on each individual encounter. Think of it this way, we can let an AI system review outgoing claims in bulk, and flag which ones look structurally similar to prior denials, or which payers are consistently pushing back. That information then feeds education and template changes, not a new layer of “you may not bill this.”

In my own practice management thinking, AI is most valuable when it makes the billing team more productive, and gives them better tools to do what they already do, (instead of threatening to replace them). I’ve seen offices use AI to run weekly “clean claim” checks that catch missing modifiers, absent time statements on prolonged services, or forgotten links between procedures and qualifying diagnoses. Over several months, their denial rate drops and cash flow smooths out, not because the AI took over, but because the humans adjusted how they document and submit. This is exactly the kind of change that lets a practice expand without adding another full-time biller. The same staff simply gets more leverage.

Documentation improvement: making what you actually did visible

Documentation burden is the part of this conversation that clicks fastest with physicians. Most of us feel like we spend too much time writing notes that are too long, too repetitive, and mostly aimed at auditors and payers. In the diabetic foot world, I’ve written about how accurate documentation is the backbone of sustainable care, because if the record doesn’t capture the patient’s real complexity, the coding and reimbursement lag behind what we’re actually dealing with. Again, it is the old adage, if it isn’t in the note, it didn’t happen.

AI can help here, but only if it is pointed at quality, not at upcoding. Ambient tools that listen to the visit and draft a structured note already shift the balance a bit, instead of the provider typing while the patient talks. This allows the provider to focus on the exam and the conversation, not button clicking. Later, you review the draft, fix the odd phrasing, and make sure the assessment and plan match what you actually decided. What I’ve seen in practice, is that this kind of capture often picks up small but important details that shorthand notes miss. For example, a subtle change in how a patient describes pain, or a comment about recent falls, that matters for risk and follow-up.

On the back end, AI-driven clinical documentation improvement (CDI) tools can read through notes and flag missing elements for a given level of service, conflicting statements across the chart, or identify vague terms that don’t match coding rules. I’ve had systems prompt, “You mention neuropathy but never document a sensory exam; do you want to clarify this?” That prompt is annoying the first time and helpful the second. The important part is control. If the suggestion feels wrong or pushy, I ignore it. I am still the one deciding whether the note is accurate and honest. When practices roll this out, I recommend three things: 1. Make it opt-in by default, 2. Tune it by specialty, and 3. Give clinicians veto power. When people know they can say no to a suggestion without penalty, they’re more willing to try the tool.

Revenue cycle and denial management: spreadsheet on steroids

Revenue cycle teams already live in spreadsheets and dashboards. Denial codes, payer mix, aging A/R, you name it. AI sits very naturally there as a pattern spotter. If you hand a few years of 835/837 data to a decent AI pipeline, it will analyze the data, and what surfaces are things your RCM staff sort of knows in their gut but couldn’t prove at scale. In one case I am aware of, an AI analysis showed that a single payer was denying a disproportionate share of claims tied to a specific code, even on audit. Turns out it was because the documentation consistently missed one required phrase. The team knew that payer was “difficult”, but AI gave them the proverbial smoking gun. They changed the template, added a prompt, and a big chunk of those denials disappeared within a quarter, while passing post payment audits.

This is where AI feels the least controversial to physicians. It’s not touching individual patient decisions. It’s looking at the fabric of the business and telling you where the friction lives. Vendors are already baking these kind of analytics into RCM packages. I’m fine with that as long as the practice owns the decisions about what to fix and how. My guidance to colleagues is simple. Ask your RCM vendor, or your internal data people, questions, but not, “Can AI solve denials?” Instead, “Can it show us exactly which combinations of payer, code, and documentation language are costing us the most.” That is a much more actionable question.

Admin and front office: the invisible last mile

There’s a whole layer of practice life that clinicians barely see, but can make or break the day. This includes intake, scheduling, messages, referrals, prior auth paperwork, and internal reporting. Nothing to do with the practice of medicine (and we wonder why healthcare costs so much). I’ve watched front desks and call centers drown in portal messages, refill requests, and, “I just have a quick question” calls. AI can help here without ever getting near the exam room.

Smart intake can turn five pages of patient-completed forms into a one-page snapshot. Think of a one-page summary of the key problems, meds, allergies, prior surgeries, and a short risk summary, even before walking into the treatment room. Triage tools can sort inbound messages into the right chart instead of asking one person to read every portal note and email. Referral and prior auth tools can turn messy free text and scanned PDFs into structured forms that staff can submit without retyping anything. Internal reporting tools can generate clear dashboards on visit volumes, no-show rates, payer mix, and productivity, so practice leaders aren’t manually building reports at 10 PM.

None of this is glamorous. It’s not going to end up in a glossy brochure about AI saving healthcare. But in a practice that doesn’t want AI “in the room,” these back-office use cases are often the first wins. They soak up staff time in a way many physicians don’t appreciate until the systems improve. When that work becomes smoother, the same staff can support more clinicians or more locations without burning out. That’s how practices use AI to expand without growing headcount linearly.

Governance: the boring part that keeps you out of trouble

If you’re going to put AI into your workflow, you need a basic governance policy. This isn’t something I’d recommend writing on a napkin. Sit down with your legal counsel and your compliance folks and make it part of your larger documentation and technology policy. In my own thinking, a solid AI governance document for a practice should include the following:

1. It should clearly state that AI does not make the diagnosis or treatment choice. It may summarize, propose codes, or suggest wording, but a licensed human clinician or coder is responsible for the final decisions and the signed record.
2. It should define where AI is used in the workflow and how that use is documented. For instance: “Ambient listening used to generate draft notes” or “AI coding suggestions reviewed by certified coders before claims submission.”
3. It should spell out an error-handling process. If the AI gets something wrong, who catches it, how is it corrected, and how is that correction auditable if an auditor or patient questions the record later.
4. It should require human-in-the-loop review for any AI-generated content that enters the official record or goes out as a claim. No fully autonomous claims, no notes signed without human review.
5. It should address patient communication and consent. If visits are being recorded or if data is being processed by an AI system, patients should be informed in plain language, with an opt-out process where possible that doesn’t penalize them for saying no.

I’ve found that once practices nail down this governance piece, clinicians relax a bit. They know the machine is not secretly practicing medicine or quietly coding behind their backs. They know there is a human gate at the end of the process. And they know that if something goes sideways, there is a documented way to fix it.

Staffing, productivity, and practice growth

One myth I try to kill early when I talk with colleagues about AI is the idea that it will let them fire half their staff. In my experience, that is not how this plays out. What actually happens is more subtle and more useful. The same staff becomes more productive, less buried in repetitive tasks, and more available for higher-level work. That, in turn, lets the practice expand without adding full-time employees at the same rate.

If you can reduce denials, smooth intake, and shorten documentation cycles, you often can add clinic sessions, open a satellite location, or increase patient volume within the same staffing footprint. The MA who used to spend an hour a day chasing missing forms can now help with risk-based recalls. The biller who used to manually rework every denial can spend their time on higher-value appeals. The physician who used to spend two hours every night finishing notes can reallocate some of that time to seeing an extra patient, teaching, or frankly just resting. That kind of quiet productivity change matters more than the fantasy of AI cutting staffing costs in half.

How I’d tell this to a practice owner

If I were sitting with a practice owner in my office, I wouldn’t start with diagrams or buzzwords. I’d say something like this: You already have a third party in every exam room, the EHR. AI is about to join it. You have a choice. You can let AI become another source of noise, or you can deliberately assign it the clerk job it’s actually good at. Let it listen, extract, classify, and organize. Then have your clinicians, coders, and managers review and approve. Keep the clinical decisions and the final record in human hands. Put the policy in writing. Then, start with two or three very boring, very practical use cases. Think cleaner claims, better documentation, and calmer front-office workflows.

From there, watch what happens to your denial rates, your documentation time, and your staff stress levels over six to twelve months. If those move in the right direction, then you can decide how much more ambient listening or AI-driven coding support you want. If they don’t, you can dial it back or change tools without ever having put AI in charge of patient care.

Dr. Mikel Daniels is a board-certified podiatrist and healthcare executive with more than two decades of experience in foot and ankle surgery, wound care, and medical economics. As President and Chief Medical Officer of WeTreatFeet Podiatry, he has grown the practice from one office into a regional network of surgical centers, and retail health services across Maryland, Pennsylvania, and Washington, D.C.

Dr. Daniels earned his Doctor of Podiatric Medicine from Temple University and an MBA in Healthcare Administration, combining clinical expertise with business strategy to deliver efficient, patient-centered care. His work focuses on complex reconstructive procedures, diabetic limb salvage, sports injuries, and minimally invasive techniques designed to accelerate recovery.

A Fellow of the American College of Foot and Ankle Surgeons and the American Professional Wound Care Association, Dr. Daniels also consults for biomedical technology firms and serves as a principal investigator in clinical research. His insights have appeared in Forbes, Parade Magazine, and CNN, and through his writing and mentorship, he continues to advance innovation.