Build, Don’t Buy: AI Chart Review for Billing

We get asked all the time whether Distillemr will run billing chart reviews for our customers. Our answer is no, and we’d rather show you how to build it yourself. Here’s why, and here’s exactly how.

The question we get asked the most

“Can you also review our charts for billing?” It comes up in nearly every conversation we have with clinical and operations leaders. And we understand why. Billing accuracy is money on the table, and anything that promises to find missed charges sounds like an easy win.

Most vendors would happily say yes. It’s recurring revenue, it’s sticky, and it makes you dependent on them for something you do every single day. We see it differently. Billing chart review is one of the few things in this industry that is genuinely simple to automate. Because it’s simple, you shouldn’t be paying anyone a per-chart fee, a platform fee, or a percentage of collections to do it for you. That’s money you could keep in-house. And you certainly shouldn’t be handing a core function to an outside party.

So instead of selling it to you, we’re going to show you how to build it yourself. It costs us a potential line of business to say this out loud. We think it earns us something more valuable: your trust.

Billing review is the easy part

Billing follows rules. Codes map to documented elements, payers publish their guidelines, and the question “does this note support this level of service?” has a defensible, largely deterministic answer. That structure is exactly what modern language models are good at: take a chart, apply an explicit set of criteria, and return a consistent, explainable result.

This is not where the hard problems in healthcare live. Coding rules are written down. They don’t require years of clinical judgment to interpret, which is exactly why this work is handled by dedicated coding and billing specialists rather than clinicians, and why it’s a far more affordable function to staff. Once you’ve encoded your guidelines into a good prompt, an LLM can apply them across thousands of charts with a consistency no human team can match.

Why billing belongs in-house

There’s an old operating principle: if it’s a core function, don’t outsource it. Billing is about as core as it gets. It is mission-critical to your organization’s financial health, it follows guidelines specific to your payers and your contracts, and it changes as those guidelines change. When something is that central to how you survive, you want full control, full visibility, and no dependency on a third party who can raise prices or disappear.

There’s also a simple economics argument. Billing and coding review is work that’s well-suited to trained coding and revenue-cycle specialists, a cost-effective resource that most organizations already have or can readily staff. When you automate the first pass with an LLM and let your billing team validate and act on the output, you get speed, consistency, and cost efficiency all at once. Keeping it in-house isn’t the compromise. It’s the better operating model.

Where the hard work actually is

Here’s the part most vendors won’t tell you: the reason we don’t want your billing work is that it’s not where the real difficulty, or the real value, lies.

The art of medicine is messy. Clinical quality assurance isn’t a lookup table; it requires actual clinical expertise to judge whether the right decisions were made, whether a workup was appropriate, whether a diagnosis was reasonable given what was in front of the clinician. Upskilling the physicians, NPs, and PAs who deliver that care has to be done intentionally and thoughtfully. It demands scarce, expensive physician time, which is precisely why most vendors can’t do it well and quietly steer customers toward the easy, rules-based work instead.

That hard problem is our entire focus. Distillemr exists to build a strong foundation of clinical excellence. That foundation is what lets organizations scale more efficiently. It’s what lets them supervise APPs and early-career clinicians with confidence. And it’s what helps them mitigate safety risk before it becomes harm. That’s the niche we’re proud to own, and it’s why we’re glad to give the billing playbook away.

How to build it

The entire concept fits in one line:

You take the relevant text of a patient chart, combine it with a prompt that spells out your billing criteria, send both to a language model, and get back a structured assessment: the supported level of service, the documentation that justifies it, and any gaps that would put a claim at risk. That’s it. No platform required to prove the concept, and you can run your first review today.

The quality of the result comes almost entirely from the prompt. A good prompt names the coding framework you’re using, lays out the criteria for each level, tells the model exactly what to extract, and asks for its reasoning so a human can verify it. We’ve done the work of building one for outpatient primary care E/M coding, and we’re happy to give it away for free.

What goes into a solid prompt

You don’t have to take our word for it. Here are the parts that matter most, pulled straight from the prompt we use, so you can see how it’s built and start adapting it to your own payers and guidelines.

1. Define the role, and set a guardrail against upcoding. This is the single most important instruction. It keeps the model focused on accuracy and defensibility instead of maximizing revenue, which is what makes the output trustworthy.

You are a certified medical biller/coder with expertise in outpatient primary care E/M coding, CPT code selection, diagnosis-code accuracy, and documentation support.

Focus on billing and coding accuracy, not revenue maximization. Do not recommend a higher CPT code unless the documentation supports it. Do not suggest diagnosis changes solely to increase reimbursement.

2. Constrain the scope to this encounter. Without this, the model drifts into a full chart audit and flags unrelated chronic conditions or care gaps. You want a focused, billing-specific review of the visit in front of it.

Strictly limit feedback to the problems explicitly detailed in the HPI and managed in the Assessment & Plan for this specific encounter. Do not introduce feedback on unrelated chronic conditions, preventive care gaps, or problem-list items unless they directly affect the CPT/E/M level, diagnosis-code accuracy, medical necessity, documentation support, or safety of the plan.

The purpose is not a full chart audit. The purpose is a billing-focused encounter review.

3. Give it the coding map. Spell out exactly how CPT codes correspond to MDM complexity, and what factors push a visit up or down. The model is only as accurate as the rules you hand it.

For established outpatient visits:
  99212 = straightforward MDM
  99213 = low-complexity MDM
  99214 = moderate-complexity MDM
  99215 = high-complexity MDM

Factors that may support 99214/99204:
- Prescription drug management
- Ordering or reviewing multiple tests
- Ordering imaging when clinically relevant
- Chronic illness with exacerbation, progression, or side effects
- Acute illness with systemic symptoms
- Moderate-risk management decisions

Factors that may support 99213/99203:
- Stable, uncomplicated problem
- Limited data reviewed or ordered
- Low-risk management, no prescription drug management

4. Dictate the output structure. Asking for a consistent format makes results easy to scan, easy to validate, and easy to compare across charts.

Return feedback in three sections:
- Key Takeaways: 1-3 concise, high-impact recommendations
- Diagnosis Feedback: whether diagnoses are specific enough and
  support the services ordered
- Billing Feedback: whether the billed CPT code is supported,
  under/over/appropriately coded, and the MDM rationale

5. Set the tone. Collaborative, hedged language keeps the output constructive and reminds everyone that a human makes the final call.

Use careful, collaborative language:
"Based on the available documentation..."
"This appears more consistent with..."
"Consider reviewing whether..."
"This code may be difficult to support unless..."

Those five pieces are the backbone of a reliable billing review prompt. The full version we share adds detailed evaluation criteria, worked examples for each CPT level, and diagnosis-specificity guidance. Access it here and you can be running reviews the same day.

A sample chart to try

Use the sample chart below to test the prompt yourself. Paste the chart and the prompt into the LLM of your choice and compare what you get to our output.

What the output looks like

Here’s what we get back when we run the prompt against the sample chart: a clear recommendation, the documentation that supports it, and the gaps worth a second look.

The takeaway

Build your billing chart review in-house. It’s simple, it’s cheap, and it’s too important to outsource. Keep your billing team focused on validating and acting on the output, and keep control of a function your financial health depends on.

Then let us help with the part that’s genuinely hard: building clinical excellence at scale. That’s where physician expertise matters, where the stakes are highest, and where we’ve chosen to focus everything we do.

CPT® is a registered trademark of the American Medical Association, and any commercial or product use of the CPT code set requires a license from the AMA. Before you deploy this in production, review the AMA’s CPT licensing FAQs and its guidance on licensing CPT for AI, and obtain the appropriate license.