AI is in your billing department now
Over the past 18 months, AI-powered billing tools have moved from pilot programs to production use in thousands of outpatient practices. Your billing company is probably using one. Your EHR vendor is probably selling one. And if you have talked to a healthcare IT vendor recently, someone has pitched you one.
Some of these tools are genuinely useful. They catch errors that humans miss, they speed up processes that used to take hours, and they reduce denial rates in measurable ways.
Others create new problems. Compliance risks from automated coding decisions nobody reviews. Integration failures with legacy systems. Promises of savings that do not materialize once you account for the subscription cost and the implementation time.
Here is what we are actually seeing across the practices we work with.
What is available right now
The AI billing tool market has consolidated around five main categories. Each one addresses a different part of the revenue cycle.
AI-assisted coding. These tools read clinical documentation (provider notes, procedure records) and suggest CPT and ICD-10 codes. The provider or coder reviews the suggestion and accepts, modifies, or rejects it. The better tools also flag documentation gaps that would result in a denial.
Automated claim scrubbing. Before a claim is submitted to a payer, the AI checks it against payer-specific rules, coding guidelines, and historical denial patterns. It flags claims that are likely to be denied and tells you why. Think of it as a pre-submission quality check.
Denial prediction. Using historical data from your practice and industry-wide patterns, these tools predict which claims are most likely to be denied and why. The idea is to fix problems before submission rather than chasing denials after the fact.
Eligibility verification. AI-powered tools that check patient insurance eligibility in real time, flag coverage gaps, estimate patient responsibility, and identify coordination-of-benefits issues before the patient is seen.
Patient payment estimation. Tools that use benefit data and historical payment patterns to generate accurate out-of-pocket cost estimates for patients before their visit. Better estimates mean fewer surprise bills and higher point-of-service collections.
What actually works well
Claim scrubbing
This is the category where AI delivers the most consistent, measurable value. A good claim scrubber catches coding errors, missing modifiers, diagnosis-procedure mismatches, and payer-specific edits before the claim goes out the door.
The financial impact is straightforward. Every claim that gets denied costs you money to rework. The average cost of reworking a denied claim is $25-30 in staff time. If your practice submits 2,000 claims per month and has a 10% denial rate, that is 200 denials per month, costing $5,000-$6,000 in rework. If AI scrubbing prevents even half of those denials, you save $2,500-$3,000 per month in rework costs plus you collect the money faster.
A dermatology group we work with (anonymized Sorso client, Southwest, four locations, ~$5M revenue) implemented AI claim scrubbing last year. Their first-pass claim acceptance rate went from 88% to 94%. On 3,200 monthly claims, that is 192 fewer denials per month. At $28 per denial rework, that is $5,376 per month in saved rework costs. The tool costs $1,200 per month. The net savings is about $50,000 per year.
That is a clean ROI, and it is the kind of result we see consistently with claim scrubbing tools.
Eligibility verification automation
Checking insurance eligibility used to require a phone call or a manual portal lookup for every patient. AI-powered eligibility tools run batch checks overnight, flag patients with coverage issues, and identify coordination-of-benefits problems before the patient arrives.
The financial impact here is about preventing claims that will be denied for eligibility reasons. Eligibility denials are the most common denial category in most practices, representing 20-30% of all denials. They are also the most preventable.
A multi-location urgent care group (anonymized Sorso client, Midwest, five locations) reduced eligibility-related denials by 60% after implementing automated verification. On their volume, that translated to $4,200 per month in recovered revenue that would have been denied and potentially never reworked.
Coding suggestions for documentation
AI coding assistants that read provider notes and suggest codes are genuinely useful for one specific thing: speed. A coder who normally processes 20 charts per hour can process 30-35 with AI suggestions. The AI does the first pass, the coder reviews and adjusts. This is not replacing the coder. It is making the coder faster.
For practices that have a coding bottleneck (charts sitting in a queue for 3-5 days before being coded and billed), this speed improvement means claims get submitted faster, which means money arrives sooner. If your average lag from date of service to claim submission drops from 5 days to 2 days, you accelerate your entire revenue cycle.
What does not work well yet
Complex coding decisions
AI is good at straightforward coding. A 99213 office visit with a clear diagnosis and a standard treatment plan. The AI reads the note, suggests the code, and it is right 90-95% of the time.
Where it falls apart is nuance. E/M level selection when the documentation is borderline between a 99214 and a 99215. Modifier usage when multiple procedures are performed in the same session. Time-based coding where the documentation supports it but the AI does not pick up on the relevant time entries. Incident-to billing rules. Split-shared visit documentation.
These decisions require understanding payer contracts, local coverage determinations, and the specific documentation habits of each provider. Current AI tools do not have that context, and when they get it wrong, the consequences are real.
An AI that consistently up-codes (selects a higher-level code than the documentation supports) creates a compliance risk. An AI that consistently down-codes (selects a lower-level code) leaves money on the table. Both happen.
We reviewed one practice (anonymized Sorso client, primary care, two locations) where the AI coding tool was suggesting 99215 for visits that clearly should have been 99214 based on the documentation. The coder was accepting the suggestions without adequate review because "the AI said so." When we audited a sample of 50 charts, 12 were up-coded. That is a 24% error rate on high-value codes. If an auditor or a payer reviews those charts, the practice faces recoupment demands and potentially a fraud investigation.
Appeal letter generation
Several tools offer AI-generated appeal letters for denied claims. The idea sounds good: the AI reads the denial reason, pulls relevant documentation, and drafts an appeal.
In practice, the letters are generic. They hit the right talking points but lack the specificity that makes appeals successful. A good appeal references the exact clinical documentation, cites the specific payer policy, and makes a precise argument for why the denial should be overturned. Current AI tools produce letters that read like templates with patient data inserted.
We have not seen compelling evidence that AI-generated appeals perform better than appeals written by experienced billing staff. They are faster to produce, but speed does not matter if the appeal gets denied.
Payer-specific rules
Every payer has different requirements for prior auth, different documentation standards, different modifier rules, and different processing quirks. Blue Cross Blue Shield of Texas has different rules than Blue Cross Blue Shield of Illinois, even though they share a name.
AI tools trained on general coding guidelines do not capture payer-specific variations well. They know that a 59 modifier indicates a distinct procedural service, but they do not know that Payer X requires an XE modifier instead, or that Payer Y does not recognize the 59 modifier at all for that CPT code pair.
This is improving, but it is still a meaningful gap. If your practice deals with 10-15 different payers, the payer-specific rules represent a significant portion of your coding and billing decisions.
The financial case, honestly
Let me lay out realistic numbers for a practice doing $3M in annual revenue.
Potential savings from AI billing tools:
- Reduced denials from claim scrubbing: $25,000-$40,000/year
- Faster coding and submission: $10,000-$20,000/year in accelerated revenue
- Reduced eligibility denials: $15,000-$25,000/year
- Administrative time savings: $10,000-$15,000/year
Total potential benefit: $60,000-$100,000/year.
Costs:
- Software subscriptions: $12,000-$36,000/year (varies widely)
- Implementation and training: $5,000-$15,000 one-time
- Ongoing management and configuration: $3,000-$8,000/year
Total cost: $15,000-$44,000/year (implementation cost is one-time; ongoing is $15K-$44K after year one).
Realistic net benefit: $16,000-$85,000/year for a $3M practice, depending on your starting point and which tools you adopt. Most practices land in the $25K-$50K range.
That is meaningful, but it is not transformational. Worth doing, but not the silver bullet that some vendors promise.
The practices that see the highest ROI are the ones with high denial rates (above 10%), slow claim submission workflows, and manual eligibility verification. If your denial rate is already at 5% and your claims go out within 48 hours, the AI tools have less room to improve.
What your CFO should watch
If you implement AI billing tools, here is what to monitor.
Audit the AI's coding decisions. Randomly sample 20-30 charts per month and compare the AI's suggested codes against what a senior coder would select. Track the agreement rate and the direction of disagreement (up-coding vs down-coding). If the AI consistently disagrees with your best coder, one of them is wrong. Figure out which one.
Track your denial rate before and after. This is the clearest measure of whether the tool is working. Pull your denial rate by category for the six months before implementation and compare to the six months after. If denial rates have not dropped measurably, the tool is not delivering.
Watch for compliance creep. AI tools that suggest codes can subtly shift your coding patterns over time. If your average E/M level distribution changes after implementing an AI coder, investigate. A sudden increase in the percentage of 99215s billed is a red flag, whether a human or an AI is making the coding decision.
Do not accept black-box decisions. If the AI suggests a code, you should be able to see why. What documentation did it use? What coding rule did it apply? If the tool cannot explain its reasoning, you cannot defend the code in an audit. "The computer told us to" is not a defense.
Measure the real cost. The subscription fee is just part of it. Account for implementation time, training time, workflow disruption during the transition, and ongoing management. If the true all-in cost is $40,000 per year and the measurable savings is $45,000, the ROI is real but thin. Make sure you are measuring against actual results, not the vendor's projections.
My take
The tools are useful. The best ones, particularly claim scrubbing and eligibility verification, deliver measurable financial value. They belong in most practice workflows.
But they do not replace a human who understands your payer contracts, your provider documentation patterns, and your billing team's strengths and weaknesses. The AI catches patterns. A good billing manager catches context. A good CFO catches the financial impact.
The biggest risk I see is practices buying these tools and reducing billing oversight because "the AI handles it." The AI handles the routine work. The exceptions, the complex cases, the payer-specific quirks, and the compliance monitoring still require experienced people watching the numbers.
Use the tools. Track the results. Keep your humans.
If you want help evaluating whether AI billing tools make financial sense for your practice, take the free assessment. We will look at your denial patterns, submission timelines, and billing costs to see where automation would actually move the needle.
