A small pediatric practice had been losing revenue for months and could not see why. Patient volume was steady, the staff was working as hard as ever, and every report from the practice management system looked reasonable on its own. The owner suspected everything from billing errors to embezzlement.
The answer was in accounts receivable, and finding it took hours, not weeks. We ran AI-assisted analysis against exports from the practice's billing system and reconciled the whole receivables file at once. It held $123,000 in aged receivables that split into four separate problems. None of them showed in the standard reports, because those reports never segmented the full file.
Why revenue falls when patient volume holds steady
When volume holds and cash drops, the money is usually stuck between the visit and the payment, and summary reports hide where. The practice's reports grouped receivables by age and by payer. Grouped that way, $123K of different problems looked like one lump of "slow payment."
The practice was not missing data. It was missing the reconciliation that turns data into a diagnosis. Checking every claim against payer status, hold codes, and filing deadlines by hand is weeks of work. Nobody had those weeks, so the work never happened.
What was inside the $123K
The $123K was not one problem. Segmented by payer and by failure mode, it split into four categories, each with a different fix and a different clock.
Claims blocked by enrollment. Enrollment status with specific payers stopped some claims upstream, so they never entered adjudication. Payers can require providers to be enrolled before they pay. Medicaid, for example, requires ordering and referring providers to enroll as participating providers.
Claims held before submission. Some claims sat in the billing platform's own pre-submission hold queues, stopped by scrub rules. Everyone believed they were "out for payment." They had never reached a payer.
Claims near the timely filing limit. A timely filing limit is the deadline a payer sets for submitting a claim after the date of service. Federal Medicaid rules, for example, set it at 12 months from the date of service. Past that date, a receivable quietly becomes a write-off.
Ordinary collectible AR. The rest was normal receivables that needed someone to work them.
What did the AI actually do?
AI-assisted analysis made a weeks-long reconciliation cheap enough to happen. That was its whole contribution, and it is the one that changed the outcome.
Matching thousands of claim records against payer status, hold-queue codes, and filing windows is exhaustive, mechanical work with little judgment in it. That is the kind of work AI speeds up. The diagnosis needed no new information. It needed the existing information processed at a cost the practice could afford. Until then, that cost was the binding constraint.
The AI did not decide which questions mattered. It did not decide that enrollment-blocked claims needed a different escalation path than scrubber holds. It did not sit with the owner to rank what to chase first against the filing calendar. People did that, informed by domain research. And the AI did not "save the practice." It found the problem fast. The fix is operational work: enrollment follow-ups, hold-queue releases, resubmissions.
We hold that line on purpose, because stories like this one inflate in the retelling. The claim we stand behind is AI-accelerated diagnosis: structure that stayed invisible for months, surfaced in hours. Anyone selling AI-accelerated rescue is skipping the part where people do the work the diagnosis prescribes.
Where the same problem hides in other businesses
Most businesses we look at have a version of this: a problem you can feel in the outcomes but cannot see in the reports. The reports summarize along dimensions someone chose years ago. The problem lives along a dimension nobody chose, and finding it takes a reconciliation of the full data set that nobody could budget weeks for.
That reconciliation used to be a consulting project. With the right person driving AI-assisted analysis against your real exports, it is now often a matter of hours. Most organizations have not updated their sense of what a diagnosis costs. What an AI readiness assessment actually tells you covers the structured version of the same move.
The diagnosis is the cheap part now. If your numbers say something is wrong and your reports say everything is fine, our AI for healthcare team can usually tell you in one conversation whether your problem is findable the same way. The AI readiness assessment is where that work starts.