A biller in one of the healthcare practices we work with hits an authorization gap. She spends forty minutes on it: a payer portal that times out, a modifier reworked, a claim resubmitted. When it's done, the record keeps two things. A denial code, and, weeks later, a payment. Everything between those two entries is gone.
That vanished middle is where the money goes. Our research and data, spread across hundreds of healthcare practices and revenue cycle management businesses, show that 75% of revenue-cycle complexity is here. Not in the software. Not in billing mechanics. Not even mostly with the payers everyone likes to blame. It sits in the operational work, and no system of record was built to see it.
None of this is a healthcare quirk. Every system of record, in every industry, keeps the outcome and drops the work that produced it. A CRM marks a deal lost and holds nothing about the six weeks that lost it. A dispatcher spends an afternoon rerouting trucks around a closed dock, and the transport system keeps only the late delivery and the penalty.
Now hold that next to the AI your vendors are selling. Almost every "AI for your industry" pitch is trained on one raw material: the record. The tidy rows a system keeps once the work is finished. That's the exhaust. Here's the part the demo skips over. Outputs don't contain their causes. A model trained on records can react to the outcome. It can draft the next document or flag a risky pattern. Useful. What it can't do is see why the bad outcome keeps happening, because the cause never made it into the record. In the healthcare practice, that's the denial that comes back every month. Automating the transaction doesn't reduce the complexity that produced it because the cause was never in the data.
Why does this matter now, when it's been true for years? Two things have arrived in short order. AI finally made it cheap to instrument the work while it happens, instead of reading the record after the fact. And the vendors turned the pooled record into their whole strategy. Healthcare shows it first, because the data is richest there. Prompt, a physical-therapy EMR, rebranded in 2025 into an "AI-powered clinic operations platform", and its AI now reads every note and CPT code across its customer base. Cedar describes the field's denial-prediction models as trained on historical claims and denial data. Ten years ago that data sat in a database doing nothing. Now it's a training corpus, and it isn't yours.
Which points at the actual problem, and it's gnarlier than it sounds. The traces are there, in a way. A Slack thread about the deal that slipped, an email about the denial code, the one-line reason a rep typed into the CRM after the loss. People do talk about why a deal died or why a claim bounced. But none of it is a usable record of the cause. It's scattered across a dozen tools and a few people's heads, none of it labeled, nothing tying this obstacle to that outcome. And the reason someone writes down is the one they settled on afterward. "Lost on price." "Auth required." A tidy label with three weeks of real causes buried under it. You can mine all of it and still not learn why the pattern repeats, because the record you'd need was never made.
For healthcare, we're building the Observatory for exactly that. It captures the work as it happens, at the point the complexity is made, logging the cause as structured data in seconds, without leaving the task: what the obstacle was, what it cost, where it sat in the sequence. Over time that becomes a dataset a system of record can't produce, because the record only ever shows up after the work is done.
Which changes how to hear the next AI pitch. Don't ask whether the vendor's model has seen data like yours. Assume it's seen all of it. Your system of record already holds everything you've ever filed in it. The harder question: was the cause of your problem ever recorded, by anyone? The forty minutes, the workaround, the real reason the claim bounced three times. None of it was written down while it happened. A model with full access to your data can't learn what your data never contained.
That's the first test. The second is the money. If the record can't hold your causes, why is it worth so much to the vendor? Because records at scale are a different asset. A million denial codes will train a model that predicts denials, drafts appeals, flags risk. Sellable, all of it. What a million denial codes won't do is cut the denials at your desk. Prediction reads outcomes; prevention needs causes, and the causes were never written down.
So your records move their number. Improvements in your margins happen when the causes start coming out of the work. And causes come out only when someone records them, in the moment, inside your own operation. That's the one investment no vendor can make for you. The payoff is better margins.
The more you invest and protect your own instrumentation of work and the data that comes from it, the better you can guide a vendor’s AI, with open source models becoming good, you may never have to train a vendor at your cost. Make AI work for you, get better margins, and use the custom-trained AI to attract acquirers that will pay better multiples.