TL;DR
- AI models have become much cheaper and can now stay on a task for hours. Most AI systems in mid-market companies still do the narrow job they were scoped for two years ago.
- The limits were set to fit a weaker, costlier model and a six-week project, and a newer model gets the same limits.
- Cost controls make it worse. Companies cut spending on models, which is the cost that’s falling fastest, and leave alone the people who check and finish the work.
- Time saved only becomes EBITDA when someone uses it to take on more work or recover money the company was losing.
- Someone has to keep asking what the system can take on next. In most companies nobody owns that question.
Where Moative fits
Moative works as an AI operating partner with mid-market companies in vertical industries. Our forward deployed teams don’t hand over a system and leave. They stay with the work after the first launch. Each time the models improve, we test them against the client’s hardest cases and give the system a bigger job when the results hold up. When part of the work settles into a simple, repeatable answer, we turn it into a plain rule so the client stops paying a model to work it out again. Because we work with several companies in the same industry, what we learn at one client shortens the work at the next.
If you have an AI system that still does the same job it did the day it went live, we’re glad to look at it with you and give you a straight opinion on what today’s models could take on. Get in touch.
A few weeks ago, I was at a meeting with the leadership team of a revenue cycle company, which handles billing and collections for a few dozen practices. Their operations head showed me an AI tool that they put in about two years ago. Every payment that comes back from an insurer goes through the tool. When a claim is denied or underpaid, the tool reads the paperwork and records the reason.
I asked what happens next. The large claims go to a team that works them by hand, and the team pulls the payer contract and the patient’s record so it can write an appeal. The small claims fall below a limit that the company set years ago. Most of them are written off, because chasing them would cost more than the company can hope to recover.
I also asked which AI model the tool uses. Nobody in the room knew. The vendor had upgraded the model twice, and each upgrade made the recorded reasons a little more accurate. Otherwise, the tool does the same job that it did on the first day.
AI models underneath have changed a lot in two years. Epoch, a research group that tracks AI prices, estimates that the same level of capability gets more than 90 percent cheaper every year. The models can also stay on one task for hours, where two years ago they managed minutes. We see both in our own client work, where jobs we had to cut into small steps in 2024 now run from start to finish.
The model inside that tool could probably gather the contract and the record for many of the small claims and draft an appeal for a person to review. No one has asked it to.
Intelligence of this kind will never be more expensive than it is today. Yet most of the systems we see were built as if it were scarce, with every task cut down to fit a weaker and costlier model. The scope was set by what that model could handle and by what a six-week project could deliver. When the vendor swaps in a newer model, the newer model inherits the same narrow scope.
Some limits on the scope have good reasons. An appeal that goes out under a doctor’s name should be approved by a person. Other limits exist only because the old model couldn’t do the work, and no one has gone back to sort one kind from the other. Those same limits also decided what the people around the system spend their day on, which we looked at in Accidental Job Designers.
The owners of these businesses want better margins. They want to take on more practices without hiring another billing team. What they’re usually offered is one more tool that handles one more step, scoped as narrowly as the last one. It is the pattern we described in The Department Is the Last Legacy System: each step automated well, inside boundaries drawn long before the AI arrived.
Cost controls make it worse. Companies watch the model bill because it comes every month, so they move to cheaper models and cap usage. They’re cutting the one cost that’s already falling fastest. The larger cost is the people who check the model’s work and finish what it couldn’t, and once you count them, the better model is often the cheaper one.
Saved time isn’t profit until someone uses it, by taking on more clients without hiring or by recovering money the company was losing. Every written-off claim is a payment the practice never gets. Since revenue cycle companies are usually paid a share of what they collect, it’s also a fee the company never earns.
Most companies we work with have no one who owns the question of what their AI system should take on next. The vendor was paid when the project ended, so it has little reason to come back, and IT’s job is to keep the system running. So the models improve every few months, and the system keeps doing the work it was scoped for on the first day. Keeping AI improving is a different job from keeping it running, which is why we argued that your IT department should not be in the IT room anymore.
The write-off limit at the company I spoke to was set for a team that worked every appeal by hand, and the tool was scoped for a model that could do little more than read. Both made sense when they were set. Two years and two model upgrades later, both are still in place.
Frequently asked questions
What is a write-off limit in revenue cycle management?
A write-off limit is the claim value below which a revenue cycle company stops chasing a denied or underpaid claim. Chasing those small claims would cost more than the company can hope to recover, so most are written off. Each one is a payment the practice never gets and, because revenue cycle companies are usually paid a share of what they collect, a fee the company never earns.
Why do AI systems keep doing the same narrow job after model upgrades?
Most were scoped to fit the model and the project of the time: a weaker, costlier model and a six-week delivery. When the vendor swaps in a newer model, it inherits the same narrow scope. Meanwhile, by Epoch’s estimate, the same level of capability gets more than 90 percent cheaper every year, and models can now stay on one task for hours instead of minutes.
Is switching to a cheaper AI model a good way to cut costs?
Often not. The model bill gets attention because it arrives every month, but it is the cost that is already falling fastest. The larger cost is the people who check the model’s work and finish what it couldn’t, and once you count them, the better model is often the cheaper one.
How does time saved by AI turn into EBITDA?
Saved time is not profit until someone uses it. That means taking on more clients without hiring, or recovering money the company was losing. In revenue cycle work, a model could gather the payer contract and patient record for small denied claims and draft appeals for a person to review, turning write-offs into collections.
Who should own what an AI system takes on next?
In most companies, nobody does. The vendor was paid when the project ended, so it has little reason to come back, and IT’s job is to keep the system running. Someone has to keep asking what the system can take on next, because the models improve every few months while the scope stays fixed.
What should a company do with an AI system that hasn’t changed since it went live?
Go back through its limits and sort them. Some have good reasons, such as an appeal under a doctor’s name needing a person’s approval. Others exist only because the old model couldn’t do the work, and those are candidates for a bigger job once a newer model holds up on the hardest cases.
