Epic's AI bill is now variable, so every AI feature needs a value gate

Part of the bill for AI inside Epic is now variable. Becker's reports that it moves with how heavily staff use each feature, which AI model runs underneath it, and what patients do with their own records, and that the tools to forecast that spending are still being built. When a cost line can swing that far, the budget question stops being "how much is AI" and becomes "which features have earned their spend."

Ochsner Health CIO Amy Trainor put the range plainly at Epic's Users Group Meeting: "Each feature, hundreds of features, will show what that looks like. It could be $10. It could be $10,000, depending on what the usage and the price model looks like." I take that quote as a design requirement. If cost is decided per feature, governance has to be decided per feature too.

Why the annual budget stops working

Three things move the number, and Becker's CIO interviews name all of them.

The price model is mixed. Trainor told Becker's that Epic's own AI features are billed inconsistently, some flat and some consumption-based. Baptist Health South Florida's Sha Edathumparampil described a fixed fee and then a variable charge that depends on which frontier model you use.

Volume follows adoption. Baptist Health's Aaron Miri asked the question every finance team will ask: if the system goes "hog wild" with Agent Factory, "who's paying the piper on the back end, paying for the tokens?" A tool that succeeds gets used more, so success and cost pressure arrive together.

Visibility is partial. Trainor said Epic's tracking is not yet down to the user: "Right now it's by function … and by time." Epic said in August it is building individualized cost-tracking and forecasting tools.

There is a quieter fourth factor. RWJBarnabas Health CIO Rob Adamson told healthsystemCIO that flat monthly fees for Epic AI tools have concealed what individual tools could cost under the consumption pricing he expects in 2028. A flat fee feels safe. It also means nobody has built the per-feature habit yet.

What a value gate is

Adamson's version is the clearest I have seen. Good governance asks the sponsor what success looks like, sets a minimum threshold the tool must reach, and asks whether it can reach that threshold within 90 days. If the answer is no, the model gets shut down and its compute goes to the tools waiting in line. His line on why: "The only way that scales is shutting things off."

He built a volume model that asks operational leaders to estimate how much each tool will run. For discharge summarization alone, it put the annual cost at about $312,000 under consumption pricing, and that is one of roughly 80 AI tools he has to account for. The same article cites a Gartner survey in which only 17% of CIOs reported guidelines for abandoning low-value AI already in production. Most organizations have an approval step. Few have an exit.

He is also clear that money is not the only gate. "ROI can be one of the value gates, but I don't think it's the only value gate," he said, and his C-suite also responds to clinical value such as lives saved through an early-warning model, even when it resists a dollar figure.

The per-feature worksheet

Here is the worksheet I would keep for every AI feature, one row each. It fits in a spreadsheet. The discipline is filling in every column before go-live, not after the first invoice.

1. Feature and owner. The feature name as the vendor bills it, and one named business sponsor who will defend it at review. Not a committee.
2. Pricing type. Flat, consumption, or both. If consumption, the unit (tokens, calls, encounters) and which model sits underneath, because a model change can change the rate.
3. Volume driver. What makes usage go up: users, encounters, notes, patient messages. Ask the operational leader for an estimate, the way Adamson does, and write down who made it.
4. Projected cost. Volume times unit rate, for today's price model and for the price model you expect next. A feature on a flat fee today still gets a consumption estimate.
5. Success measure. One measurable outcome the sponsor agrees to: minutes saved per note, denials avoided, alerts acted on. Clinical value counts even when it has no dollar figure.
6. Minimum threshold. The number below which the feature is not worth its cost. Write it before launch, so nobody moves it later.
7. Gate date. Ninety days from go-live, on the calendar, with the sponsor invited.
8. Usage cap or alert. A monthly limit or a spike alert where the tool allows it, so a surprise shows up in weeks rather than at true-up.
9. Decision and reason. Keep, change, or switch off, with one sentence on why.

The decision rule

At the gate, the rule is short.

If the success measure meets the threshold and actual cost is within the projection, keep it and set the next review.

If the measure meets the threshold but cost is running above projection, keep it only with a change: a cheaper model, narrower access, or a usage cap. Adamson's routing advice applies here, sending each request to the least sophisticated model that can answer it.

If the measure misses the threshold, switch it off and release the budget to the next feature in line. A sponsor can bring it back with a new measure and a new threshold, but it goes back through intake.

If nobody can produce the measure at all, treat that as a miss. A feature with no evidence of value is a cost line, not a pilot.

The hard part is the third branch. Turning off a tool someone likes is a political act, which is why the threshold has to be written down before launch by the person who asked for the tool.

Where to start this week

You do not need Epic's forecasting tools to begin. Pull the list of AI features currently turned on, add the nine columns, and fill in owner, pricing type, and success measure first. Any row where the sponsor cannot name a success measure is your first gate.

I wrote about the same pattern outside the EHR in The seat no longer caps your AI bill, where usage meters are showing up behind seat licenses. The fix is the same in both places: put cost next to delivered value on a schedule, and decide. How many of your AI features could pass a 90-day gate today?