When AI Goes Dark: Downtime Planning for an Automated Revenue Cycle

Picture a Monday morning six months after your hospital moved most of its outpatient coding to an autonomous AI tool. The coding team that used to handle that volume has been cut to a handful of auditors. Then the AI vendor's model provider has an outage. Or the vendor pushes a model update, and the coding still flows, but the accuracy drops, and nobody notices until denials start coming back three weeks later.

The charts don't stop coming. The timely filing clock doesn't stop either.

Every hospital I know has a downtime procedure for the EHR. Most have tabletop exercises for ransomware. Very few have thought through what happens when the AI doing real revenue cycle work disappears, or keeps running while it gets worse. A group of health system technology leaders spoke openly about that question in recent interviews with Becker's, and their answers are quite telling.

What happens to the revenue cycle when AI goes down?

Daniel Kortsch, MD, associate chief AI and digital health officer at Denver Health, put the problem in one sentence: "If that system goes down, the work does not disappear, it returns to a team whose capacity may already be fully committed."

In the revenue cycle, the work that returns has deadlines. Uncoded charts hold up claims. Unsubmitted prior authorizations hold up scheduled procedures. Denial appeals have filing windows. Every day a backlog sits, cash slows down, and some of it may never come in.

And recovery has its own risks. Shaun Miller, MD, chief health informatics officer at Cedars-Sinai, told Becker's that queued messages, partially generated documentation and delayed tasks may all need to be reconciled when a service comes back. In revenue cycle terms, that means partially coded accounts, half-drafted appeals, and work queues that need sorting before anyone can trust them again.

When does an AI tool need a continuity plan?

Dr. Kortsch sorts AI tools into two groups. Some help people who could still do the work by hand. Others have taken on work that nobody routinely does anymore.

"The trigger for needing a continuity plan is the moment the AI tool stops being an accelerator and becomes the only path," he said.

Ambient clinical documentation is usually in the first group. If it fails, clinicians can go back to typing their notes. It's slower, but the work continues. An autonomous coding tool that replaced most of your coders is in the second group. So is any agent that took over prior auth submissions, eligibility follow-up, or appeal drafting after you reduced the staff who used to do it.

Cleveland Clinic CIO Sarah Hatchett set the bar this way: "Once the loss of an AI capability can impact patient care, operational throughput, revenue, or safety, it should be treated like any other mission-critical system: documented, tested, and recoverable." Revenue is in that category. Cleveland Clinic tiers its AI and other software by how critical the function is, how long an outage it can tolerate, and how quickly it needs to come back.

Cedars-Sinai looks at patient safety, how dependent the workflow is, how widely the tool is used, how time-sensitive the work is, whether a safe alternative exists and what happens if the tool stays down for a long time. Every one of those applies to revenue cycle tools.

This matters more as some organizations go much further with automation. One AI company CEO told a Becker's conference audience that his organization has replaced roughly 80% of its workforce with AI agents running around the clock, including revenue cycle coding. Whatever you think of that approach, the higher the share of work that runs only through AI, the more a continuity plan is required.

What does an AI failure look like if nothing actually crashes?

Traditional downtime planning asks one question: is the system up or down? AI adds a harder one. The tool can stay online while the results get worse.

Dr. Kortsch listed the ways an EHR-integrated AI tool can be disrupted: an EHR outage, a vendor outage, a model provider outage, an integration problem, or a change in the underlying model or data. Only the first three look like a traditional outage. The last two can change what the AI produces without setting off a single alert.

His example was an AI tool that summarizes clinical information and starts producing less accurate summaries after a model or data change, while appearing to run normally. "That is a failure mode the industry is still learning how to identify," he said.

In revenue cycle, this kind of failure shows up late. A coding model that drifts doesn't throw an error. It shows up weeks later as a denial trend, an audit finding, or a payer study like the one Blue Cross Blue Shield released in September. A denial-prediction model that drifts lets bad claims through. By the time anyone sees the pattern, the claims are already out the door.

"As an industry we have spent years building rigor around pre-deployment validation and comparatively less on what comes next," Dr. Kortsch said. He added that ongoing monitoring creates recurring work and cost, and that the responsibility may fall to the health system rather than the vendor.

What is an AI kill switch, and who gets to pull it?

Hasan Ahmad, MD, associate CMIO at Parkview Health, learned the hard way what happens without one. Parkview spent years untangling EHR alerts that had piled up over 15 years, with no record of who owned them or why they existed. Now, he told Becker's, the kill switch is "not optional, it's part of the design requirements." Before a tool goes live, Parkview agrees on the metric it's supposed to move and what failure looks like, so turning it off doesn't require a negotiation.

Brigham and Women's Hospital sets a date at the beginning. "We say once we hit these metrics, this is the date support stops," said CIO Micheal Sweet. The tool either earns its way into production or it gets shut down, "so they don't run forever."

Mayo Clinic's chief digital officer, Nari Gopala, asks three questions before an AI tool reaches a patient: who built it, who decided it was ready for this use, and who owns its operation, meaning "who's getting that 2 a.m. call if the model misbehaves."

Seattle Children's scores every AI use case on its reach, whether a human stays in the loop, how reversible its actions are and how much harm it could do. The score decides whether a tool can simply be released or needs a full pilot with a clear go/no-go threshold.

For revenue cycle, I'd add one question to Mayo's list: who owns the work when the AI stops? That has to be a named team with the skills and capacity to pick it up, not "revenue cycle" in general.

How should CFOs decide whether to scale, pause, or kill an AI tool?

At Becker's IT + Revenue Cycle Conference in September, several health system leaders described their frameworks.

Northwestern Medicine agrees on success metrics before a pilot starts, and holds to them. "If it doesn't [hit those metrics], it needs to be shut down. And we shut it down," said Doug King, Northwestern's chief digital information officer. He warned about the most common trap: "Well, the next version will be better, or just give us two more weeks." You can do that forever, he said, and never get to the next project. At Northwestern, the decision to end a pilot includes King, the CFO, the COO, and a senior clinician, because "people get tied to these things because they want it to work."

Endeavor Health's Nirav Shah, MD, runs more than 100 AI tools through three tests: Is the model performing? Does it fit into the workflow people already use? Does it create real value? If a tool fails one, Endeavor optimizes it, pauses it, or kills it. Sometimes a pause is the right call. Endeavor's chart summaries didn't work for case managers until the team cleaned up its messy notes. "Sometimes we go fast by going slow and fixing the underlying infrastructure," Dr. Shah said.

Aspen Valley Health, a 25-bed critical access hospital, negotiates 30-day no-cause termination clauses into AI vendor contracts so it can exit quickly when a tool stops earning its keep.

King also made a point that every CFO should hear. AI value isn't fixed. Token costs change, models change, and a core platform may release something that makes a third-party tool unnecessary. "You need to revisit it," he said. "Historically that's something that healthcare has done poorly. We just implement something and then we're off to the next thing." At hospital margins averaging about 0.75%, per Dr. Shah, no one can afford AI that stopped paying for itself last quarter.

A revenue cycle AI continuity checklist

Here is what I'd put in place for every AI tool that touches revenue.

1. Inventory your AI and label each tool "accelerator" or "only path." Use Dr. Kortsch's test. If your staff could do the work by hand tomorrow, it's an accelerator. If they couldn't, it's the only path, and it needs a continuity plan.

2. Map the dependencies. Cedars-Sinai is building a map of which workflows use which models, vendors, cloud services, data feeds and interfaces, and what happens if any link breaks. Do the same for revenue cycle. Most organizations don't know which of their RCM vendors depend on the same model provider.

3. Keep minimum human capacity for every "only path" tool. Decide how many trained people you need to keep critical work moving during a multi-day outage, and keep them trained. This is the hardest item on the list, because it costs money while everything is working.

4. Monitor quality, not just uptime. Track coding accuracy, denial rates and audit results for each AI tool weekly. Set a threshold that triggers a review. Silent drift is the failure you won't see on a status page.

5. Build the kill switch before go-live. Write down the success metric, the failure threshold, the review date and who has authority to turn it off. Put a no-cause exit clause in the contract.

6. Rehearse an AI outage. Add AI scenarios to your tabletop exercises: losing a model provider, an AI tool becoming unreliable, several AI tools failing at once. Cedars-Sinai recommends all three.

7. Plan the recovery. Decide in advance how you'll reconcile half-finished work when the tool comes back: partially coded accounts, queued submissions, drafts no one approved.

What does this mean for the revenue cycle workforce in 2030?

In the RCM 2030 Workforce Modernization companion guide, I described the revenue cycle of 2030 as working less like a production line and more like a quality control lab. One of the first roles I recommended investing in is the automation oversight analyst: the person who monitors automated workflows and validates AI output.

AI continuity planning makes that role non-negotiable. Someone has to watch for drift, own the kill switch and know how to do the work when the AI can't. If you cut the people who understand the work before you've built the people who oversee the automation, your first AI outage will cost more than the savings. I wrote about that gap in We're hiring for the technology and firing for the technology in the same fiscal year.

Cleveland Clinic's Hatchett put it this way: "While downtime is something we hope to avoid, we must be realistic that it happens." Your AI will go down someday. Make sure your revenue cycle still works when it does.

Frequently asked questions

Does every AI tool in the revenue cycle need a continuity plan?

No. Denver Health's test is whether the tool is an accelerator, meaning staff could still do the work manually, or the only path. Tools that have become the only path for critical work need a documented, tested continuity plan.

What is an AI kill switch in healthcare?

An AI kill switch is a predefined plan for turning off an AI tool: the metric it must meet, what counts as failure, the review date and who has the authority to shut it down. Health systems like Parkview and Brigham and Women's now require it before a tool goes live.

How can AI fail without an outage?

A change to the underlying model or data can make an AI tool produce worse results while it appears to run normally. In revenue cycle, this often shows up weeks later as rising denials or audit findings.

Who is responsible for monitoring AI after it goes live?

Often the health system, not the vendor. Denver Health notes that ongoing monitoring creates recurring work and costs that health systems need to plan and budget for.

Not sure which AI tools are worth keeping?

Prove It or Pull the Plug: The AI Value Audit for Health System Leaders is a free seven-page field guide with three tools (the Root Cause Test, the Training Delta Audit and an ROI Scorecard) to decide which AI investments deserve continued funding within 30 days.

For the full picture of the 2030 revenue cycle workforce, read the RCM 2030 Companion Guide: Workforce Modernization.

I cover what this means for health systems every Sunday in RCM 2030 Weekly. Subscribe on LinkedIn.