How many AI pilots should a revenue cycle team run at once?

  • 4 of 33AI proofs of concept that reached production in IDC's researchSource
  • 2%of provider executives with AI or automation across their whole revenue cycleSource
  • Halfas many AI initiatives at the companies getting the most value from AISource

Short answer

Two or three, each with one rev cycle owner, one measure tied to cash and a stop-or-scale date. With five unfinished, keep the two closest to production and pause the rest in writing.

Key takeaways

  • Run two or three revenue cycle AI pilots at once, and never more than you have operations leaders with time to own them.
  • Give each pilot one measure tied to cash, a baseline, and a stop-or-scale date 8 to 12 weeks out.
  • Don't run two pilots in the same work queue. You won't be able to tell which one moved the number.
  • Start a new pilot only when one of the current ones ships or stops.

A revenue cycle team should run two or three AI pilots at once, and never more than it has operations leaders with time to own them. Five unfinished pilots usually means each one is missing the same three things: an owner inside rev cycle operations, a measure tied to cash, and a date when someone decides to scale it or stop it. Keep the two closest to production, pause the rest in writing, and start a new one only when one of those ships or stops.

Why two or three

The limit isn’t the AI. It’s the people each pilot borrows. Every revenue cycle pilot needs a manager who can change the workflow, billers, coders or appeals specialists to check the output, an EHR analyst to put it into the work queue, and someone in finance to sign off on the baseline. In most hospital and medical group rev cycle teams those are the same few people, and they are already stretched. Spread them across five pilots and none gets enough of them to finish.

The research points the same way:

  • BCG’s survey of 1,000 executives found that the companies getting the most value from AI pursue about half as many opportunities as their peers and expect more than twice the return. Its advice for getting started is one to three high-return initiatives.
  • IDC found that for every 33 AI proofs of concept a company launched, four reached production.
  • In Guidehouse and HFMA’s survey of 191 provider executives, 39% were implementing point solutions in the revenue cycle and about 2% had AI or automation integrated across all of it.
  • L.E.K.’s 2026 survey found most AI use cases, including prior authorization, coding and denial management, piloted at 70% to 80% of health systems but rolled out enterprise-wide at only 15% to 30%.

Large health systems do run dozens of AI applications. Universal Health Services has more than 50 live, and it gets there through four stages (proof of concept, pilot, limited deployment, enterprise-wide), with success criteria and an executive sponsor set before each pilot starts. What matters is how many pilots each team carries, not the enterprise total. A system with a dedicated revenue cycle automation team can carry one pilot per segment (patient access, mid-cycle coding and CDI, back-end denials and follow-up) as long as each has its own owner.

How to cut five pilots to two

  1. Write one line per pilot. The owner, the work queue it touches, the measure it should move, the baseline, the result so far, and what stands between it and production: EHR integration, accuracy, staff adoption or the contract.
  2. Put a monthly dollar figure on each. For a denials pilot, that’s the write-offs and rework in the denial categories it covers. HFMA reports 11.65% of claims were denied on first pass in 2025, and Premier puts the average cost of fighting a denial at $57.23 per claim, so back-end pilots usually have the easiest math.
  3. Keep the two with the most dollars and the shortest path to production. If two pilots sit in the same work queue, keep one. When both touch the same claims you can’t tell which one moved the number.
  4. Pause the others in writing. Note what each taught you, keep the data and test cases, and write down what would restart it.
  5. Set a production gate for each pilot you keep. Accuracy on a set of real claims your specialists already worked, the share of the queue it handles, the measure against baseline, output landing in the Epic, Oracle Health or MEDITECH work queue staff already use, and a named person who runs it after go-live.
  6. Pick a decision date 8 to 12 weeks out. On that date the pilot scales, gets one fixed round of changes, or stops. Mount Sinai ran its AI scribe decision this way: three vendors over three months, scored on accuracy, cost, technical integration and user satisfaction.
  7. Hold the limit. The next pilot starts when one of the two ships or stops.

What keeps rev cycle pilots from finishing

  • The owner sits outside rev cycle. A pilot run by IT or an innovation group can work technically and still change nothing, because nobody with authority over the work queue changed it. Health system leaders told Becker’s that some early pilots performed well technically but didn’t gain traction because they added friction to workflows.
  • The measure counts activity. Appeal letters drafted isn’t a result. Overturned denials, cash collected, cost to collect, days in A/R and touches per claim are.
  • The tool lives in a separate screen. If staff have to leave the work queue to use it, it won’t become part of the job. In Qventus’s 2026 survey of more than 60 health system technology leaders, 74% named EHR dependency as a top barrier, 45% named difficulty scaling pilots, and only 4% had scaled AI with measurable outcomes.
  • Vendor pilots have no end date. Open-ended pilots rarely have exit criteria. “We no longer treat launching pilots as success,” Lisa Stump, Mount Sinai’s chief digital information officer, told Becker’s. Mount Sinai defines value up front and makes reaching it the bar for continuing.
  • Nobody tests on cases the system hasn’t seen. A demo on hand-picked claims tells you little. Test on a held-out set of recent claims from your own payers, worked by your own specialists.

What one finished pilot looks like

One workflow, one measure, one owner. For a revenue cycle management company, we built agents that research each denied claim, run eligibility and claim status checks, apply the right SOP and draft the next action with citations for a specialist to approve. It is measured on outsourced denial-research capacity, which it cut by 25%, and it is expected to deliver $4M in run-rate savings by the end of 2026 (case study).

On timing, our referral evaluation work at a health system reached its first automation in three months. That is about the window a well-scoped pilot needs, and why a decision date 8 to 12 weeks out is fair to the pilot and to the team waiting on the people it borrows.

Questions

Related questions

Should we stop the revenue cycle AI pilots that are stuck?

Pause them in writing instead of letting them run in the background. Record what each one taught you, keep its data and test cases, and note what would have to be true to restart it. A pilot with no owner and no measure has already stopped, and making it official gives its people back to the pilots you kept.

How do I tell real AI from rules-based automation a vendor rebranded?

Give it 50 to 100 of your own recent denials it hasn't seen, including unusual payers and denial reasons, and compare its output with what your specialists did. Ask what happens when a payer changes a policy, since rules need someone to rewrite them. Ask for the error rate on your claims, not the automation rate on theirs.

Should we buy a denials product or have something custom built on our own data?

Buy when the workflow is standard and a product already fits your EHR and payer mix. Build when the value sits in your own SOPs, payer contracts and claim history, or when the products you tried cover only part of the work. Either way, hold it to the same production gate and decision date.

Do we need a revenue cycle consulting firm to decide which pilots to keep?

You can do the triage below in a few working sessions with your rev cycle leaders and finance. Firms like Guidehouse, Huron, Chartis and Tegria take on broader revenue cycle transformation and EHR optimization, and KLAS's 2025 report credited Huron, Chartis, Healthrise and Impact Advisors with finding high-value areas to optimize. To get one AI workflow into production, look for a team that builds it inside your work queues and stays to run it.

Tell us about the work you want off your team’s plate