A hospital’s first AI project should be a single administrative workflow, run for 90 days, owned by one operational leader, and judged against one baseline number. Pick work that is high-volume, rules-based, and measurable — eligibility checks, prior authorization preparation, claim scrubbing — then scale or stop based on what the number does.
Days 1–30: Pick the work, not the tool
The first month is selection, and selection is most of the outcome. Inventory the repetitive, rules-based tasks your teams carry every week and rank them by hours consumed and error cost. The raw material is not scarce: administration accounts for roughly a quarter of U.S. hospital spending (Health Affairs), physicians average 13 hours per week on prior authorization alone (AMA, 2024), and health systems spend $19.7 billion a year pursuing denial appeals (Premier Inc.), at $25 to $181 of rework per denied claim (MGMA).
The right first project has four properties:
- Volume. The task happens hundreds of times a week, so results are statistically visible inside a quarter.
- Rules. The work follows definable logic with a manageable set of exceptions.
- A number. There is one metric — denials prevented, hours returned, days in A/R — that the project will be judged on.
- An owner who wants it fixed. A director or VP who lives with this workflow and will show up to the weekly review.
Before anything is purchased, record the baseline. What does this work cost today, in hours and dollars, measured the same way you intend to measure it afterward? A project without a before-number cannot prove an after-number. This is the single most common omission, and it is fatal to the business case.
Days 31–60: Implement narrow
The formula for the middle month is one workflow, one owner, one metric. Resist every invitation to widen scope; breadth is what turns 90-day projects into 18-month stalls. The AI handles the routine volume, and the exceptions route to a human — define that boundary explicitly, because how the system fails matters more than how it demos.
Hold vendors to operational terms. If a vendor cannot state which number will move and by when, that is a marketing claim, not a business case — the standard laid out in Deciding Where AI Fits, a CFO Guide. Insist on integration with the systems your staff already use, and put the baseline owner and the vendor in the same weekly meeting, looking at the same number.
Days 61–90: Measure and decide
The final month is a comparison against the baseline, not a feelings check. Published ROI averages trace to industry compilations rather than primary research, so the comparison that matters is against your own recorded baseline (2025 benchmarks). For a first project, the practical bar is 2x: if the deployment is not on a credible path to returning at least twice its cost, it does not scale.
Ninety days ends in one of three decisions. Scale it — the number moved, extend the same workflow to more sites or volume. Fix it — the mechanism works but the scope was slightly wrong, adjust and re-measure. Stop it — the number did not move, shut it down and take the selection lesson. All three are wins compared to the alternative, which is a pilot that runs for a year and a half with no defined success criteria.
Where first projects go wrong
About 75 percent of health systems now run at least one AI application, yet 95 percent of enterprise AI pilots produce no measurable return (benchmarks). That gap is not a technology gap. It is project selection: buying a platform first and hunting for problems second, skipping the baseline, assigning ownership to a committee, or judging an operations project as an IT installation. AI is not a software purchase; it is a workflow change that happens to involve software.
One more selection signal: when 43 health systems were surveyed on their top goal for AI, 72 percent said reducing caregiver burden — far ahead of margin improvement at 12 percent (JAMIA, 2025). A first project framed as giving staff their hours back, with the financial return as the consequence, gets cooperation instead of resistance. That framing is also simply accurate: the work AI absorbs best is the work nobody went into healthcare to do.
Common questions
What should a hospital's first AI project be?
The best first project is a single administrative workflow that is high-volume, rules-based, and measurable: eligibility verification, prior authorization preparation, claim scrubbing before submission, or scheduling follow-ups. It should not be clinical diagnosis. Administrative work carries no clinical risk, has a clear baseline cost, and produces a result you can count in dollars or hours.
How long should a first hospital AI project take?
Ninety days to a measurable result. Typical payback on healthcare AI investments runs 12 to 18 months, but a well-scoped first project should show clear movement in one quarter. If nothing measurable has moved in 90 days, the scope was wrong, not the timeline.
What ROI should a hospital expect from a first AI project?
Published ROI averages for healthcare AI trace to industry compilations rather than primary research, so the only figure worth planning against is your own. The practical bar for a first project is 2x: if the deployment is not on a credible path to returning at least twice its cost after the 90-day proof window, fix the scope or stop. Both outcomes beat an 18-month stall.
Who should own a hospital's first AI project?
One operational leader who owns the workflow being changed, such as the director of patient access or the revenue cycle VP. Not IT alone, and not a committee. The formula is one workflow, one owner, one metric. Projects with a committee for an owner produce reports; projects with an operational owner produce numbers.
Want your first project picked in one afternoon?
Schedule a Strategy Call. Twenty minutes, no cost. We’ll identify which workflow in your operation meets the four criteria and what its baseline should be.
Schedule a Strategy Call