AI in Medical Practices: What the Adoption Surveys Actually Show

Medical Group Management Association (MGMA) Stat polling in 2025 found that AI tools have become the top technology priority for practice leaders, named by 32% of respondents. That figure put AI ahead of EHR usability improvements, which had led the same question in prior years. The shift is notable less for the specific percentage than for what it displaced: EHR complaints have been a fixture of practice-management surveys for a decade, and a new topic overtaking them is a real change in what administrators are paying attention to.

That attention has not translated into a settled picture of what AI is doing for practices that use it. The adoption numbers and the outcome numbers answer different questions, and conflating them is common in how AI tools get marketed to practices. This post separates the two, using MGMA Stat's own data, and ends with a way to check the difference for a specific practice rather than take either claim on faith.

What the adoption numbers actually say

Roughly half of the practices MGMA surveyed report they already use at least one AI tool somewhere in their operations. That is a broad category. It covers everything from ambient documentation tools that draft notes during a visit, to scheduling and intake automation, to AI-assisted coding suggestions, to chatbots handling patient messages. A practice that piloted one AI feature in one department counts the same as a practice with AI embedded across several workflows.

Looking forward, more than half of surveyed practices say they plan to apply AI specifically to revenue-cycle-management tasks. That is the single most commonly cited planned-use category in the survey, ahead of clinical documentation and patient communication. It is a reasonable place to start: RCM tasks like claims scrubbing, denial follow-up, and payment posting are rules-based, high-volume, and already partly automated in most practice-management systems, so adding AI to them is an incremental step rather than a new category of software.

Both of these findings describe adoption: whether a practice has picked up a tool and where it plans to point it next. Neither describes what happened after adoption. A survey question that asks "do you use AI" or "where do you plan to use it" cannot answer "did it work," and the MGMA data does not claim otherwise.

The question MGMA itself has not answered

MGMA's own Stat coverage has raised a direct question about its adoption findings: is AI actually reducing staff workload, or is it shifting the work to a different point in the process without reducing the total amount of it? This is presented as an open question in MGMA's reporting, not as a conclusion in either direction, and it is worth taking at face value.

The distinction matters operationally. An ambient documentation tool that drafts a note during a visit can save a clinician time at the point of dictation while adding a new task: reviewing and correcting the draft before it is signed. A coding-assist tool can speed up code selection while adding a review step for a biller who now has to check the AI's suggestions instead of assigning codes from scratch. In both cases, a practice could adopt the tool, report using AI in a survey, and see no change in total labor hours or in the total time a claim or a note takes to finish. The work does not disappear. It moves.

None of this means AI tools provide no benefit. It means "we use an AI tool" and "our costs went down" are separate claims, and only one of them is what most practices have actually measured so far. MGMA's polling captures the first claim. It does not, by MGMA's own account, currently answer the second.

Why the interest is rising now

The context for the AI findings is a separate MGMA Stat result from August 2025: cost and margin pressure ranked as practices' top overall priority, named by 41% of respondents, and 90% of surveyed groups reported higher operating costs than the prior year. That is the backdrop against which the AI numbers should be read. Administrators are not adopting AI tools because the technology has proven itself against a cost baseline. They are adopting it, or considering it, because operating costs are up almost across the board and AI is the current candidate for bringing them down.

A candidate is not the same as a proven fix. Being the most-discussed technology option in a year when practices are under unusual cost pressure says something about where attention is going. It does not say the tools deliver the savings that attention implies. A practice leader evaluating AI spending in this environment is choosing between an unproven option and no option, which is a different decision than choosing between a proven option and no option, and the survey data does not let the two get confused.

A framework for checking whether it actually worked

Vendor-reported averages describe results across a vendor's customer base, aggregated in ways that are rarely disclosed in detail and never specific to one practice's payer mix, staffing, or workflow. A practice that wants to know whether a specific AI tool is worth its cost needs a measurement built on its own numbers, taken before and after adoption of that specific tool for that specific task.

• Pick one task, not the whole workflow. "AI helped with documentation" is not measurable. "Average time from end of visit to note completion" is.
• Baseline it before adoption. Record the metric for several weeks under the current process, using the practice's own staff and its own patient mix, before any tool is introduced.
• Measure the same metric after, not a different one. If the baseline was time-to-note-completion, the follow-up measurement has to be time-to-note-completion, not "clinician satisfaction" or "notes drafted per day," which are different measurements that can move for reasons unrelated to the tool.
• Include the new work the tool creates. If a coding-assist tool requires a biller to review its suggestions, that review time counts against the tool's total time budget. Measuring only the AI's output time and ignoring the human review step attached to it overstates the benefit.
• Hold staffing and volume steady across the comparison period. A change in patient volume, a new hire, or a shift in payer mix during the measurement window can move the metric independent of the tool, and any of those changes should be noted alongside the result rather than folded into it.
• Compare against the practice's own baseline, not a vendor's published average. A vendor's reported time savings reflects its customer base, not this practice's task mix, EHR configuration, or staff experience with the tool.

This approach takes longer than reading a vendor case study, and it produces a smaller, more specific answer: whether one tool changed one measured task in one practice. That is a narrower claim than "AI reduces administrative burden," but it is the claim a practice can actually verify with its own data, and it is the one the current survey data does not yet answer on their behalf.