AI in CDI Creates More Work, Not Fewer CDI Jobs

Ask a CDI manager whether AI in CDI is going to eliminate the department and you usually get a tired look. Ask the same manager whether AI in CDI has added to the team’s workload this year and you get a much longer answer. The distance between those two reactions is the honest state of AI in CDI right now: the technology is absorbing specific tasks, and nearly every task it absorbs hands back a different one.

Someone still has to check what the model produced, document why the tool was trusted, and defend the resulting record when a payer’s own algorithm challenges it. None of that appeared in a CDI job description ten years ago. It exists now, and it lands on the same team that was already reviewing charts.

Will AI in CDI replace CDI specialists? What the labor data says

The replacement fear around AI in CDI is not irrational, and it deserves a straight answer rather than reassurance. The U.S. Bureau of Labor Statistics names the risk directly. In its Occupational Outlook Handbook entry for medical records specialists, BLS writes that “the increase in adoption of artificial intelligence (AI)-powered solutions that make the medical coding process more efficient may affect the demand for these workers.”

That is the federal statistical agency saying, in print, that automation is a factor. In the same entry, BLS projects the occupation will grow 7% from 2024 to 2034, much faster than the average for all occupations, rising from 194,800 jobs in 2024 to 208,600 in 2034, with about 14,200 openings a year over the decade. Median pay was $50,250 in May 2024.

The adjacent occupation looks stronger still. BLS projects health information technologists and medical registrars to grow 15% from 2024 to 2034, from 41,900 jobs to 48,100, with roughly 3,200 openings a year and median pay of $67,310 in May 2024.

Projections are projections, and BLS revises them every two years. Read them for direction rather than precision. Still, the agency that flagged AI as a demand pressure expects both occupations to add jobs over the decade, not shed them. Anyone telling a CDI team that AI in CDI is about to make them redundant is making a claim the current labor data does not support.

AI in CDI is already standard, not coming

The second problem with the replacement story is timing. Treating AI in CDI as a future event is out of date. Software that reads charts and suggests findings has been ordinary equipment in CDI departments for years.

In the ACDIS 2025 CDI Week Industry Survey, which drew 783 respondents, the most-used CDI software was electronic grouper software at 78.25%, followed by chart prioritization at 77.47%, electronic querying at 77.31%, and computer-assisted coding at 76.83%. Roughly three out of four responding programs already run tools that read a record and point a human at something worth a second look.

Those departments did not disappear. They got busier. That is the pattern worth studying before anyone builds a headcount plan around AI in CDI.

The larger shift is upstream of CDI entirely. In the American Medical Association’s 2026 Physician Survey on Augmented Intelligence, which surveyed nearly 1,700 physicians and was published in March 2026, 81% reported using AI professionally, more than double the 2023 rate. Thirty percent used it to create discharge instructions, care plans, or progress notes. Twenty-eight percent used it for documentation of billing codes, medical charts, or visit notes, and another 28% for generating chart summaries.

Read those numbers from a CDI chair. The chart a specialist opens tomorrow morning is increasingly a document a model helped write. AI in CDI is no longer only a question about the tools a CDI department buys. It is a question about the documentation a CDI department receives.

Where AI in CDI creates new work

AI in CDI generates work in three specific places. None of them are hypothetical, and all three land on the same headcount.

Someone has to read what the model wrote

Published evaluations of ambient documentation tools do not support signing a machine-generated note unread. In a 2025 study in Mayo Clinic Proceedings: Digital Health, researchers ran 14 simulated ambulatory encounters through five commercial ambient scribe platforms and compared the output against professional transcription. They reported a mean clinical note error rate of 26.3% (95% CI, 17.0% to 31.0%), an average of 3 errors per case carrying potential for moderate to severe harm, and only 35.8% of correct clinical elements appearing consistently across all five platforms.

A 2026 perspective in npj Digital Medicine on scaling ambient AI scribes describes a related problem the authors call note bloat: records that grow longer without growing more useful, so time saved during the visit gets spent later by whoever has to read the chart. In a hospital, whoever reads the chart is very often CDI.

None of this means the tools are bad. It means the output is a draft. Drafts require review, review is labor, and that labor is the first place AI in CDI quietly adds hours instead of saving them.

Governance became part of the job

The federal government now treats predictive AI inside the EHR as something that has to be documented and monitored. Under the ASTP/ONC HTI-1 final rule, certified health IT containing a predictive decision support intervention must expose 31 source attributes covering how the model was developed, what data trained it, how it performs, how fairness was assessed, and how it is maintained. Evidence-based interventions require 13. Developers had to meet the criteria by December 31, 2024, with maintenance of certification obligations beginning January 1, 2025, and must keep risk analysis, risk mitigation, and governance practices in place.

AHIMA’s guidance directs provider organizations to evaluate the predictive tools they implement for fairness, appropriateness, validity, effectiveness, and safety. Someone inside the hospital has to read those source attributes, ask whether the model was validated on a population that resembles theirs, and decide whether a suggestion can be trusted for their case mix. In many health systems that person reports to health information management or CDI. AI in CDI created a governance workload with no pre-AI equivalent.

Payers automated too, and CDI defends the record

The most underestimated source of new work sits on the other side of the claim. Premier surveyed 280 hospitals across 23 states representing more than 48,000 acute care beds and analyzed 2023 claims data. The average initial denial rate was close to 15%, with some organizations as high as 49%. About 70% of denied claims were eventually paid after appeal. The cost of adjudicating a claim rose from $43.84 in 2022 to $57.23 in 2023, which Premier put at $25.7 billion nationally, with roughly $18 billion of that spent arguing over claims that arguably should have been paid the first time. Denials issued after a prior authorization had already been approved rose from 3.2% in 2022 to 10.4% in 2023.

Every overturned denial represents documentation that was defensible and got challenged anyway. Someone assembled that appeal. Increasingly it is CDI. In the 2025 ACDIS industry report, about 55% of CDI programs said they take part in the denials and appeals process, and another 8.65% said they are not involved yet but plan to be.

That is the part of the AI in CDI conversation that gets skipped. Automation on the payer side generates review volume on the provider side, and the provider side answers with clinical judgment written down by a human.

What AI in CDI changes about the CDI job

AI in CDI changes the job. It does not erase it.

Finding is the part software is genuinely good at. A prioritization engine can surface the accounts most likely to hold an opportunity far faster than a worklist sorted by admit date. What software does not do is decide whether a clinical picture supports a diagnosis, write a compliant non-leading query, hold a hallway conversation with a hospitalist who is already behind, or explain to a payer’s medical director why the record says what it says.

The 2025 ACDIS survey data points the same direction. Query rate goals of 21% to 30% were the most common reported, and 71.22% of respondents said their program has a full-time or part-time physician advisor. Programs are not thinning out the clinical layer. They are adding people whose whole contribution is judgment, which is the opposite of what you would expect inside a function being automated away.

The realistic version of AI in CDI is a specialist who covers more charts per shift, spends less time hunting and more time deciding, and carries responsibility for the accuracy of output they did not personally produce. That is a harder job than the old one, and it argues for more experienced people rather than fewer.

How to staff for AI in CDI

Programs adopting AI in CDI tend to hit the same staffing problem in the same order. The tool arrives with a business case built on efficiency. The department gets held to that efficiency before the workflow is stable. The validation, governance, and denials work shows up anyway. The gap gets covered with overtime until someone leaves.

Medovent Solutions is a US-based revenue cycle staffing and consulting firm headquartered in St. Petersburg, Florida, with no offshore delivery, working across five service lines: CDI, HIM, case management and utilization review, oncology data management, and trauma registry. Our clinical documentation integrity services include remote and onsite CDI specialists, outpatient CDI, CDI educators and auditors, interim CDI leadership, and quality and process improvement support. On the coding side, our HIM team covers inpatient and outpatient coding, HCC and risk adjustment, and external quality and compliance audits.

The practical use for a program in the middle of an AI rollout is coverage and a second set of eyes. Experienced, credentialed specialists can hold the review queue steady while the department learns the tool. Independent auditors can tell you whether the AI-assisted output is actually accurate before a payer tells you it is not. Large enough to scale, small enough to care.

Frequently Asked Questions

Will AI replace CDI specialists?

No current labor data shows that happening. The Bureau of Labor Statistics projects medical records specialists to grow 7% from 2024 to 2034 and health information technologists and medical registrars to grow 15%, even while BLS explicitly names AI adoption as a factor that may affect demand for coding work. What AI does replace is specific tasks, mainly chart prioritization and first-pass identification of documentation gaps. Clinical judgment, query construction, provider education, and denial defense stay with people.

How does AI in CDI create more work rather than less?

Three ways. AI-generated documentation has to be reviewed, and a 2025 study in Mayo Clinic Proceedings: Digital Health found a mean clinical note error rate of 26.3% across five ambient scribe platforms. Predictive AI in certified EHRs now carries federal transparency requirements, including 31 source attributes for predictive decision support interventions under the ONC HTI-1 rule, which someone at the hospital has to read and evaluate. And payers use their own automation in claims review, which raises appeal and defense volume for the CDI teams that increasingly own denials work.

Should we cut CDI headcount after implementing an AI tool?

Not before the workflow has been measured. Most staffing decisions about AI in CDI get made too early. The efficiency projected in a vendor business case usually assumes the reviewing, validating, and governance work is free. It is not. A safer sequence is to run the tool alongside existing staffing, measure query quality and denial outcomes for at least two quarters, then adjust based on numbers you observed rather than numbers you were shown.

What skills matter most for CDI teams working with AI?

The ability to disagree with the tool. That takes current clinical knowledge, current coding and query compliance knowledge, and enough understanding of how a model produces a suggestion to recognize when the suggestion does not fit the patient in front of you. Auditing skill matters more than it used to, because a larger share of the work is now checking output rather than generating it.