Healthcare AI consulting pays first in the administrative back office of a provider group, and the clearest evidence is the size of the paperwork. The American Medical Association's 2025 survey of 1,000 practicing physicians found that physicians complete an average of 40 prior authorizations a week, that the work consumes 13 hours of physician and staff time each week, and that 40 percent of physicians employ staff dedicated only to prior authorization. CAQH's 2025 Index puts the remaining savings from automating the medical industry's administrative transactions at $18.7 billion. Prior authorization, eligibility, claims, denials, coding support and intake are where an AI engagement can be measured and paid for. Clinical decisions are not, yet.
Where healthcare AI consulting pays
The back office of a physician practice is a set of transactions with payers, and CAQH has measured them for more than a decade. In its 2025 Index, built from data supplied by more than 600 provider organizations and health plans covering 63 percent of insured lives, medical spending on administrative transactions fell 9 percent to $75 billion. Medical providers averaged $8.03 for each transaction done by hand and $2.65 for each one done electronically, across the transaction types the Index tracks. Health plans reported that only 40 percent of prior authorizations were fully electronic, and the rest were handled partly or wholly by hand, through portals, phone calls, fax and email.
Those numbers point to five processes, each with a clear unit of work.
Prior authorization is first. A request needs the patient's coverage, the payer's rules for the procedure, and the clinical documentation the payer asks for, gathered from the EHR and entered into a payer portal or a fax form. An agent can assemble the request, submit it, track its status and bring back the decision, and a person signs off on anything that turns on clinical judgment.
Eligibility and benefit checks come next. The check is knowable before the visit, and a coverage gap found then costs a phone call rather than a denied claim. Patient intake sits on the same work: reading the insurance card, confirming the demographics, and checking coverage before the patient arrives.
Claims and denials are the third. We walked through the denial workflow, and why a denied claim loses value the longer it sits, in process redesign in healthcare revenue-cycle management. An AI engagement that works the denial queue at the rate denials arrive is paid for out of revenue that would otherwise be written off.
Coding support is the last. An agent can read the encounter note and propose codes, and flag documentation that a payer is likely to question. The certified coder makes the call. The value is in the coder's time, spent on the claims that need judgment.
Where it doesn't pay yet
Clinical decisions are a different kind of work. A diagnosis, a treatment plan or a decision about medical necessity carries liability for the clinician, and it can bring the software under the Food and Drug Administration's rules for medical devices. The evidence a buyer needs to trust a model there is a clinical study, not a before-and-after measure of cost per claim. For a mid-market provider group, that makes clinical AI a long, uncertain project with a payback the operator can't measure on a spreadsheet.
The CAQH data show where providers have started. More than half of health plans and a quarter of providers now use AI, and CAQH reports that provider use is focused on administrative tasks such as eligibility checks, note-taking and patient communications. The administrative work has a measurable baseline and a person who can review the output. That's where an operator should spend the first budget.
What the CMS prior authorization rule changes
The Centers for Medicare and Medicaid Services' Interoperability and Prior Authorization Final Rule, known as CMS-0057-F, changes the payer side of prior authorization for Medicare Advantage plans, Medicaid and CHIP programs and their managed care plans, and plans on the federally facilitated marketplaces. Beginning in 2026, most of those payers must decide urgent requests within 72 hours and standard requests within seven calendar days, and must give a specific reason when they deny one. From January 1, 2027, they must run a Prior Authorization API, so that a provider's systems can ask whether authorization is needed, see what documentation is required, and submit the request electronically. The rule does not apply to prior authorization for drugs.
For a provider group, the rule shortens the wait and makes the denial reason something a system can read. It does not do the provider's half of the work. Someone still has to gather the documentation and submit it, and CAQH found that only 20 percent of providers had begun implementing the new FHIR-based interfaces for electronic prior authorization. Employer-sponsored commercial plans and plans on state-run exchanges aren't covered by the rule at all. A practice will be working through a mix of interfaces for years, and an agent that works across APIs and portals alike fits that mix. We covered the mechanics of working through a portal with no API in writing back to legacy systems without an API.
The cost of implementing AI in healthcare
No research firm publishes a reliable benchmark for what a mid-market provider pays to put AI into its back office, and the figures that circulate come from vendors. What can be said is where the cost comes from, and the model is the smallest part of it.
Integration is the largest. The agent has to read from the EHR and the practice-management system, and it has to work in each payer's portal, which differ by payer and change without notice. Every system it touches is a connection someone has to build and maintain.
HIPAA is the second. Under the federal rule's definition of a business associate, a vendor that creates, receives, maintains or transmits protected health information on a provider's behalf, including for claims processing, billing or practice management, is a business associate. So is that vendor's subcontractor. A practice needs a business associate agreement with the AI vendor, and the vendor needs agreements with the model and hosting providers behind it. Access controls, audit logs and data retention are built into the work from the first day, and they cost time.
Human review is the third. A payer can audit a claim, and a regulator can audit a practice, so the decisions that carry clinical or financial authority stay with a person. The engagement has to design where those reviews happen and what the reviewer sees. We set out the controls that sit around an agent's actions in controls for autonomous agent actions.
How to buy healthcare AI consulting
A provider group buying this work can hold a vendor to a short sequence.
- Pick one process with a volume and a cost the practice already counts, such as prior authorizations per week or denials worked per month.
- Measure that process from system records before the work begins.
- Confirm the business associate agreements for the vendor and every subcontractor that will touch patient data.
- Agree which decisions stay with staff, and what the reviewer will see when a case reaches them.
- Tie the fee to the measured change in the process, with the integration and maintenance included.
A vendor that can't meet the second and fifth steps is selling time, and the practice carries the risk of the result. We wrote about the questions that separate the two in how to evaluate an AI delivery vendor.
A reasonable counter, answered
A reasonable counter is that payers are using AI too, so automating the provider side only speeds up a contest the provider can't win. The AMA found that six in ten physicians worry that AI may further increase denial rates, and 32 percent already report that prior authorization requests are often or always denied. The concern is fair, and no provider-side system changes a payer's policy. But the contest the provider loses today is mostly one of volume and time. A request missing a document is denied for the missing document, a denial worked past its deadline is lost by rule, and a practice that answers each request completely and on the day it arrives wins the cases that were winnable. Payer AI raises the volume of decisions a practice has to answer. That raises the value of the provider's own capacity to answer them.
Common questions
- What does a healthcare AI consulting engagement cover?
- For a provider group or physician practice, the engagements that pay soonest cover administrative work: prior authorization, eligibility and benefit checks, claim submission and denials, coding support and patient intake. The work is high in volume, follows payer rules, and is measured in every revenue-cycle benchmark. Clinical decisions, such as diagnosis or treatment choices, carry regulatory and liability questions that administrative work doesn't, and they're a poor place to start. A good engagement names the process, measures it before the work begins, and ends with a system in production rather than a recommendation.
- What is the cost of implementing AI in healthcare?
- No research firm publishes a reliable benchmark for what a mid-market provider pays to put AI into its back office, so be wary of any single figure. The cost is driven by three things: connecting to the EHR, the practice-management system and each payer's portal; meeting HIPAA, including a business associate agreement with every vendor that handles patient data; and the human review that stays on decisions a payer or a regulator will check. The software and model costs are usually the smaller part. A buyer should ask for the price tied to one process, with the integration and review work included.
- Does HIPAA allow a practice to use an AI vendor on patient data?
- Yes, under the same rules that apply to a billing company or a clearinghouse. A vendor that creates, receives, maintains or transmits protected health information on a practice's behalf is a business associate under HIPAA, and the practice needs a business associate agreement with it before any patient data moves. The rule reaches subcontractors too, so the AI vendor's own model and hosting providers have to be covered by agreements of their own. A practice should ask for the full chain of agreements, including the vendor's own.