Your AI Success Partner

Why most AI implementations fail

Most medical practices that try AI do not fail because the technology does not work. They fail because the project was set up around a tool instead of a result.

The pattern is familiar. A vendor demo looks promising, a subscription is signed, and six months later the tool is either switched off or quietly ignored. Looking across the implementations we have assessed, five causes come up again and again.

Starting with a tool. The practice asks "which AI should we buy?" before asking "which number are we trying to move?" Without a baseline (collection rate, days to the next available visit, calls answered), there is no way to know whether the tool worked, and no reason to keep paying for it.

Adding complexity. A new system with its own logins, rules and exceptions puts more on a front desk that was already at capacity. The tools that last are the ones that remove steps rather than add them.

Nobody owns the outcome. The vendor owns the software. The office manager owns the schedule. Nobody owns the result, so when the collection rate does not move, there is no one accountable for finding out why.

Dependence on one champion. Many practice AI projects live or die with one enthusiastic person. When that person leaves or gets busy, the project stalls.

Launch and leave. Payers change rules every January. Providers join and leave. Vendors push updates that quietly change behavior. An AI tool configured once and never revisited drifts out of date within months, and the first sign is usually a wave of denials.

The common thread is that implementation is roughly a third of the work. Research from BCG puts about 30 percent of AI value in tools, technology and data, and about 70 percent in people and process. The weekly monitoring, retesting after vendor updates, measuring against the baseline and keeping the tool current with the practice: that ongoing work is where the return actually comes from.

What succeeds looks different from the start. It begins with a measured baseline, picks the one workflow with the largest gap, pilots side by side with the current process before anything switches over, and assigns a named owner for the number. Done this way, AI in a medical practice is not a technology project. It is an operating discipline with technology inside it.

The AI receptionist: what patients say, and where it is heading

Patients say they want a human. Practices say they cannot staff the phones. Both are right, and the data shows the gap is narrowing.

Start with the patient view. In a 2026 survey of more than 2,000 U.S. adults, three in four said a healthcare call is stressful, and among patients who had dealt with AI on a health call, nearly a quarter gave up without getting what they needed. A clear majority want a human for sensitive matters such as test results. Patients are not wrong to be skeptical. Early phone agents misheard names, could not handle accents, and had no way to actually book a visit.

Now the practice view. The same phones are often the weakest point in the operation. In a recent assessment of an eight-provider specialty practice, roughly one in four new-patient calls was abandoned during the busiest months, live coverage ended at 4:30 pm, and an outside call center cost about $163,000 a year while patients still complained about accents and hearing. Each unanswered call is a patient who may book elsewhere.

Adoption is moving quickly in response. MGMA polling shows 83 percent of medical groups now use AI somewhere in patient visits, up from 71 percent a year earlier. Asked where they plan to add automation next, practices put scheduling (31 percent) and calls (27 percent) at the top of the list. A 2026 survey of healthcare executives found 91 percent of large practices and 72 percent of midsize practices already automate at least one front-office function, and among those using it, 79 percent cite improved patient experience as the most common outcome.

Two things changed. The current generation of voice agents speaks clearly in English and Spanish, understands natural speech, and reads and writes to the practice management system, so it can book, reschedule and confirm rather than take a message. And practices have learned how to deploy them: for routine calls first, with a clear path to a person, and with every call logged by reason so leadership can finally count demand.

The direction is not an automated front office. In that same executive survey, none of the respondents chose a predominantly automated front desk; 85 percent favored a hybrid model. The practical standard emerging is simple: every call answered, at any hour, with a human available for anything clinical or sensitive. Patients who object to AI on the phone mostly object to bad AI on the phone, and to not being able to reach anyone at all.

Three places revenue leaks in the revenue cycle, and what they are worth

Most practices believe their billing is fine because claims go out on time. In the assessments we conduct, claims going out is rarely the problem. The money is lost in the work that happens after.

Three gaps appear in almost every practice we review.

Denials that never get worked. A denied claim is not a lost claim; it is a claim waiting for someone to fix it. But working denials is tedious, and in a busy office it is the first task to slip. In a recent assessment of an eight-provider specialty practice, roughly 700 denials were open and about $500,000 of accounts receivable had gone more than 30 days without a note. A well-run office of that size has fewer than 100 open denials. Unworked denials age, and aged denials become write-offs.

Visits that were never billed. A patient is seen, the note is signed, and no charge is ever entered. It happens through scheduling edits, provider changes and visits that fall between systems. The same practice had roughly 20,000 appointments on record with no charge attached. Not every one is billable, but 1 to 3 percent of revenue in a practice this size can sit in visits that never became a bill.

Payer variance nobody is watching. Collection rates are usually reported as one number. Broken out by payer, the picture changes. In this case two major commercial payers collected at 91 percent while a Medicare Advantage plan collected at 82 percent. A gap that wide points to a contract issue, a denial pattern or systematic underpayment, and none of them will fix themselves.

What these gaps are worth. The simplest measure is net collection rate, the share of allowed payments the practice actually collects. Best-in-class practices run at 97 to 99 percent. The practice above was at 94.6 percent, and each point of collection rate was worth about $70,000 a year at its current volume. Moving from 94.6 to 97 percent would be worth roughly $140,000 to $210,000 annually, with no new patients and no new providers. Money the practice had already earned.

None of this requires new software. It requires a weekly scorecard with three numbers (net collection rate, days in A/R, open denials), a defined owner for each, and a process for working the oldest and largest items first. AI helps most in the matching work, such as reconciling appointments to charges, where a person reviewing exceptions can cover what used to take a team.