Healthcare organizations say artificial intelligence is paying for itself fastest in the back office — billing, claims processing and other administrative work — while AI at the bedside is still struggling to earn clinicians’ trust, according to The 2026 Healthcare AI ROI Scorecard, published September 30, 2026, by venture capital firm Bessemer Venture Partners. The report is based on a survey, conducted with Bain & Company, of 226 executives at health insurers (payers), healthcare providers and pharmaceutical companies, covering 65 AI use cases. It follows Bessemer’s 2025 Healthcare AI Adoption Index, which surveyed more than 400 executives using a similar framework.

The headline finding is speed. Respondents had budgeted for AI investments to pay back in roughly 24 months, but report realizing returns in about 12, at an average of 3.5 times cost, with results beating expectations in roughly 40 percent of use cases. Fifty-four percent reported material return on investment inside the first year. Ranked by reported return, provider revenue cycle work led at 4.0 times cost, followed by payer claims and pharmaceutical preclinical discovery at 3.4 times, payer provider-network functions at 3.3 times, member engagement at 3.2 times, provider front-office work at 3.1 times, provider clinical uses at 2.9 times, pharmaceutical commercial uses at 2.6 times and clinical drug development at 2.3 times.

The gap between administrative and clinical AI shows up most clearly in how much work organizations have handed over to autonomous systems. Two-thirds of provider revenue cycle respondents now run semi- or fully autonomous AI agents, as do 62 percent of payers in member engagement, compared with just 4 percent in provider clinical work. The fastest-growing use cases were all administrative: payer contracting management rose from 19 percent to 37 percent adoption, provider credentialing and enrollment from 35 percent to 56 percent, provider contracting from 40 percent to 60 percent and prior authorization from 32 percent to 46 percent. Still, the share of use cases in full production barely moved over the year, and for payers it held flat at 6 percent.

The survey also points to workforce effects. Half of the organizations surveyed said they have already reduced headcount as a direct result of AI or plan to within six months, with cuts averaging 8 to 13 percent of the affected functions. Among payers reporting reductions, 71 percent named claims processing and 68 percent named member services; among providers, 73 percent named revenue cycle and medical billing. Bessemer estimates, as a rough calculation of its own, that this implies about $100 billion in labor-spend disruption across roughly $1 trillion in annual healthcare administrative spending. The authors also argue that healthcare is short of workers in many of these same functions, citing an American Medical Association estimate of a roughly 30 percent medical coder shortage, and suggest that nurses now staffing prior authorization and utilization review could be redeployed to care management rather than simply eliminated.

For the claims side of the business, the report highlights how much AI is being aimed at the same disputes from both directions. Providers ranked denial and appeals management as their top pain point (78 percent) and prior authorization second (61 percent), while payers placed appeals management in their top three (42 percent). Bessemer characterizes the result as “AI-to-AI combat” on mirrored workflows, and argues the larger savings would come from eliminating duplicative processes altogether through aligned incentives and faster payment — something it says AI alone cannot deliver. For scale, it points to a 2021 JAMA Viewpoint, Administrative Simplification and the Potential for Saving a Quarter-Trillion Dollars in Health Care, which estimated roughly $265 billion in potential annual administrative savings (Bessemer cites the figure as $260 billion).

Clinical AI faces steeper obstacles. Only 46 percent of respondents said they trust AI-generated tools to support clinical decision-making, and 77 percent said clinicians override AI suggestions more than half the time. Fifty-eight percent said the treating clinician bears primary accountability for AI-influenced decisions, half cited medico-legal exposure as the main barrier to scaling beyond proof of concept, and 36 percent of providers have delayed or canceled a clinical AI deployment over regulatory and compliance uncertainty. Payment is another hurdle: about 48 percent of payers said they are unwilling to reimburse fully autonomous AI care, versus about 5 percent who object to AI-assisted care in which a clinician reviews and confirms the output. Roughly 44 percent of payers expressed willingness to create payment pathways in narrower areas such as primary and preventive care, diagnostic support and medication management. Bessemer frames the stakes against workforce shortages, citing the Association of American Medical Colleges’ projection in The Complexities of Physician Supply and Demand: Projections From 2021 to 2036 of a shortfall of up to 86,000 physicians by 2036, including 20,200 to 40,400 in primary care.

Buyers are also changing how they purchase AI. Forty-two percent have completed or are running a vendor consolidation, 72 percent now have an AI governance committee, and respondents named regulatory and compliance uncertainty as the top reason deployments are delayed or canceled. More than half of next year’s incremental AI spending is earmarked for scaling proven use cases (35 percent) and integrating them into existing workflows (20 percent). Internal development has faded: 61 percent of organizations said half or fewer of their internally built AI tools are still maintained and in use.

For insurers, claims administrators and employers, the findings suggest that AI-driven automation of claims processing, prior authorization and utilization review is moving from pilots into routine operations, with real staffing consequences, while the unresolved questions around clinical AI — who is liable when a clinician follows or overrides an AI recommendation, and whether payers will pay for AI-delivered care — remain open. The survey covers group health payers, providers and pharmaceutical companies; it does not address workers’ compensation or property-casualty claims specifically.

Several caveats apply. The return figures are self-reported by respondents and reflect deployments that Bessemer says are roughly two years old; the firm presents them as a floor rather than a ceiling, but they have not been independently audited. The published article does not detail how respondents were selected or the survey’s margin of error. Bessemer is a venture investor in healthcare AI companies, and the report names portfolio companies, including Abridge and SmarterDx, among the market’s winners; the firm’s own disclaimer notes that companies discussed may be current or former portfolio companies. Characterizations of trends and opportunities in the report reflect Bessemer’s analysis and investment perspective rather than independent findings.