The 2026 Healthcare AI ROI Scorecard

Last year, 408 healthcare executives told us the opportunity set for AI was wide, and that adoption moved through their industry faster than electronic health records ever did. It was an imperative at the board level for all providers, pharma, and payers to adopt. In 2025, 84% expected GenAI to transform clinical treatment decisions within three to five years. One year later, 71% say it already has. The question is no longer whether AI takes hold in healthcare. The question is which use cases AI can be applied to most effectively, and what value it returns.

In 2026, the answer split in two, with nuances across provider, payer, and pharma cohorts. Value is concentrated in administrative work for providers and payers. Revenue cycle, claims, and the back office are where returns land, and where autonomous agents now run the work, with two-thirds of provider revenue cycle respondents running semi- or fully autonomous agents, versus 4% in clinical work. That layer has crossed into production and is returning 3-4x its cost within 12 months. For pharma, the highest adoption has landed in preclinical opportunities. Clinical AI, with the promise to increase access and reduce clinical services cost, is much earlier in value realization. It reaches proof of concept at the same rate as administrative AI and converts to full deployment at half that rate, held back not by model capability but by trust, liability, and reimbursement.

The early drivers of that return are productivity and revenue gains, and both are landing faster and larger than executives underwrote. The more consequential finding is forward-looking: most of the value still lies ahead. Executives expect materially more to be realized over the next 12 months, and the leading indicators are already visible, such as early cost-reduction capture and planned FTE reductions concentrated in administrative functions.

By and large, healthcare buyers are starting to pick their bets and their partners. Investment is shifting toward scaling what already works: 42% are consolidating vendors to simplify their stack and cut costs, and over half of next year’s incremental AI investment will go toward scaling pilots and integrating them into existing workflows. Providers and pharma are furthest along in this push to deploy at scale piloted and vetted solutions, while payers are still in ideation.

In the back office, the baton has been passed to semi- and fully autonomous agents, and the returns are clear. At the bedside, clinical AI suffers from high override and low trust because liability still sits with the providers, and today’s payment model doesn’t account for AI-delivered care. But Clinical AI isn’t short on value, which is why this is one of the many emerging opportunities for the industry.

This is what Bessemer Venture Partners and Bain & Company found in our 2026 survey of 226 executives across 65 use cases, run on last year's framework with expanded scope in clinical AI and ROI. The report covers what payers, providers, and pharma enterprises are prioritizing: where healthcare has crossed the chasm into scaled deployment, how buyers are consolidating toward best-of-platform solutions, the ROI Scorecard we built to track where value is actually landing, and the multi-trillion-dollar opportunities to replatform our healthcare system ahead.

Key insights on the state of healthcare AI adoption

• The three-year transformation forecast has become present tense. In 2025, 84% of executives expected Gen AI to transform clinical treatment decisions within three to five years. Just a year later, 71% say it already has.
• Healthcare AI ROI arrived twice as fast and at a higher level than expected. Buyers projected ~24 months, and realized it in ~12, averaging 3.5x, exceeding expectations in ~40% of use cases. 54% see material ROI inside the first year.
• More value is on the horizon because of planned cost and headcount reductions. Half of the organizations surveyed have already cut headcount as a direct result of AI or plan to within six months, at an average reduction of 8% to 13% of the affected labor for use cases. The savings booked to date are only the leading edge, not the total.
• The returns are concentrated in the back office. Revenue cycle management is leading on every measure with 4x ROI and 67% of solutions running as semi- or fully autonomous agents. Provider use cases for clinical sit at 2.9x and 4%.
• Buyers are consolidating and pivoting from building it themselves. 42% have consolidated or are consolidating AI vendors; 60% have named a primary foundation-model provider. Decay of the ‘internal build’ strategy we talked about last year. <50% of internally developed solutions are maintained and in use today.
• Innovative healthcare AI vendors have taken a significant share of the re-shift. The share of development shifted to healthcare-specific AI vendors (+12 p.p.) and AI labs (+2 p.p.), moving away from internal building (-16 p.p.). In last year’s findings, healthcare organizations were building their own; this year, most of those tools are no longer maintained.
• Clinical AI is used, but not fully trusted—looking for clarity in liability, reimbursement, and compliance. Only 46% trust AI-generated tools to support clinical decision-making, and 77% override its suggestions more than half the time. 58% say the treating clinician bears primary accountability for AI-influenced decisions. And ~48% of payers are unwilling to reimburse fully autonomous AI care, against ~5% who object to AI-assisted care with a clinician-in-the-loop. But given the imperative to have system-wide deflationary impact, stakeholders are focused on finding the right business models to realize the clinical AI promise.

What’s next: crossing the chasm into scaled deployment

2025 was the year of the PoC and pilot. In 2026, many of those pilots are completed and beginning to graduate into scaled production. The segments with the highest share increase into implementation/production are front-office provider (14%), RCM provider (10%), and member services payer (10%). So healthcare enterprises today are no longer focused on moving from proof of concept to pilot; they're focused on picking winners and scaling into production—and few have crossed that chasm. Providers are leading the scale-up.

Over half of next year's incremental AI investment will go toward deployment rather than discovery: 35% to scaling proven use cases and 20% to integrating them into existing workflows, against 13% for infrastructure and 11% for foundation model licensing.

Regulatory and compliance uncertainty will be the top driver of delayed or canceled deployments. As 72% of organizations now have an AI governance committee, purchasing, procurement, and deployment decisions run through these bodies, and decisions are increasingly centralized compared to the experimentation days of 2025. (See: Consolidation toward the best of platform, for buyer-seller dynamics)

Takeaway: AI adoption matures along the implementation curve; healthcare executives look to AI partners to realize implementation and full-scale production and deployment.

ROI showed up early—introducing the ROI Scorecard

ROI is arriving twice as fast as executives underwrote it: typically occurring within ~12 months, averaging 3.5x and exceeding expectations in ~40% of use cases. That survey finding is consequential, as nearly every AI budget approved over the past two years was built on a payback assumption that turned out to be wrong by half. Organizations still modeling two-year paybacks are underinvesting against their own results.

Every function we surveyed clears or approaches the three-times threshold that enterprise buyers typically require before funding at scale. The decision is no longer whether to fund AI; rather, the ROI ranking tells executives which function to fund first.

This is based on realized ROI, not projected ROI, and it came early—these deployments are roughly two years old. The following returns are what buyers have already booked in that window, so we see them as a floor rather than a ceiling. More is contracted to come, as the headcount reductions that convert automation into permanent cost savings are planned but largely not yet executed (See: Where automation meets attrition trends)

1. Provider revenue cycle: 4.0X.
2. Payer claims and pharma preclinical discovery: 3.4X
3. Payer provider network: 3.3X
4. Member engagement: 3.2X
5. Provider front office returns: 3.1X
6. Provider clinical: 2.9X
7. Pharma commercial: 2.6X
8. Clinical development: 2.3X

But not all returns are the same. When asked which benefits matter most for each activity, executives describe different sources of value for administrative and clinical work. This difference explains most of the ROI gap between them. In the revenue cycle, the leading drivers are improved speed to output and increased revenue, with reduced FTE cost close behind. In the commercial payer function, reduced FTE cost is the single largest driver outright. These are the benefits that land within a year’s budget, attributable to a specific system.

AI’s value to the provider looks nothing like that. So far, its ROI is concentrated in optimized decision-making and improved quality, while reduced FTE cost and increased revenue barely register. This is the context for reading the ROI Scorecard: each use case sits in a different stage of adoption, resulting in varied stages of ROI capture.

The ROI Scorecard is the successor to last year’s AI Dx Index, mapping adoption against realized ROI for every use case. Where the AI Dx showed where the opportunity was, the scorecard shows where the return is. Providers are the furthest along on both axes, Payers are mixed, and Pharma is the earliest because attributing value is genuinely hard when drug timelines dwarf a one-year measurement window.

How to read the scorecard

Each dot is a single use case, placed on two axes:

• The horizontal axis is overall adoption, carried forward from last year’s AI Dx Index definition so the two years are comparable. It’s a weighted average of where respondents sit on the development stage for that use case, where “not yet started” counts as 0% adoption and “implementation/full roll-out” counts as 80% adoption.
• The vertical axis is realized ROI, the self-reported return multiple on money already spent. For example, a 3x use case returned three times its cost.

Read together, the top-right quadrant is where a use case is both widely deployed and paying back; the top-left is where the return is proven, but the market hasn’t moved yet; the bottom half is where value hasn’t yet shown up. The top-left quadrant is worth paying attention to as it’s where the next wave of adoption is most likely to come from.

Provider

Payer

Pharma

Why the money landed in admin

Every adoption gain of consequence this year was administrative. The fastest-growing use cases in the study were payer contracting management, which nearly doubled from 19% to 37%; provider credentialing and enrollment, from 35% to 56%; provider contracting, from 40% to 60%; and prior authorization, from 32% to 46%. Not one clinical decision support use case appears near the top of that list.

Autonomy is the tell. Two-thirds of provider revenue cycle respondents now run semi- or fully autonomous agents, and 62% do so in payer member engagement. In provider clinical work, the figure is 4%. Where organizations have handed the work over entirely, they have handed over administrative work.

Takeaway: Returns are highest in mature administrative workflows and lowest in clinical.

Consolidation toward best-of-platform and efficiency gains

The era of the internal build experiment is over as organizations begin to streamline the tech stack and reduce costs by picking use cases to prioritize and invest in to scale deployments that will capture significant ROI. In 2025, teams were building a great deal of AI themselves in the era of experimentation. Today, 61% of organizations report that half or fewer of their internally built AI tools are still actively maintained and in use, and 32% report that less than a quarter survived.

The failure is not in building, but rather in scaling these PoCs to production scale and maintaining them. An internal team can produce a working MVP. Fewer can operate it at production reliability, monitor it, keep it current with a changing model, and stand behind it in an audit. Buyers have concluded that this is not work they want to own, and the share of recent development that was internal fell by 16%.

This shift maps to new expectations between buyers and sellers

• Today, healthcare AI buyers want fewer, deeper relationships: 42% of buyers have completed or are actively running a vendor consolidation. They’re doing it to simplify their stack (37%) and to reduce cost (32%), with a further 30% removing vendors that fell short on security, compliance, or performance.
• Meanwhile, 57% say they are inundated by AI companies pitching point solutions. That gap is the defining commercial tension of this phase. Buyers have moved from evaluating capabilities to buying implementations, and most of the market is still selling the former.
• The “AI winners” in healthcare differ by segment, but systems of record aren’t where healthcare organizations expect their next AI to come from.
  • The last 12 months were about consolidation, and startups and AI-native companies were the winners. Buyers are cutting their vendor lists and picking who to scale with, and healthcare-specific vendors took more share than any other category, up 12 p.p.The companies that won scaled a wedge product into a platform, including our portfolio companies Abridge and SmarterDx, or provided co-development and deployment support like Qventus. The pattern is one we described in last year’s report: begin with a single product wedge, earn the workflow, then expand into adjacent products from inside. The co-development motion we identified is the other half of the playbook, and it’s holding.
  • Looking across individual use cases, systems of record and HCIT vendors also gained share, most visibly in revenue cycle management. Large HCIT was up 7 p.p. Overall, the second-largest gain in the data, and provider RCM is where that shows up most sharply. These vendors ship off-the-shelf features into workflows they already sit in, which captures the low-hanging fruit without new procurement. They’re a real competitor for adoption today, even as buyers look elsewhere for what comes next.
  • Going forward, decision-makers look to innovative players like AI-native startups and scaled companies, as well as AI labs and deployment platforms, for next-10 solution development. Innovative partners sit ahead of internal builds and incumbent HCIT. Notably, AI labs and platforms were close to flat on solutions in adoption today, up only 2 p.p., but they rank among the preferred partners going forward, most strongly in pharma. The next wave is being spread across a wider set of innovators rather than consolidated onto any one incumbent.
• Different segments need different types of AI partners. For example, provider and payer work is workflow-shaped and regulated, which rewards an AI-native vendor who has already solved the workflow and can help scale in compliance while showing high fidelity to capture more value. Pharma's highest value problems are research problems, where the frontier model intelligence fine-tuned with proprietary data matters more than the wrapper around it. We see the biggest share gain from AI labs here, commensurate with their prioritization to partner with life sciences companies.

Takeaway: For founders and builders in Healthcare AI, the co-develop motion we identified last year is working, and it’s working because it solves the maintenance problem that killed internal development.

Where automation meets attrition trends

The clearest evidence that healthcare AI has moved past experimentation is that it now shows up in workforce planning. Half of the organizations we surveyed have already reduced headcount as a direct result of AI or plan to within six months. Where they land, reductions average 8% to 13% of the affected functions. The healthcare system spends roughly $1 trillion on admin, most of which is labor, and taking a rough 10% of that suggests $100 billion in labor-spend dislocation.

The reductions follow the returns. Among providers reporting cuts, 73% name revenue cycle and medical billing, which posts both the study's highest return at 4.0x and its highest autonomy rate at 67%. Among payers, 71% name claims processing and 68% name member services.

Pharma is the outlier. It’s least likely to cut at all (44%), and development is its largest affected function (22%). The most plausible explanation is that pharma carries less of the administrative duplication that defines the payer and provider relationship, though our data doesn’t test this directly.

In reality, healthcare has no surplus of administrative workers to shed. It has a shortage in the exact functions AI is being pointed at, and an aging workforce that is already set to widen the same gap on a slower timeline. The American Medical Association puts the medical coder shortage at roughly 30%, and the average certified coder is over 50. This means a meaningful share of the most experienced staff will exit the workforce within the next decade regardless of what AI does. Certification takes 6-18 months, and new coders need another year or two to reach full productivity. The pipeline can’t close this gap on its own.

Healthcare, like every other industry absorbing AI into its workforce, will have to adapt, retrain, and elevate skillsets rather than simply subtract roles. The distinctive opportunity here is that healthcare is short of the people it’s automating against. A large share of administrative work is performed by staff with clinical training or clinical-adjacent capability, such as nurses in prior authorization and utilization review. When an agent takes over the queue, the question isn’t just what the savings are but where that capability gets redeployed. Moving a utilization review nurse back to care management converts an FTE reduction in one cost center into capacity in a function the organization cannot currently hire into.

The multi-trillion-dollar opportunities left unclaimed

While our survey shows where healthcare AI has already delivered, there is still a lot of work left to accomplish system-wide impact. Three opportunities account for most of it and vary across the adoption curve, each falling short for different reasons.

The $1T opportunity: paying for the same fight twice

Roughly $1 trillion a year goes to healthcare administration, with an estimated $260 billion considered waste. Much of that isn’t inefficiency inside any single organization. This is the cost of two organizations working against each other.

The pain point data makes this clear. Denial and appeals management is the single largest pain point that providers report at 78%. On the payer side, appeals management ranks in the top three at 42%. Prior authorization is the provider’s second-largest pain point at 61%, while medical necessity and claims adjudication, appeals management, and out-of-network negotiations occupy the payer’s list. These are essentially the same dispute on both sides, staffed and paid for twice.

This friction is also where AI has moved fastest, but the impact so far lands on a single stakeholder's productivity and value capture, and what we’re really seeing today is just AI-to-AI combat: prior auth and utilization management running mirrored workflows on either side of the same dispute. Automating both sides of a duplicative process shouldn’t be the goal because it simply produces a faster, cheaper version of the same fight. Under aligned incentives and real-time payments, most of this work wouldn’t need to exist at all. The near-term value is automation, and the trillion-dollar value is in making the workflow unnecessary. That requires scale, AI companies that can serve both sides to become interstitial tissue between payer and provider, and an incentive alignment that AI alone can’t deliver.

The $3T opportunity: four blockers for clinical AI

The delivery of care costs around $3 trillion, growing at ~5% year-over-year, against a clinical workforce that’s already short of demand and getting shorter. The Association of American Medical Colleges (AAMC) projects a shortfall of up to 86,000 physicians by 2036, including 20,200 to 40,400 in primary care. This is the larger opportunity and by far the earliest one. Only 4% of provider clinical solutions run as semi- or fully autonomous agents, and clinical returns are 2.9x for providers and 2.3x for pharma, which are some of the lowest figures in our data.

Four components are holding clinical back—

The first is trust. 46% of respondents trust AI-generated tools to support clinical decision-making, and 77% override the AI’s suggestions more than half the time. Clinical AI is in the workflow without being relied on. What would change that is specific and buildable:

• Stronger clinical validation evidence (67%)
• Transparent recommendation rationale (53%)
• Patient-specific performance data (47%)
• Seamless EHR integration (44%)

Second, it’s liability.58% of providers say the treating clinician bears primary responsibility for AI-influenced clinical decisions, and 50% say medico-legal exposure is what keeps AI from scaling past PoC. 36% have either delayed or canceled a clinical AI deployment outright over regulatory and compliance uncertainty. From this perspective, the 77% override isn’t an irrational caution. It’s the predictable behavior of professionals who carry the risk personally.

Third is regulatory and compliance. 36% of providers have delayed or canceled clinical AI deployments outright over regulatory and compliance uncertainty. There is no established regulatory path for pan-indication clinical intelligence and no settled post-market monitoring regime for semi-autonomous and fully autonomous clinical AI.

Then, there’s payment. Fee-for-service pays clinicians for time and specific actions. If AI does the work, or if the patient is treated without a clinician’s time at all, there’s currently no existing mechanism to be compensated. 48% of payers are unwilling to reimburse fully autonomous AI care, while only 5% object to AI-assisted care in which a clinician reviews and confirms the output. Human-in-the-loop is the reimbursable architecture for the near term.

The good news is that payers have higher intent to pay for narrow areas for cost management. Primary and preventive care, AI diagnostic support, and medication management each draw roughly ~44% payer willingness to establish a payment pathway. We expect CMS and the national payers will need to move before the broader market follows.

The $0.5T drug development and distribution opportunity: earlier ROI from bench to bedside

The US spends $500 billion a year on drugs and therapeutics, and pharma spends $150 billion per year on R&D. AI’s biggest impact to date sits at the front of that pipeline: preclinical workflows lead pharma on both adoption and return compared to clinical development and commercialization. Preclinical intelligence is where most of the frontier intelligence has been focused, and where we see the clearest fruit and the strongest appetite for more.

Preclinical development, which has been the focus of AI so far, captures only 25% of R&D spend. The harder question sits downstream. The more selected candidates graduate into the clinic, the more value shifts to raising clinical trial success rates and distributing and widening patient access to the drugs that clear them. Both are earlier in ROI capture than the administrative use cases in this report, and for a structural rather than technical reason: the timeline from bench to bedside and the multifactorial nature of moving a drug through it dwarf any annual measurement window.

What these opportunities tell builders and buyers

The throughline of this year’s data is that healthcare AI stopped being an experiment and became a reality. The landing-and-expanding strategy we described last year is serving health AI startups well. With the speed of product development and the incremental capability gains of the last 12 months, AI-first companies are taking share of spend, strategy, and partnership with enterprises. It’s wartime, and the biggest winners are the companies building fast, eating into adjacent workflows, and compounding trust with their enterprise partners. The last point matters most from here because the same trust that was sufficient for administrative work is now the binding constraint on clinical deployment.

Outside of healthcare, teams deploying AI agents into production have already converged on a standard set of supporting infrastructure around the model. This harness infrastructure includes evals, performance monitoring, guardrails, and agent orchestration, which in healthcare, maps almost exactly onto what clinicians said they would need to trust AI’s output (validation evidence, transportation rationale, patient-specific performance data) and onto what a regulator will eventually ask for in post-market surveillance. Those who build this layer deliberately will more easily scale past pilot in clinical and admin settings.