A profile of Onos Health, the behavioral health AI startup selling payers a single system for prior authorization, quality review, and program integrity.
Key Takeaways
- Onos Health sells health plans an AI “infrastructure layer” for behavioral health, structuring the raw clinical documentation that claims data leaves out. The company says on its website that it works with three of the six largest health plans in the country.
- Founder and Chief Executive Officer Akshay Agrawal frames the product less as eliminating prior authorization than as redirecting it, so that trusted providers are rewarded for quality care while concentrating review where low-quality care appears. He says the system surfaces information for reviewers so they can make data-backed clinical decisions, with the clinician remaining the decision maker and the system providing decision support.
- Onos recently closed a Series A round, its second raise in under a year, following a 2025 seed round co-led by Haystack and Pathlight Ventures. The company says the new capital is meant to keep pace with demand from national health plans.
- The company is scaling into a period of intense payer and federal scrutiny of behavioral health, including a new CMS autism-services toolkit and a wave of Medicaid ABA rate cuts and audits. That oversight environment is precisely the market Onos is selling into.
Akshay Agrawal took the call with Acuity from a conference room in San Francisco, where his company, Onos Health, is based. During our call, we discussed the complex systems American health plans use to manage behavioral health. Agrawal argues that, despite significant investment and effort, these systems often struggle to deliver the consistency and quality that plans are working toward, while creating significant administrative abrasion for providers.
Outside of routine outpatient therapy, he notes, nearly every behavioral health service a plan covers runs through authorization and quality clinical reviews. To administer that, plans hire large clinical teams whose days are spent, case by case, deciding whether treatment is appropriate, whether a member is on the right clinical pathway, and whether the care is any good. The word Agrawal returns to for the result is “abrasion”: “a lot of abrasion for members, a lot of abrasion for providers,” and costs that keep climbing anyway.
Onos Health, which Agrawal founded in 2024 with Suhaas Prasad and Josh Levitan and now runs as Founder and Chief Executive Officer, proposes to streamline key parts of that machinery. The company describes itself as “clinical AI infrastructure for health plans,” a single system in which a plan’s behavioral health clinicians would run prior authorization, quality reviews, provider benchmarking, and fraud detection.
The company recently closed a Series A round, its second raise in under a year, and it is arriving at a conspicuous moment: behavioral health has become one of the fastest-rising costs in American insurance, and the scrutiny that follows the money has landed hardest on the very services Onos was built to help plans optimize.
A Single System for Behavioral Health
Agrawal resists the framing of Onos as a single product. “We view ourselves as an infrastructure layer that all clinical workflows build off of,” he says. In practice, the company sits atop a plan’s existing data, ingesting not only claims and utilization records but also the messier material underneath: session notes, treatment plans, authorization packets, the unstructured documentation where, in his telling, the real story lives. The system structures that material and hands it back to clinicians, who then work out of Onos to make more informed decisions.
Two things, Agrawal says, set the approach apart. The first is focus. “We’re only behavioral health,” he says; the models are trained on behavioral health data alone, on the theory that a general-purpose medical tool cannot read a treatment plan the way a specialist can. The second is where the company looks for quality. Most payers, he argues, judge care through claims and utilization data, counting units and dollars. “Quality benchmarks and quality nuances actually live in clinical documentation,” he says: whether a member is progressing, whether assessment scores are improving, whether goals are being met. Claims capture “a very small part of the picture.”
Who defines quality is the harder question, and Agrawal is candid that the field lacks a common yardstick. “There aren’t clearly defined metrics on quality consistently,” he says. Onos has its own view, configures it to each plan’s preferences, and partners with outside organizations that hold “stronger perspectives” on what good care looks like, even as the wider industry still argues over whether standardized outcome data, more than any technology, is the real test. The company says its models were trained with the help of software engineers, data scientists, and behavioral health clinicians, guided by a clinical advisory board of Chief Medical Officers from large national plans.
Onos does not disclose most of its customers, but it says on its website that it works with three of the six largest health plans in the country. It reports early results it attributes to those deployments, including a 35 percent improvement in adherence to clinical guidelines, a figure it has not paired with a published methodology.
Rethinking Prior Authorization
Early in the conversation, offered the shorthand that Onos “eliminates” prior authorization, Agrawal pushed back. “That’s not quite it,” he said. Prior authorization is not going away in the near future; the point, he argues, is to aim it. “You want to get rid of prior authorization in areas where high-quality care is occurring,” he said, “but refocus those efforts on areas where low-quality care is occurring.” For the providers delivering high-quality care, he poses a question that already has a name in insurance: “What if we gold card those providers?”
Gold carding, the practice of exempting providers with strong track records from routine prior authorization, is not a new idea, and several states have enacted versions of it in recent years. What Onos proposes is to make the sorting continuous and data-driven, using the clinical record rather than approval-rate history to decide who gets waved through and whose cases get a closer look. Agrawal calls the current alternative burdensome: a plan leverages prior authorization and utilization management to address low-value care, but this can often result in care delays for members.
The burden he is describing is real and documented. Behavioral health carries a heavier prior authorization load than most of medicine. A late-2024 American Medical Association survey found physicians and their staff completing an average of 39 prior authorizations per week and spending about 13 hours on them, with the overwhelming majority reporting that the process delays care and drives up overall utilization. A Johns Hopkins review published in 2025 tied prior authorization delays to measurable patient harm, and found that in behavioral health the delays were linked to treatment interruptions and worse outcomes for psychiatric and substance use patients. In June 2025, major insurers pledged to federal health officials to scale back prior authorization across health care, including behavioral health.
The unresolved question, for providers, is the one buried in Agrawal’s framing: who decides what counts as low-quality, and how a provider learns of, or contests, being sorted into that column. It is the same question now moving through Medicaid programs, where a proposed California utilization-management overhaul and a broader tightening of prior authorization and medical-necessity review across behavioral health have already begun to redraw the rules that govern day-to-day authorizations.
Scaling Into a Scrutiny Moment
Onos could hardly have picked a more charged moment to scale. On August 4, 2026, the Centers for Medicare and Medicaid Services released a 173-page toolkit urging states to tighten oversight of applied behavior analysis, the dominant autism therapy and one of Medicaid’s fastest-growing line items. By the agency’s own accounting, Medicaid and CHIP spending on ABA rose 421 percent between 2021 and 2025, from roughly 2 billion dollars to about 10.1 billion, far outpacing growth in the number of children receiving services. Federal audits have flagged close to 200 million dollars in improper ABA payments across several states, and the toolkit points states toward exactly the levers Onos sells: prior authorization thresholds, tighter documentation, and outcomes tracking.
The pressure is not only federal. Centene, the country’s largest Medicaid managed care organization, said in late 2025 that it was forming a task force on ABA spending. States have moved on both rates and rules, from New York’s phased reduction of its Medicaid ABA rate to a commercial rate cut in Georgia, with hour caps and phasedowns landing in other states as well. For plans, tooling that can sift clinical records at scale is suddenly less a luxury than an operational necessity, which is roughly the market Onos is describing.
That same capability raises the question that shadows any AI system placed between a patient and coverage: what happens when it is wrong. Agrawal’s answer is that Onos is never the one deciding. “We never make any clinical decisions with just AI,” he says. “It’s just not something we do.” It is a stance that mirrors where much of the field has landed, that AI should augment behavioral health clinicians rather than replace them. The system, he says, surfaces documentation, points clinicians to the passage they need, and pulls the quality metric they are looking for, but the human at the plan makes the call. Accuracy can be measured several ways, Agrawal says; on retrieval accuracy, which he calls one of the best measures, Onos reports more than 98 percent on behavioral-health-specific clinical data points. He frames the tool as “an enablement function” for clinical staff rather than a decision-maker. On security he was more concrete: Onos is SOC 2 Type 2 certified, keeps its data onshore, employs no offshore staff, and, he says, is built so that “nothing ever leaves the four walls” of a plan’s system.
What Agrawal describes wanting next is, in a sense, the other side of the same bridge. Today Onos sells to payers; he talks about eventually working with providers directly, helping them understand what a given plan is looking for before they submit, increasing transparency on both clinical and procedural expectations so that the two sides “interact more seamlessly.” The end state he sketches is a single shared view of a member, “one single care team to understand what is optimal care for an individual.” The Series A, he says, is meant simply to keep up with demand.
Whether that reads as relief or as oversight depends on where one sits. To a plan, Onos promises to make an overwhelmed clinical operation faster and more focused. To a provider, the same system reads the chart, synthesizes existing data, and helps plans decide who gets waved through and who gets a closer look. Onos’s wager is that the answer to behavioral health’s quality problem was already sitting in the clinical record, waiting to be read at scale.
