Pelago, the specialty substance-use platform formerly known as Quit Genius, announced a new behavioral health platform this week that folds substance use, mental health, and behavioral addiction care onto a single employer contract and a single member record. Every member entering the platform now goes first through Sona, Pelago’s voice-based AI tool, which assesses need and recommends a level of care, with the company describing licensed clinicians as retaining supervisory authority over the process. What the public evidence does and does not establish about that assessment, both what members report afterward and whether the routing itself is accurate, is now central to evaluating the platform Pelago is selling to employers.
What launched
Pelago’s substance-use business has run on an acuity-matching model since it operated as Quit Genius: assess what a member needs, start with the least intensive support that works, and escalate when it doesn’t. This week’s launch extends that same logic across mental health and behavioral addictions like gambling and binge eating, with Sona as the shared front door for all three programs. Pelago’s own release frames the model as a corrective to a default it says routes “nearly everyone” into one-on-one therapy “regardless of how acute their needs are,” adding that “most people seeking help don’t need that level of support.” That is Pelago’s characterization of the problem it says it is solving, not an independently confirmed measure of the new platform’s own routing pattern; the company has not published what share of Sona-assessed members are directed to each level of care.
What Pelago’s economic evidence shows
Pelago has published disclosed economic evidence for its underlying substance-use model, across three separate analyses that differ in rigor, population, and review status, and the September platform announcement does not make those differences clear.
The earliest is an August 2023 claims analysis: 2,670 members with tobacco and alcohol use disorders, compared against a propensity-matched control group of non-participants, reporting $9,367 in average annual medical claims reduction and a 3.0-to-1 return on investment over 14 months, with the methodology reviewed by an unnamed third-party risk-management firm.
The strongest is a 2025 study published in the peer-reviewed journal Telemedicine and e-Health: 1,172 Pelago members with alcohol, tobacco, or opioid use disorders, enrolled between January 2021 and February 2024, matched to 6,414 control participants on age, gender, and region, using a difference-in-differences model to adjust for baseline cost gaps between the groups. It reported a 33 percent reduction in all-cause medical costs, an average of $6,758 in savings per participant over 12 months, and a 4.5-to-1 return on investment, with sensitivity testing showing the result held after removing high-cost outliers. This is the only one of the three that went through independent academic peer review before publication.
The most recent is a March 2026 analysis Pelago describes as independently conducted by Aon, the professional-services and actuarial firm: 1,132 Pelago members evaluated, using Aon’s own Cost Efficiency Measurement methodology, against a control group drawn from a multi-employer claims database, reporting $11,829 in annual medical cost savings per member and a return exceeding 5 to 1. Pelago describes this analysis as validating the 2025 peer-reviewed study’s findings. A commissioned actuarial analysis by an outside firm and a study accepted through academic peer review are both forms of external scrutiny, but they are not the same process, and Pelago’s release does not describe Aon’s underlying data or matching criteria in the level of detail the 2025 journal article discloses.
The September platform launch cites “nearly $12,000 in annual savings per participant” without naming a study. That figure is close enough to the March Aon result that it may be the same number, rounded, or it may be a different calculation; Pelago’s release does not say which. What no version of the economic evidence covers is the platform launching this month: the 2023, 2025, and March 2026 analyses all measure the existing substance-use program alone, not the combined three-program platform Sona now triages.
Symptom improvement is not the same question as accurate routing
Sona has been live with members since this summer, and Pelago has released two kinds of findings about it that the platform announcement runs together, though they answer different questions. One is whether members who talked with Sona later reported improvement in depression or anxiety symptoms. The other is whether Sona accurately identifies who needs a human therapist, who can be helped with lighter support, who needs specialty substance-use or behavioral-addiction care, and who needs immediate escalation. The September release provides evidence only on the first question and treats it as though it answers the second.
On symptom outcomes, Pelago has published two different accounts of what appears to be the same underlying pilot, in two separate documents, with two different sets of summary statistics that are never reconciled with each other. The company’s launch release states that two sequential observational studies covering more than 9,000 member conversations found that, among participants who began with clinical-range depression or anxiety scores and completed repeat assessments, close to half improved by at least five points on the PHQ-9 or GAD-7 within about a month, with improvement outnumbering worsening by more than 13 to 1, and 20 to 25 percent fewer self-reported poor mental health days per month. That release names no journal and states no peer-review status for those figures. A separate company blog post describes what reads as the same pilot differently: 71 percent of members with baseline depression or anxiety reducing at least one severity tier, with a 53 percent average symptom-score reduction, and states that this research is under peer review at JMIR. Neither document discloses a sample size, a control or comparison group, or how many total member conversations sit behind the subset who completed repeat assessments, a detail that matters because a study built only on people who returned for a second check-in will tend to look better than one that also accounts for everyone who did not return. A separate study of Sona’s suicidality-detection and escalation protocol, submitted for peer review on its own track, reports that 13 percent of participants endorsed suicidal ideation, with a clinician notified an average of 11 minutes after detection.
None of that is evidence about routing accuracy. An uncontrolled study of symptom change cannot show whether Sona correctly sorted a given member into therapy, coaching, specialty care, or escalation, only that some members who talked with it later reported feeling better, something an observational design cannot separate from regression toward the mean, the effect of being asked repeatedly how one feels, or concurrent care from elsewhere. Pelago has not published a study measuring whether Sona’s care-level recommendations match what a clinician would independently determine a member needed, and nothing in the public record establishes that they do. That is a specific, distinct evidence gap, not a restatement of the symptom-data limitations above.
How much of the assessment is automated, and how much is reviewed
Pelago describes Sona as escalating to a human clinician “the moment a situation calls for one,” language that points to risk-triggered review rather than universal review of every assessment. The company also says its clinicians “supervise Sona’s operation on an ongoing basis, reviewing performance, refining protocols, and retaining authority over every clinical decision.” Read together, those two statements describe a system that generates the initial assessment and care-level recommendation for routine cases on its own, with clinician involvement concentrated in flagged cases and in ongoing oversight of the system’s protocols and performance, not in individually signing off on every non-flagged routing decision before it takes effect. Pelago’s public materials are explicit that Sona does not diagnose or prescribe treatment. They are less explicit about whether a licensed clinician reviews a specific member’s care-level assignment before that member is routed, when nothing has triggered an escalation.
Why this matters for employers
For an employer signing this contract, the pitch is that one AI-supported intake step now governs access to three previously separate lines of behavioral health spending, in place of three separate coordination problems, one contract, and one member record. Whether that produces the savings and outcomes Pelago describes is not yet something the public evidence answers for the platform as constituted. The company has disclosed evidence, including one independently peer-reviewed study, that its acuity-matching model has worked for substance-use care specifically. It does not yet have comparable evidence that the same logic, run through Sona across three programs at once, allocates members correctly or saves what the company says it saves. Pelago is not the only employer-facing behavioral health vendor making an AI-related move this quarter: Rula Health announced an external AI ethics and safety advisory council the same day, a governance step distinct from Pelago’s operational deployment of AI-supported assessment and routing.
What to watch
Four things are worth tracking separately, because conflating them is exactly the problem with how the September announcement presents its evidence. Whether Sona’s observational symptom research is published, at JMIR or elsewhere, with a disclosed sample size, comparison design, and the denominator behind “completed repeat assessments.” Whether Pelago or an outside party publishes anything that measures Sona’s care-routing accuracy specifically, not symptom change alone. Whether an economic analysis appears for the combined three-program platform, with the same level of methodological disclosure as the 2025 peer-reviewed study, instead of a reused or unexplained savings figure. And what happens when member access to the full platform begins in November, ahead of the January 1 renewal cycle when large employers finalize next year’s benefits contracts.