On February 9, 2026, the Berkeley voice-AI company Kintsugi told its remaining staff and partners it was shutting down commercial operations, closing out seven years of work and roughly $30 million in venture funding spent building a technology that, by its own published research, worked. The company’s AI could analyze as little as 20 seconds of free-form speech, in any language, and flag acoustic patterns, shifts in pitch, pauses, cadence, and energy, that correlated with depression and anxiety. It just never got the one thing it needed most: an FDA clearance that would let health systems treat it as a validated medical device rather than a wellness gadget.
Kintsugi was founded in 2019 by technologist Grace Chang and machine learning scientist Rima Seiilova-Olson, and it raised money in stages that reflected real investor confidence: $8 million in seed funding in 2021, then a $20 million Series A in 2022 from backers including Insight Partners, Acrew Capital, and Darling Ventures. The company built real commercial relationships with telehealth platforms, health systems, and payers along the way, which made the shutdown notable precisely because it was not a story of a product nobody wanted.
The science actually held up
Kintsugi did not take shortcuts on validation. It published peer-reviewed research in the Annals of Family Medicine and ran a prospective, double-blind pivotal study comparing its voice-analysis model against the SCID-5 structured clinical interview, the diagnostic gold standard in psychiatry. In one real-world deployment with a health plan, standard screening questionnaires flagged depression symptoms in just 3% of a patient population; running the same population’s voice samples through Kintsugi’s model flagged 33% for moderate-to-high depression and 14% for severe depression, a gap that researchers and clinicians involved took seriously as evidence the tool was catching cases that standard screening methods were missing, not generating false positives.
Four years, $16 million, no submission filed
Kintsugi chose to pursue the FDA’s De Novo pathway, the route for novel medical devices without an existing predicate on the market, which would have classified its software as a Class II device subject to clinical validation requirements. The company spent four years and an estimated $16 million working through multiple FDA pre-submission meetings. It never filed the actual De Novo submission. The capital ran out before the regulatory process reached a conclusion either way. Founder and CEO Grace Chang framed the decision to shut down as a choice about integrity rather than failure: “We are choosing the integrity of the science over the limitations of a distressed market,” she said, adding that Kintsugi would open-source its voice biomarker models, its underlying scientific methodologies, and its formative research so other organizations could build on the work without starting over.
Not an isolated case
Kintsugi’s collapse echoes a pattern other venture-funded clinical AI companies have hit before it. Mindstrong, a mental health startup that raised more than $100 million, shut down after pivoting away from its original diagnostic technology toward virtual care services, laying off more than 130 employees in the process. The throughline across these cases is less about whether the underlying AI works and more about timing mismatch: venture capital typically expects returns or an exit within five to seven years, while FDA clearance for a genuinely novel AI diagnostic category, especially one touching mental health, can easily take longer than that even when a company has strong data.
The case for not cutting corners
It would be easy to read Kintsugi’s story purely as an indictment of regulatory slowness, but that is not a fair fight to pick. The FDA’s own Digital Health Advisory Committee met in November 2025 specifically to weigh guardrails for generative and AI-enabled mental health devices, and researchers who study this space, including Brenda Wiederhold’s 2026 academic analysis of AI mental health regulation, have warned that without standardized, objective metrics for what counts as a safe and effective mental health AI tool, approvals risk becoming arbitrary and the market risks being flooded with subpar products making unverified clinical claims. That is precisely the failure mode the FDA’s current rigor, slow as it is, exists to prevent: a voice-analysis tool that is wrong about a vulnerable user’s depression risk is not a minor product defect, it is a potential patient safety incident. Kintsugi’s own data was strong, but strong data from one company does not resolve the harder question of what evidentiary bar every company making similar claims should have to clear.
What’s next for the technology, if not the company
With Kintsugi’s models and research now in the public domain, other startups and academic labs can build on validated groundwork without re-running years of foundational studies, potentially shortening the path for whoever tries next. But the core bottleneck Kintsugi ran into, FDA clearance timelines that outlast venture funding cycles, remains unresolved, and the FDA has still not authorized a single AI-enabled medical device specifically for mental health despite clearing more than 1,200 across other specialties like radiology and cardiology. Until that changes, or until regulators and investors find a funding model built around multi-year review timelines, Kintsugi’s exit is likely to be read by the clinical AI industry less as a cautionary tale about bad science and more as a cautionary tale about doing everything right and still running out of time.
Photo: Egor Komarov / PEXELS via Pexels
