Arizona Built an AI Front Door for Opioid Treatment. It Inherited the Same Broken Directory.

Bill Ayars built his first list on a legal pad. His daughter Jennifer, who used drugs, died of an overdose in Cleveland in 2016, and in the weeks after her funeral he started calling every number the state’s official directory had given her — numbers he now knew, because he’d tried them himself while she was alive, didn’t all lead somewhere real. Some rang to voicemail boxes that were full. Some belonged to programs that had closed. He kept calling anyway, building his own list of places that actually answered, actually had beds, actually did what the directory said they did. That list became a nonprofit. It now serves more than 1,200 providers and over 200,000 visitors a year in Ohio, and it exists because a father decided the official version couldn’t be trusted.

Ten years later, Arizona just partnered with Google to solve the same problem with better technology.

AHCCCS just gave Arizona’s opioid treatment directory a much better interface. It did not, yet, prove the directory underneath is any more honest than the one that made Bill Ayars start calling numbers himself.

What actually launched

On July 14, 2026, the Arizona Health Care Cost Containment System — the state’s Medicaid agency — announced a generative-AI-powered Opioid Use Disorder Service Provider Locator, built with Google Public Sector and the data-platform company Syntasa. The pitch is a genuine departure from the clunky, form-field, drop-down-menu locators that have defined this category since the early internet: a chatbot running on Google’s Vertex AI platform and Gemini models that takes a plain-language question — “where can I find treatment near me,” “is there a program that takes Medicaid in Yuma” — and answers it conversationally, in whichever language the person types in, with Google Maps pins layered on top.

Kate Dobler, who coordinates AHCCCS’s Women’s Services Network as part of the state’s Opioid Treatment Authority, put the goal plainly: “Finding treatment should never be complicated.” Reymund Dumlao, Google Public Sector’s director for state and local government, framed it as a partnership built to reduce friction and stigma at the exact moment someone is ready to act. By the state’s own numbers, it’s working as an on-ramp: since launch, the tool has logged more than 100,000 page views, reached over 20,000 unique individuals, and been used in more than 120 Arizona cities, with more than a third of sessions starting on a phone — which matters, because a phone in a moment of crisis is a different device than a desktop at a case manager’s office. You already know that difference if you’ve ever been the one making the call.

That’s a real reach number for a state where, by AHCCCS’s own account, fewer than one in twenty people with a substance use disorder receive treatment in a given year. A tool that gets 20,000 people to a treatment question at all is not nothing.

The list under the chatbot is the whole ballgame

Here is the part that separates this launch from the ones that have failed before it, and it’s easy to miss because it’s not the part AHCCCS put in the headline: the AI locator doesn’t search the sprawling national facility database that has burned every locator before it. It draws from a curated set of roughly 100 vetted OUD providers that AHCCCS itself maintains. That’s a real design choice, and on its face, it’s the right one. A hundred providers a state agency actually checked beats thirteen thousand a federal database inherited from decades of self-reported survey data.

Compare that to the tool it’s implicitly competing with. FindTreatment.gov, the national locator SAMHSA has run since 2019, lists more than 13,000 licensed facilities and logs roughly 300,000 page views a month — the default most people land on when they search “rehab near me.” The HHS Office of Inspector General opened an audit specifically to determine whether the information on that locator is accurate, complete, and timely — the kind of question a federal watchdog doesn’t ask about a tool that’s working. KFF Health News’s investigation into the same locator quoted Dr. Cara Poland, an addiction medicine physician at Michigan State, describing her own patients — real people, mid-search for a bed, not a hypothetical — who called listed numbers and reached disconnected lines, facilities no longer taking new patients, and clinicians who had retired before the listing was ever updated. “It’s being treated as a gold-standard tool,” Poland said, “but it’s not… it’s scary, because if you go to the site, it’s got invalid information.” Jonathan Stoltman of the Opioid Policy Institute put the stakes in five words: “It’s crucial to get this right.” Frank Greenagel, who has spent years inside state licensure systems, explained why the data rots in the first place — state agencies verifying these listings, he said, “check only a facility’s paperwork,” not whether the facility is actually open, actually staffed, or actually taking the patient on the other end of the line.

“It’s being treated as a gold-standard tool,” Poland said, “but it’s not…

The failure mode isn’t hypothetical or abstract. In Michigan, reporting found that the state’s own locator kept a closed facility, Betsy’s Place in Caro, listed for a period even after it shut down — while Wolverine Human Services, a youth detention center, remained listed as though it were a treatment option. That is not a rounding error in a spreadsheet. That is a parent, at the worst moment of their life, being told by a government website that a juvenile lockup is a place to send their kid for care.

We already ran this experiment, and the answer was no

This isn’t the first time American health policy has assumed that publishing a list of providers is the same thing as guaranteeing access to them. In 2018 and 2019, researchers ran what amounted to a secret-shopper audit of the country’s buprenorphine treatment network — the actual doctors listed as certified to prescribe the gold-standard medication for opioid use disorder. Posing as patients, callers contacted 546 publicly listed prescribers across six of the hardest-hit states — Massachusetts, Maryland, New Hampshire, West Virginia, Ohio, and Washington, D.C. They reached a scheduler on 78% of calls. Of those, only 54% of callers posing as Medicaid patients were offered an actual new-patient appointment. Uninsured callers did slightly better, at 62%. Translate that out of percentages: for every ten people on Medicaid who found a “certified” provider on the list and called, roughly four to five hit a wall the list never warned them about.

That study is six years old now, predates every generative AI product mentioned in this article, and reached the same conclusion Cara Poland reached about FindTreatment.gov and Michigan Advance reached about Michigan’s locator: a directory is only as honest as the last time someone actually called every number on it. Technology has changed the interface three times since 2019. It has not yet changed that underlying fact.

What a confident voice costs you

If you’ve ever been the one making that call — not researching it, not case-managing someone else through it, actually dialing the number with your own hand because you or someone you love needed a bed that night — you already know the particular kind of damage a wrong answer does. It isn’t neutral. A paper list that’s wrong wastes your time. A chatbot that’s wrong does something worse: it answers you in a warm, complete sentence, with a pin on a map and a name and an address, and that confidence is exactly what makes you stop calling around to double-check. The old form-field locator at least felt uncertain — a wall of links you had to sort through yourself. A conversational AI interface is designed to feel certain. When the underlying list is wrong, that certainty is the product working exactly as built, and it’s aimed at someone in the worst hour of their week.

That’s not an argument against Arizona’s tool. A hundred agency-vetted providers is a genuinely more defensible starting point than 13,000 self-reported ones, and if AHCCCS keeps that list current the way Bill Ayars keeps his — someone on staff actually calling the numbers, actually confirming the beds, on a cadence measured in weeks and not years — this could be the rare treatment locator that earns the trust its interface projects. AHCCCS hasn’t published how often the underlying provider list gets re-verified, or by whom, or what happens when a provider stops taking new patients between checks. That’s the one number the state hasn’t given anyone yet, and it’s the only number that actually matters.

If you’re the one making the call tonight, and the AI or the website or the pamphlet says “yes, they have a bed” — Arizona’s opioid settlement funds are still flowing into naloxone and treatment access statewide regardless of any one tool’s accuracy, and calling the facility directly to confirm, out loud, before you drive there, costs you two minutes and has saved people a wasted trip on a night they didn’t have one to spare.

Bill Ayars didn’t build a better locator because he had a better algorithm. He built one because he was willing to call every number himself and throw out the ones that lied. Arizona just built the fastest, most fluent way yet to ask a directory a question. Whether it’s finally one worth trusting is a decision that gets made in the unglamorous work of someone, somewhere, still picking up the phone.

Arizona just built the fastest, most fluent way yet to ask a directory a question.