AI assessment reports in ABA are moving fast. Two private equity-backed providers asked a vendor to remove the clinician review step entirely.
Key Takeaways
- The request came from inside the industry: Two large private equity-backed ABA providers asked a vendor to finish an AI assessment tool so the clinician on the case could submit the report without reading it. The vendor built it, then declined to ship it.
- The phrase that carried the request was “insurance ready”: It sounded like a formatting requirement and functioned as a removal of the clinical review step. Adam Ventura says his team heard it as the former until the room went quiet.
- Federal auditors are already reading for exactly this pattern: The Office of Inspector General flagged identical or near-identical notes across clients, notes signed before the session ended, and narratives too vague to substantiate treatment. An unread generated document produces all three signatures at once.
- The refusal does not settle the question: Ventura declined, and the demand remains in the market for whoever builds next. Operators evaluating documentation tools should be asking who reads the output and what proves it.
The tell was the silence. Adam Ventura was on a video call with his own team about a year and a half ago, having just delivered what two large clients asked for, and nobody said anything. “Everyone at our company, we all had a moment where it got quiet in this virtual meeting,” he said at a panel during ABA CARES in Boston earlier this month, “and we all just kind of sat there, and we were all thinking the same thing. We’re like, is this a good idea?”
What they had built was a report that no longer required a human to look at it. Ventura, a Board Certified Behavior Analyst who founded and exited the multi-state provider World Evolve in 2018, now runs Intraverbal AI, which makes behavior-analysis-specific tools on top of large language models. One of them handles initial assessments. It transcribes the indirect assessment, the rating scales and checklists and parent interviews. It ingests intake information and summarizes what matters. It suggests questions to ask during the interview. Then it assembles the whole thing into the format payers want.
The category is crowded and moving fast. CentralReach markets NoteDraftAI and RethinkBH markets Session Note AI, both of which draft session documentation from collected data for a clinician to review. Mosaic Pediatric Therapy’s lead data scientist has predicted that operators will build their own tools in-house within five years, and a MassABA panel spent an hour last spring sorting which of these workflows the ethics code and the Illinois disclosure law actually permit. Buyers are choosing among tools faster than anyone is auditing them.
That product already existed and already saved time. What the two clients wanted was the last inch. “We started working with two very big private equity-backed companies,” Ventura said, “and they’re like, we need you to just create the report so that it’s insurance ready.”
What “Insurance Ready” Asked an AI Assessment Tool to Remove
Ventura is candid that the phrase did not register as a problem at first. “That sounded like a great phrase, insurance ready,” he said. “When I heard that, I’m like, yeah, we’ll make it insurance ready.” His team built it. Then they sat in the silence and worked out what they had made: “Should we create a document that is insurance ready that really requires the BCBA on the case not to look at it at all, just to press buttons and then send it over to someone?”
They went back to both companies and said no. Ventura frames the decision as a cost question rather than a moral one, which serves operators weighing the same tools better than a lecture would. “So, do we want to save their time? Yes, absolutely,” he said. “But we also want them to put their human voice on this. I can’t think of a field that’s more human than ours.”
Neither company was named, and Acuity has not independently confirmed their identities. The shape of the ask is verifiable, and Ventura described it on the record in a room full of the operators it came from.
Why OIG Documentation Audits Turn on Who Reads the Report
The reason this matters beyond one vendor’s conscience is that federal auditors have spent the past year reading ABA documentation for signs that nobody was home. Office of Inspector General audits across four states produced more than $123 million in federal recoupment recommendations, and the findings were operational rather than clinical. Auditors flagged identical or near-identical notes across sessions and across clients. They flagged notes referencing the wrong client, typically a copy-paste artifact. They flagged narratives that recorded generic activity without specific targets, prompts, or data. And they flagged notes signed before the scheduled session ended, on the straightforward logic that a document signed early cannot authenticate what it describes.
A generated assessment report that a clinician submits without reading is a machine for producing all of those patterns at scale and with perfect consistency. The document will be well formatted. It will use the right words. It will also be, in the sense auditors care about, unattested. The signature block will say a qualified professional reviewed a document that a qualified professional did not review.
The commercial pressure behind the request is not mysterious. Clinician hours are the scarcest input in this business and the supply of certified analysts has not kept pace with demand for them, which squeezes hardest at the organizations carrying the most caseloads. Private equity firms acquired 574 autism service delivery sites across 42 states between 2015 and 2024 (the JAMA Pediatrics research letter behind that count is publicly available). At a hundred locations, a saved hour per assessment is a line in a model.
The enforcement backdrop is not hypothetical either. Massachusetts providers are facing six-figure recoupments tied to supervision documentation, Indiana has rebuilt its credentialing and supervision infrastructure by bulletin, and the broader accountability era arriving for autism services assumes a level of documentation discipline that an unread report cannot supply. A tool that makes the paperwork faster and the attestation emptier moves a provider in exactly the wrong direction against that trend.
Where the AI in ABA Panel Agreed, and Where It Split
Ventura told the story alongside three panelists who broadly agreed with him and disagreed about how much that agreement is worth. Shelby Dorsey, a Board Certified Behavior Analyst who uses AI extensively in her own practice, described the discipline she tries to teach: keep the technology in the category of tool. “The screwdriver is doing the job you tell it to do,” she said. “It is not telling you what exactly to make.” She named the decision directly: “That insurance ready sounds super cool and would make a lot of people a lot of money. That doesn’t mean it’s a good idea.”
Dan Dube, who came to the panel from technology rather than behavior analysis, offered the operational version. “AI should be your copilot, not your autopilot,” he said. “Don’t just blindly trust it.” His practice is to ask a model to cite the sources behind any recommendation before acting on it, on the reasoning that a system with no answer will produce one anyway.
Ellie Kazemi, who moderated, pushed against the comfort of the framing. Her argument was that the tool category has already been overtaken in practice, because people are pasting output without reading it. “You are suggesting that it’s just a tool, and I’m saying I love that,” she said, “but evidence shows that it’s become more than the tool. It’s become the decision maker because people are copying to see what it does.” She cited research she said showed more than 80 percent of people interacting with these systems copying and pasting output without reading the detail, and noted that some researchers have begun planting phrases in documents to see whether they resurface unread. She did not name the study from the stage. The behavior she described is precisely what Ventura was asked to productize.
Questions ABA Operators Should Ask AI Documentation Vendors
Ventura’s own reframe of the trade-off travels further than the anecdote does. The vocabulary around time is monetized, he pointed out: people save it, spend it, waste it. The question that follows a time saving went unasked by both clients. “At what cost?” he said. “We need human beings to work with other human beings on how to be better human beings. That is the essence of what ABA is, and if you’re outsourcing all of that to something that’s not human, then I think you lose a lot of what you’re trying to accomplish.”
The practical residue of the story is a short list. Ask a documentation vendor what its product does when the clinician does not read the draft: whether the workflow permits submission without an affirmative review step, whether that step is logged, and what the audit trail shows six months later when a payer asks who reviewed what. Ask what the vendor has declined to build, and why. Ventura answers that one with unusual specificity, which makes it checkable.
The refusal itself is the least durable part of the story. One vendor said no to two clients. The clients still want the thing, other vendors are building in the same category, and the reading of a submitted document does not appear anywhere on a claim form. “Yes, can we save the time of the BCBA? Absolutely,” Ventura said. “But we have to consider at what cost that’s going to.” The cost, in the current enforcement climate, is denominated in recoupments.
