When AI Writes Your Session Notes: What ABA Leaders Need to Know

AI-generated session note summaries are becoming common in ABA software, helping agencies save time and streamline documentation. That convenience carries real compliance risk. This article walks through how AI is being used in ABA session notes today, what we've found in actual audits, and what agency leaders need to do to keep their documentation accurate, ethical, and defensible.

Can ABA agencies use AI to write session notes?

Yes, but AI can only draft, never finalize. Under the guidelines from the Artificial Intelligence Consortium for Applied Behavior Analysis (AIC-ABA), AI use in ABA must be human-led: a qualified professional reviews, edits, and approves every AI-generated summary and remains responsible for the final content. The moment AI drafts part of the clinical record, that record becomes part of your compliance risk profile. AI doesn't own the note. You do.

AI already sits inside many of the systems ABA agencies rely on every day. Most modern EHR platforms now offer AI-generated narrative summaries built from session data. The promise is simple: faster documentation, cleaner notes, less burden on staff. And it can deliver on that. The risk shows up when agencies treat the output as finished instead of as a draft.

Audits That Raised Red Flags

We've audited multiple ABA agencies that had fully adopted AI-generated session summaries, but built no meaningful safeguards around them. They had:

• No "human in the middle" review process
• No indication in the notes that AI generated the summaries
• No client or caregiver consent for AI use in medical records
• Full reliance on the exact prompt recommended by their EHR vendor

From each agency's perspective, they were doing what they thought they should. They trusted the system and assumed their vendors had done the hard work. That assumption created real exposure.

From a compliance standpoint, "we followed the vendor's instructions" isn't a defense. Responsibility for documentation doesn't transfer to the software company. It stays with the provider organization. Here's what the audits found in the summaries the AI produced.

What the AI Was Actually Producing

When we reviewed the notes, we saw consistent patterns of risk. These weren't isolated glitches. They showed up across multiple notes and providers, in four categories, each illustrated with real examples.

1. AI Making Clinical Judgments Instead of Reporting Facts

"Social skills were also enhanced…"
"The client reached a high accuracy level of between 75% to 78.6% across different targets."
"Activities such as sorting items into categories and requesting help facilitated communication development…"
"These observations highlight the areas requiring ongoing attention and intervention to further reduce the frequency of these behaviors…"

Each statement sounds appropriate on the surface. They share one problem: they interpret outcomes rather than document observable events. The AI summarizes what it thinks the data means instead of sticking to what occurred.

In several cases, the underlying data didn't clearly support these conclusions. The language made progress sound more definitive than the data justified. Session notes are supposed to be objective and defensible. When AI introduces interpretation without verification, it raises the risk of overstating progress or misrepresenting clinical outcomes. Without a human reviewing the content, these judgment calls go straight into the record.

2. Assertions Without a Clear Basis in the Record

"He successfully said 'thank you' to the aide and his sister with prompts."
"CLIENT demonstrated progress in several programs, including Manners, Transitions, and Identifying Letter Sounds."
"The focus remained on addressing these behaviors consistently without any additional unwanted behaviors being observed."
"The BCBA particularly focused on the programs where attention and task engagement were parts of the key skills being targeted."

These statements introduce a second risk: unsupported assertions. The AI included details that either weren't documented or couldn't be traced back to specific data points from those encounters.

The sibling reference is especially telling. Nothing from the provider who delivered the intervention indicated a sibling was present. The AI inserted that detail to make the narrative feel complete. The others make broad claims about progress or focus without tying them to measurable data reported by an actual provider. This kind of language slips through because it sounds natural. If it can't be substantiated, it weakens the note's credibility and creates exposure during audits or payer reviews. That gap between what a note says and what it can prove is the same problem we examine in Your Note Describes the Session. Does It Defend the Bill?.

3. Inclusion of Non-Billable or Irrelevant Activities

"The BCBA placed orders for all necessary items to ensure sessions can be conducted with fidelity and consistency."

These activities may well have occurred. That's not the problem. The problem is that they don't belong in a session note tied to billable services, at least not the way they're presented.

AI doesn't understand payer rules or documentation standards. It pulls in activities that sound relevant without distinguishing whether they support the billed service. The result blends clinical care with supervision, planning, or administrative work, and muddies the connection between what was documented and what was billed. Keeping that line clean matters most in codes where the provider and activity must match precisely, a point we cover in 97155 Solo Sessions.

4. Fabricated or Speculative Content

"During their discussion, the BCBA and BT likely focused on refining intervention strategies…"
"Although specific program modifications by the BCBA aren't provided, typical adjustments might involve tailoring interventions…"
"While specifics of the client's response to treatment weren't provided, the supervision session likely included reviewing progress…"
"The BCBA and BT likely discussed challenges faced and emphasized data collection accuracy…"

This is the most serious category. Here the AI is no longer summarizing. It's generating hypothetical content, filling missing information with what usually happens in similar situations.

Words like "likely," "might," and "typically" signal that the content isn't based on documented events. This isn't just inaccurate. It's fabricated. Once that language enters a signed session note, it can seriously undermine the defensibility of the record.

The Prompt Problem Behind the Scenes

One of the most important findings wasn't just what the AI produced. It was why.

The agencies used the exact prompts their EHR vendors recommended, and assumed those prompts were enough for compliant documentation. They weren't. The prompts didn't restrict inference, didn't require alignment with source data, and didn't prevent speculative language.

So the AI did what it was designed to do. It generated polished, complete-sounding narratives. They just weren't always accurate. And again, responsibility for that doesn't sit with the vendor. It sits with each ABA agency.

What the Guidelines Make Clear

The AIC-ABA guidelines emphasize several expectations:

• AI must be human-led, with clear accountability for outputs.
• Use must be transparent, with appropriate disclosure and consent.
• Outputs must be monitored for accuracy, safety, and bias.
• Organizations must be prepared to pause or discontinue use if risks appear.

In these cases, none of those safeguards were in place. That's what turned a helpful feature into a compliance liability. The Council of Autism Service Providers (CASP) practice parameters for AI use in ABA reach the same conclusion from the payer and regulatory side, spelling out organizational oversight, monitoring, and auditing expectations for agencies deploying AI.

A Better Way to Stay in Control

AI isn't the problem. Unstructured documentation is.

If your team doesn't have a clear standard for what belongs in a session note, AI will fill the gaps for you, and as we've seen, it won't always do that in a compliant way. This is really a question of building the right oversight into your compliance program from the start, which is where knowing what to prioritize first pays off.

That's why we created our ABA Session Note Frameworks. They define what should and shouldn't go into a session note, making it easier to review AI-generated content and catch issues before they become risks. If you're using AI, this gives you a way to stay in control of your documentation. Learn more about the ABA Session Note Frameworks.

Frequently Asked Questions

Can ABA agencies use AI to write session notes?
Yes, but only as a drafting aid. AIC-ABA guidelines require AI use in ABA to be human-led, meaning a qualified professional reviews, edits, and approves every AI-generated summary and stays responsible for the final content. AI can speed up a first draft; it cannot produce the final, signed note on its own.

Is AI-generated ABA documentation compliant?
Only when a qualified human reviews and verifies it against the actual session data. In audits, AI-generated summaries have interpreted outcomes, asserted details with no basis in the record, included non-billable activities, and even fabricated speculative content using words like "likely" and "typically." Any of those can make a signed note indefensible. Compliance depends on the review process, not the software.

Who is responsible when AI writes an inaccurate session note?
The provider organization. Responsibility for documentation does not transfer to the EHR or AI vendor. "We followed the vendor's recommended prompt" is not a defense. If it's in the note, your agency owns it.

Do we need client consent to use AI in session notes?
AIC-ABA guidelines call for transparency, including appropriate disclosure and consent for AI use in medical records. In the audits we reviewed, missing consent and missing disclosure that AI generated the summary were among the most common gaps.

Why do vendor-recommended AI prompts still produce risky notes?
Because most vendor prompts don't restrict inference, don't require alignment with source data, and don't prevent speculative language. The AI then does what it's built to do: generate a polished, complete-sounding narrative, whether or not the data supports it. A structured note standard plus human review is what closes that gap.

Final Thought

AI is going to be part of ABA documentation moving forward. The question is whether your agency is managing it or trusting it.

Because no matter what your vendor says, and no matter how the note is generated, if it's in the note, your organization is responsible for it.