AI Progress Notes for Therapists: What They Get Right and Wrong

AI-generated progress notes are no longer a novelty. In 2026, nearly every major EHR platform has some form of AI documentation, and standalone AI note tools have multiplied rapidly. The question is no longer whether AI can generate therapy notes. It is whether the notes it generates are actually good.

The answer, honestly, is complicated. AI notes solve some problems elegantly and create others. They save time in some areas and cost time in others. They are sometimes clinically impressive and sometimes clinically embarrassing.

This is not a sales pitch for any particular tool. It is a clear-eyed assessment of what AI-generated progress notes get right, what they get wrong, and how to evaluate whether an AI tool is actually helping your clinical documentation or just creating a different kind of work.

What AI Notes Get Right

Speed of First Draft

The most obvious advantage is speed. AI can produce a structured progress note in seconds -- whether from session audio, therapist-entered observations, or a combination of both. For a therapist who spends 10 to 15 minutes per note manually, getting a complete first draft in under a minute is genuinely transformative.

This matters because documentation time is the most common source of therapist burnout that is not about clinical work itself. A full-time therapist seeing 25 sessions per week and spending 10 minutes per note is spending more than four hours weekly on documentation. If AI cuts that to two or three minutes of review and editing per note, you recover 2 to 3 hours per week. Over a year, that is 100 to 150 hours -- roughly three to four full working weeks.

The speed advantage is real, and it is the primary reason therapists adopt AI notes.

Consistent Structure

Human note-writing is inconsistent. On Monday morning, your notes are detailed and well-structured. By Friday afternoon, they are compressed, vague, and occasionally incomplete. This is normal -- cognitive fatigue affects documentation quality even when clinical skill remains high.

AI produces consistent structure every time. The format is the same for your 8 AM session and your 5 PM session. Every note includes the sections you need, in the order you expect, with nothing omitted because you were tired.

This consistency has real clinical value. Consistent documentation is easier to review when preparing for sessions, easier for a covering therapist to follow, and more defensible if notes are ever examined by a licensing board or in a legal context.

Capturing Details You Might Miss

This advantage is less obvious but significant. During a 50-minute session, a lot happens. Even a skilled therapist writing notes 10 minutes later will miss some details -- the exact phrasing a client used, a somatic shift that happened mid-session, a briefly mentioned stressor that turned out to be significant.

AI tools that work from session audio capture everything that was said. They do not forget, get distracted, or run out of cognitive bandwidth. The raw material is complete, even if the processing of that material varies in quality.

For therapists who have ever wished they could remember exactly how a client described something three weeks ago, the completeness of AI-captured session data is a genuine benefit.

Reducing the "Blank Page" Problem

Many therapists report that the hardest part of note-writing is starting. The blank note template creates a psychological barrier, especially at the end of a long clinical day. AI eliminates that barrier by providing a draft to react to rather than a blank space to fill.

Editing is cognitively easier than creating. Reviewing an AI draft and making corrections is a different kind of mental task than composing a note from scratch. For therapists who dread documentation, the shift from creation to curation can make the entire process feel more manageable.

What AI Notes Get Wrong

The Modality Problem

This is the most significant limitation of current AI note tools, and it is the one that most directly affects clinical quality.

Most AI note tools are modality-blind. They transcribe what was said and organize it into a standard format (usually SOAP). They do not understand the difference between cognitive restructuring in CBT and defusion in ACT. They do not recognize parts language in IFS. They do not know what a desensitization phase is in EMDR or what a SUD rating means.

The result is that a CBT session, an IFS session, and an EMDR session all produce essentially the same kind of note: a summary of what was discussed, organized into Subjective/Objective/Assessment/Plan.

For a CBT session, the note might say "client discussed negative thoughts about work" instead of "identified all-or-nothing thinking: 'I got one piece of critical feedback, so I'm completely incompetent.'" For an IFS session, it might say "client explored feelings of emotional numbness" instead of "contacted a manager part (The Controller) protecting an exile carrying a burden of emotional neglect." For an EMDR session, it might say "client processed a traumatic memory using eye movements" instead of documenting the SUD trajectory from 7 to 3 across twelve sets of bilateral stimulation.

The clinical vocabulary that makes notes meaningful gets flattened into generic language. You end up with notes that are technically accurate as summaries but clinically shallow as documentation.

The "Sounds Right, Isn't Right" Problem

AI is very good at producing text that reads well. The grammar is correct, the sentences flow, the structure is professional. This creates a subtle risk: notes that sound clinically competent but contain inaccuracies or fabrications.

Common examples:

• Invented specificity. The AI generates a specific homework assignment that sounds plausible but was not actually discussed in the session.
• Wrong clinical vocabulary. The AI uses a term like "cognitive distortion" in a session where the therapist was doing psychodynamic work, because the term pattern-matches to the session content.
• Inflated progress. AI tends to be optimistic in assessment sections, describing clients as "making progress" or "demonstrating insight" even when the session was difficult and progress was minimal.
• Hallucinated measurements. Some AI tools generate specific scores (PHQ-9 ratings, SUD levels) that were never discussed or measured.

Because the text reads smoothly, these errors can slip past a therapist who is reviewing quickly. A note that says "client scored 12 on the PHQ-9, down from 16 last session" sounds great -- unless you did not administer the PHQ-9 this session. The confidence of AI-generated text can make fabrications harder to catch than gaps.

The Context Problem

AI tools process each session independently. They do not have the context that you, as the therapist, carry: the client's treatment arc, the themes that have been building across months, the significance of a small comment that connects to something from six sessions ago.

A client who says "I talked to my mother this week" might be reporting something routine, or it might be the first contact after a two-year estrangement that you have been working toward for months. The AI does not know which one it is. It will document the statement without the context that gives it clinical significance.

This is not a flaw in the AI -- it is a structural limitation of session-by-session processing. But it means that the therapist's clinical judgment is essential for contextualizing what the AI produces. AI notes should be treated as drafts that need human context, not finished products.

The Consent and Therapeutic Alliance Problem

If you are using AI that records or processes session audio, there is a clinical dimension beyond HIPAA compliance: the effect on the therapeutic relationship.

Some clients are comfortable with AI-assisted documentation. Others are not. Some will say they are comfortable because they do not want to inconvenience their therapist, while internally they are editing what they share. The mere presence of recording technology can alter what clients disclose, which alters the clinical data that generates your notes.

This is not an argument against AI notes. It is an argument for thoughtful implementation. Informed consent should be genuine -- not a form signed during intake that the client forgets about. The impact on disclosure should be monitored clinically, not assumed away.

The Deskilling Risk

There is a longer-term concern worth acknowledging: if therapists rely on AI for documentation over years, will their ability to write clinical notes degrade?

Note-writing is not just administrative work. The process of translating a clinical session into structured documentation is itself a form of clinical reasoning. When you write a progress note, you are organizing your clinical thinking -- identifying what was significant, connecting it to the treatment plan, and articulating your conceptualization. If AI does this for you, that cognitive process may atrophy.

This is a theoretical concern, and it is not a reason to avoid AI notes. But it is a reason to remain actively engaged with the notes AI generates rather than approving them without review. The note is a clinical document. The therapist should understand every word in it.

How to Evaluate an AI Notes Tool

If you are considering AI notes or evaluating your current tool, here are the questions that actually differentiate quality from marketing.

Does It Understand Your Modality?

Ask for sample output from a session in your specific therapeutic framework. If you practice IFS, the sample should include parts language, Self-energy assessment, and system dynamics. If you practice EMDR, it should track SUD levels, bilateral stimulation sets, and phase transitions. If you practice CBT, it should name cognitive distortions and document homework.

If the sample output looks the same regardless of the modality described, the AI is doing transcription-to-formatting, not clinical documentation.

What Is Your Actual Edit Time?

Generation speed is a marketing metric. Edit time is the clinically relevant metric. If the AI generates a note in 30 seconds but you spend 8 minutes correcting it, your total documentation time is 8.5 minutes. That is better than 12 minutes from scratch, but not transformative.

Ask to test the tool with a real session (or a detailed mock session) and measure how long you spend editing. If you are rewriting entire sections, adding clinical vocabulary the AI missed, or correcting fabricated details, the tool is not saving you as much time as it claims.

Where Does the Audio Go?

If the AI processes session audio, you need to know: where is the audio stored? For how long? Who has access? Is there a Business Associate Agreement (BAA) with the AI provider? Does the AI provider retain data for training purposes?

HIPAA compliance is not a checkbox. It is a set of specific technical and legal requirements. A vendor saying "we are HIPAA compliant" is not the same as a vendor explaining exactly how they handle PHI: zero data retention, encrypted transmission, BAA in place, and no use of clinical data for model training. If you want to go deeper on this topic, our guide on AI therapy notes and HIPAA compliance covers what to ask and what the answers should look like.

Is the Output a Draft or a Finished Product?

The best AI note tools position their output as drafts for therapist review. The worst ones encourage therapists to accept notes without editing. This distinction matters clinically and legally.

You are responsible for every word in your progress notes, regardless of who or what generated them. If an AI produces a note that contains a fabricated measurement, a wrong clinical term, or an inaccurate assessment, and you sign it, that is your note. Treat AI output as a first draft that needs your clinical judgment applied to it.

Tools that make review easy -- highlighting uncertain content, providing edit interfaces, flagging potential issues -- are better than tools that try to minimize friction by making approval a single click.

Does It Handle What Was Not Said?

A skilled therapist notices what a client does not say as much as what they do say. A client who always discusses their marriage but suddenly stops mentioning it. A client who has been improving but shows up with flat affect and gives short answers. A client who changes the subject every time a particular topic approaches.

AI is constrained by what is explicitly said. It cannot document meaningful silences, topic avoidance, or the absence of expected content. Your notes need to capture these clinical observations, which means you need to add them after the AI generates its draft.

The best AI tools leave space for therapist observations that go beyond transcription. The worst ones present a seamless, complete-looking note that discourages additions.

The Middle Path: AI as Clinical Tool, Not Clinical Replacement

The healthiest relationship with AI notes is neither wholesale adoption nor wholesale rejection. It is treating AI as a clinical tool that does certain things well and other things poorly, and using it accordingly.

What this looks like in practice:

Use AI for the first draft. Let it handle structure, formatting, and the bulk of session content. This is where AI saves the most time with the least risk.

Review with your clinical brain engaged. Do not skim. Read the note as if you were reviewing a supervisee's documentation. Is the clinical vocabulary correct? Are the assessments accurate? Is anything fabricated? Is anything clinically significant missing?

Add context the AI cannot know. The treatment arc, the significance of certain statements, your clinical impressions that go beyond what was said. These are the elements that make a progress note a clinical document rather than a meeting summary.

Edit the modality-specific elements. If your AI does not understand your modality, this is where you will spend the most editing time. The distortion names in CBT, the parts language in IFS, the SUD tracking in EMDR -- these need to be accurate, and generic AI often gets them wrong.

Monitor your own engagement. If you notice that you are approving notes without reading them, you have crossed from using AI as a tool to delegating clinical judgment to software. Reset the habit.

TherapyDesk takes a different approach to AI notes by building modality awareness into the system. Rather than generating generic summaries that you edit into clinical documents, it produces drafts in your therapeutic framework's vocabulary -- CBT notes that name distortions, IFS notes that track parts, EMDR notes that document SUD levels and phases. The editing step becomes fine-tuning rather than rewriting. If you want to see the difference, the demo takes two minutes.

What the Next Generation of AI Notes Should Look Like

The current state of AI notes is a starting point, not an endpoint. Here is what meaningful improvement looks like:

Modality awareness as standard. AI that understands therapeutic frameworks, not just transcription formatting. This is the biggest gap in current tools and the one that would most reduce edit time.

Cross-session context. AI that understands the client's treatment arc, not just the current session. Notes that reference last session's homework, track measurement trends across time, and connect today's content to the treatment plan.

Clinical flagging. AI that identifies potential concerns -- risk factors mentioned in passing, inconsistencies with previous sessions, measurements that suggest treatment plan revision -- and brings them to the therapist's attention.

Transparent uncertainty. AI that tells you when it is unsure, rather than generating confident text for everything. "Client may have mentioned a SUD rating of 3 [uncertain -- verify]" is more useful than confidently stating "SUD: 3" when the AI is guessing.

Therapist-in-the-loop design. Tools designed around the assumption that the therapist will review and edit, with interfaces that make review efficient rather than tools designed to minimize therapist involvement.

Some of these capabilities exist in early forms. None are universal. The therapist who evaluates AI tools critically today will make better choices than the one who accepts marketing claims at face value.

Conclusion

AI progress notes are a genuine advancement in clinical documentation. They save time, provide consistent structure, and reduce the blank-page barrier that makes documentation feel burdensome. These are real benefits, and they matter for therapist wellbeing and practice sustainability.

But AI notes also flatten clinical vocabulary, fabricate plausible-sounding details, lack cross-session context, and cannot understand what was left unsaid. These are real limitations, and they matter for clinical quality and documentation integrity.

The therapists who will benefit most from AI notes are the ones who treat them as tools to be evaluated critically, not technologies to be adopted uncritically. Ask about modality awareness. Measure your edit time. Understand where your data goes. And never approve a note you have not read.

AI should make your documentation better and faster. If it is only making it faster, you have the wrong tool.