The first reaction most people have when they hear that AI helps draft a counseling response is suspicion. Reasonable suspicion. Counseling is supposed to be human; the last chatbot they spoke to was generic at best and uncanny at worst; the whole point of the work is to be heard by a person. I share the suspicion in some forms and not in others. What I want to do in this piece is name precisely where AI fits in the work I do, why the fit is structural rather than incidental, and what the research literature has actually found when text-based, asynchronous, AI-assisted counseling has been studied seriously. The conclusion I have arrived at, after two years of building this service and reading the evidence carefully, is that AI is not a compromise position for systemic counseling delivered by email. It is, for this medium and this approach, the better tool.
Language is the intervention
There is a remark of Steve de Shazer’s I keep returning to. De Shazer, the co-founder of solution-focused brief therapy, wrote in 1994 that “words were originally magic.” He was making a technical point. In the systemic and solution-focused traditions, the therapeutic mechanism is not insight, not catharsis, not the relationship as such. It is the precise use of language to make a pattern visible that the client could not previously see, or to introduce a difference into a stuck description that the client cannot easily reverse. The intervention is the sentence.
This claim travels well to writing. In fact, it travels better to writing than to speech, because in writing the sentence is durable, re-readable, and editable before it is delivered. The Milan school’s circular questions, Michael White’s externalizing language, the SFBT miracle question — all of these are constructions that benefit from being composed with care rather than improvised live. When a counselor sits across from a client in real time, the linguistic precision the approach requires is in tension with the social pressure to respond promptly. In writing, that pressure dissolves. You can compose, revise, throw out, start again, and arrive at the formulation that does what you want it to do.
What the AI is, and what it isn’t
The work AI does in this practice is bounded and specific. It does not decide what to say. It drafts. I have trained it on the systemic counseling literature and on the particular demands of asynchronous email work, so that given the email a person has written, it produces a draft that takes the right systemic posture, asks questions instead of giving advice, mirrors the client’s own language back where useful, and avoids the directive over-prescription that Xu et al. (2025) identified as the most common failure mode of off-the-shelf LLM counseling.
Then I read it. I edit it against what the person actually wrote, against my own reading of what the situation is asking for, and against the standards of the Lebens- und Sozialberatung credential I work under. Sometimes I rewrite paragraphs. Sometimes I keep almost the whole draft. Occasionally I throw it away and start again. The point is that the AI gives me a starting point that is already on register, which means the working session I have with each email is the editorial pass, not the cold-start composition. That pass is where the counseling actually happens. The arithmetic of who is responsible for the reply is unambiguous: a licensed systemic counselor approves every word before it leaves.
What the evidence actually shows
The relevant studies have only existed for the last few years, so it is worth naming them rather than gesturing at “research.”
Van Lotringen et al. (2021), a scoping review of text-based digital psychotherapy across twenty-three studies, found a Working Alliance Inventory score of 5.66 out of 7.0 in async text contexts. That score is in the same clinically meaningful range as in-person therapy benchmarks. The working alliance — the bond between counselor and client, the agreement on goals, the felt sense of being held — is the single most robust predictor of therapeutic outcome across modalities and approaches. The headline finding is that it forms in text, asynchronously, without any nonverbal cues at all.
Bendig et al. (2023) compared conversational mental-health applications against non-conversational ones and found the conversational element roughly doubled the effect size, from a Hedges g of 0.28 to 0.53. The dialogue itself does work; it is not only the content that matters.
Alavi et al. (2025) ran the first meta-analysis specifically of generative AI chatbots in mental health: fourteen randomized trials, around six thousand participants, effect size 0.30. Modest but statistically real, and roughly comparable to what early digital mental-health interventions achieved before LLMs existed.
What these studies do not yet directly measure is the configuration I am describing: AI-drafted, human-reviewed, asynchronous, multi-exchange, systemic. Each component has evidence; the combination sits at the frontier of what has been clinically tested. The honest claim is not that this is proven superior to a fifty-minute in-person session. The honest claim is that the combination is designed to inherit the best of each part: the linguistic precision and tirelessness of the drafting model, the clinical judgment and accountability of the counselor, the reflective discipline of writing, and the cumulative durability of an exchange the client can return to.
What the medium itself adds
A few things are worth naming directly.
Re-readability. When you receive an email from me, you can read it again at 11 pm three weeks later when the situation recurs. Spoken counseling has no such artifact. The narrative therapist David Epston used to write therapeutic letters to clients between sessions, and his own informal accounting estimated that one letter was worth roughly four and a half sessions of in-person work. Nancy Moules, writing in 2003, documented that families returned to those letters dozens of times over months and years. The written reply keeps working long after the moment it was sent.
Reflection time on both sides. The client writes alone, without performance pressure, often saying things they would not say aloud — the “online disinhibition effect” that John Suler named in 2004 is a real and clinically useful affordance, not a defect. The counselor reads, drafts, edits, and sends without the time pressure that a live session imposes. Murphy et al. (2009) found that predictable response windows — not faster responses — correlate with higher client satisfaction. The 24-hour turnaround is a feature.
Distance as a feature. Writing about a difficult situation produces psychological distance from it that is therapeutic in itself. Pennebaker’s decades of expressive-writing research have documented this. The same principle is what makes externalization in narrative therapy work, and it shows up more naturally in writing than in speech.
Writing style as clinical signal. The way a person writes — what they capitalize, where they hesitate, what they bury in a parenthesis, where the typos cluster — carries information about emotional state that is not available in a scheduled session slot. A draft model trained to read for those features can flag them; a counselor reading the draft against the original email can decide what to do with them.
The scaling problem. A working systemic counselor with a Vienna practice can hold a finite number of clients. Waiting lists for good systemic work are long, in Austria and elsewhere. Most people who would benefit from a careful, theory-grounded reading of their situation will never get one, because the supply is bounded by the hours in a week. AI drafting, with human review, changes that arithmetic without thinning the work. The counselor still reads every email, still edits every reply, still carries the clinical responsibility. What changes is the starting point of the editorial pass.
The honest limits
This is not the right form for everything, and the AI in the workflow does not change that. It is not psychotherapy. It does not diagnose. It does not treat depression, anxiety disorders, trauma, addiction, or any clinical condition. It is not appropriate for crisis, where a 24-hour response window is structurally wrong. Some people need the embodied, present, in-the-room work of a face-to-face counselor or therapist, and for those people I name that directly in the reply. The credential under which I practice — Austrian Lebens- und Sozialberatung — defines the scope, and the scope is real. The full account of what the work is and is not lives in the companion post on what systemic counseling actually is.
What is left
What is left, after the honest limits are named, is the very large space of situations most people actually live in. The stuck conversation. The pattern they cannot yet put words to. The relational strain that does not rise to clinical territory but is not going away on its own. For those situations, the combination of a written medium, a systemic frame, an AI drafting layer trained on the relevant literature, and a counselor who reads and edits every word before it leaves is the most useful instrument I have ever worked with. I am not embarrassed by the AI in the workflow. The AI is the reason the work scales without thinning — and, in some specific ways the research is beginning to make legible, the reason it can be done at all in this medium at the quality I want.