Dr. med. Maja Dobelsek Psychiatry · AI · Implementation
All articles

Essay · Mental health services

25% more patients, no new staff: why AI should help psychiatry end treatment on time

Before AI touches diagnosis or psychotherapy, it should help us end outpatient treatment at the right time.

Dr. med. Maja DobelsekPsychiatrist · PhD candidate, HSG 15 November 2026 · 6 min read
Personal perspective. This essay reflects my own view as a clinician and researcher, not a settled truth and not the position of my employer.

A friend recently asked me where AI should be implemented in psychiatry next. My first answer was: not in diagnosis, and not in psychotherapy. At least not yet.

Before AI touches the core of our clinical work, I believe it should do two more modest things. First, take the repetitive tasks off our desks. Second, help us reduce the number of patients each clinician carries in the public healthcare system. This article is about the second point, and about one specific lever: ending outpatient treatment at the right time.

What AI does not need to do first

Most public attention goes to AI that diagnoses or delivers therapy. I understand why: these look like the heart of psychiatry. In my daily work, though, they are rarely what limits the care we can offer.

What limits it is time. The first step is already under way: ambient scribes that take over documentation, which is the focus of my own research. They give clinicians back minutes per consultation. The next step is to look at the consultations themselves, and at how many of them are still needed.

The real bottleneck: caseload

Reducing the number of consultations per treatment by 20% increases a clinic’s capacity by 25%. Any practitioner senses this intuitively, but the arithmetic is worth spelling out.

Take a clinician with 100 appointments a month. If a treatment episode takes 10 consultations on average, those appointments serve 10 episodes. If episodes take 8 consultations instead, the same 100 appointments serve 12.5 episodes: 25% more patients, with no new staff.

Put generally: if each treatment needs a share r fewer consultations, capacity grows by 1 / (1 − r) − 1. With r = 0.2, that is 0.25.

The effect grows faster than the reduction itself: cutting 30% of consultations gives 43% more capacity. Two conditions apply. The freed appointments have to go to people on the waiting list, and the consultations removed have to be ones that no longer bring benefit.

Why ending treatment is so hard

Ending an ongoing outpatient treatment with a stable patient is one of the hardest conversations in our work. The reason is rarely clinical. It is the relationship that has grown between clinician and patient, and the fear that suggesting an end will feel like rejection.

Research confirms that this is not an individual weakness. A qualitative study in three Dutch mental health services found that discussions about ending long-term treatment were highly unstructured and clear decisions were rare1. The authors describe continuation as the easy default, so ending requires deliberate effort against it.

Psychotherapy research points in the same direction. In UK routine care, patients with planned endings reached similar recovery rates whether they attended few or many sessions2. The “good-enough level” model explains this: people tend to stop once their gains are sufficient, at very different speeds3. These findings come from psychotherapy rather than psychiatric outpatient care, but they suggest that more sessions do not automatically mean more benefit.

I see a small natural experiment in my own clinic. In the months before a clinician leaves, they are asked to close as many cases as possible, and many cases then turn out to be closable. That tells me the criteria for ending often exist. What is missing is a reason to apply them that does not come from the relationship.

What AI could do: an external reference point

AI could apply shared, explicit criteria across a whole caseload and flag treatments that may be ready to end. The clinician no longer has to say “I think we should stop.” They can say “Our clinic reviews every treatment against the same criteria, and yours now meets them. Let’s look at it together.”

This externalizes the need for ending without putting the therapeutic relationship at risk. It works the way a lab value or a guideline does when we taper a medication: the reason sits outside the two people in the room.

The reason for ending sits outside the two people in the room.

In practice, such a system could:

  • track stability over time, for example months without crisis contacts, unchanged medication and treatment goals reached;
  • notice long stretches of consultations in which the documentation shows no new problem and no change in plan;
  • prepare a short list for the team’s case conference, so that decisions about ending become a routine agenda item;
  • flag the opposite case too: patients who are not improving and may need a different or more intensive treatment.

That last point matters. Feedback systems in psychotherapy help most for patients who are not on track4. The same meta-analysis found no effect of feedback on treatment duration. Measuring progress alone does not shorten treatment. The decision to end still has to be made, and this is where explicit ending criteria would add something new.

Guardrails: what this must not become

An ending tool only helps patients if it stays a decision aid. Without clear limits, it can easily turn into a cost-cutting machine. I would set five conditions.

  1. The clinician decides. The system proposes a review; it never ends a treatment. Externalizing the reason is not the same as hiding behind it.
  2. Patients see the criteria. Feedback works better when it reaches both clinicians and patients5. Ending becomes a shared look at progress, not a verdict.
  3. Criteria fit the diagnosis. Some conditions, such as psychotic or bipolar disorders, often need long-term care. One threshold for everyone would be clinically wrong.
  4. Ending has a safe landing. A clear way back, a handover to the GP or to psychiatric home care (Spitex) for relapse prevention, and a documented plan. Ending well is the opposite of abandonment6.
  5. We check for unequal effects. We need to monitor whether certain groups, for example people with less German or fewer resources, are flagged for ending more often than others.

And the freed capacity has to go to the people waiting for a first appointment. Otherwise the gain stays on paper.

What it would take

The hardest part is not the AI. It is agreeing, as clinicians, on what “ready to end” means. That step needs no technology at all, and it should come first.

After that, three things are needed:

  • Data. Many clinics, mine included, do not collect structured outcome scores such as the PHQ-9 routinely. Either we start doing so, or we extract trajectories from clinical notes. The second option only works if documentation quality is measured, not assumed.
  • Integration. The flag has to appear where clinicians already work, in the electronic health record and in the case conference. In many hospitals, connecting a new tool to the record is the slowest part.
  • Evaluation. A pilot should measure waiting times, re-referrals within a year, and how patients and clinicians experienced the ending. Capacity gained at the cost of more relapses would be no gain.

Closing

AI that helps us end treatment at the right time is not the most exciting idea in psychiatry. I still think it is one of the most useful. It would give more people access to the clinicians who are already there, and it would let those clinicians end good treatments in a good way.

This is my personal view as a psychiatry resident and AI researcher, and I would be glad to hear where colleagues see it differently.

References

  1. Koekkoek B, van Meijel B, Perquin A, Hutschemaekers G. Decision making on (dis)continuation of long-term treatment in mental health services is an interpersonal negotiation rather than an objective process: qualitative study. BMC Psychiatry. 2019;19:92. doi:10.1186/s12888-019-2072-0 · PMC
  2. Stiles WB, Barkham M, Wheeler S. Duration of psychological therapy: relation to recovery and improvement rates in UK routine practice. Br J Psychiatry. 2015;207(2):115–122. doi:10.1192/bjp.bp.114.145565 · Cambridge Core
  3. Barkham M, Connell J, Stiles WB, Miles JNV, Margison F, Evans C, Mellor-Clark J. Dose-effect relations and responsive regulation of treatment duration: the good enough level. J Consult Clin Psychol. 2006;74(1):160–167. doi:10.1037/0022-006X.74.1.160
  4. de Jong K, Conijn JM, Gallagher RAV, Reshetnikova AS, Heij M, Lutz MC. Using progress feedback to improve outcomes and reduce drop-out, treatment duration, and deterioration: a multilevel meta-analysis. Clin Psychol Rev. 2021;85:102002. doi:10.1016/j.cpr.2021.102002 · Leiden University
  5. Gondek D, Edbrooke-Childs J, Fink E, Deighton J, Wolpert M. Feedback from outcome measures and treatment effectiveness, treatment efficiency, and collaborative practice: a systematic review. Adm Policy Ment Health. 2016;43:325–343. doi:10.1007/s10488-015-0710-5 · PMC
  6. Psychiatric News. Ending a treatment relationship. 2023. doi:10.1176/appi.pn.2023.06.6.17
Share LinkedIn Copy link Email
Cite this essay Dobelsek M. 25% more patients, no new staff: why AI should help psychiatry end treatment on time. majadobelsek.com. 15 November 2026. Available at: majadobelsek.com/articles/end-treatment-on-time
About the author Dr. med. Maja Dobelsek is a psychiatrist at Psychiatrie St. Gallen and a PhD candidate at the University of St. Gallen (HSG), where she leads a controlled trial of an AI ambient scribe in routine psychiatric care. More about me

Keep reading

Essay · Ambient scribes When does a medical AI scribe start being a decision tool? Essay How the work of a psychiatrist will change once AI tools are everywhere