Dr. med. Maja Dobelsek Psychiatry · AI · Implementation

Safe AI in psychiatry · Implementation guide

AI ambient scribes in psychiatry: the new clinical workflow

4 recommendations on clinical workflow and governance that protect quality of care, patient safety and clinical competence, with practical tools for implementation.

Dr. med. Maja DobelsekPsychiatry resident and AI researcher, University of St. GallenNovember 20268 min read
ReliefCompetence and safety

The problem

Ambient scribes save time. In psychiatry, however, writing the note has always been part of clinical reasoning, and handing the note to a machine creates problems on two levels.

The first concerns clinical competence. The scribe only hears, so it can only infer the mental state, and a fluent draft invites a quick signature. Trainees who document with AI from day one may never build the reasoning that writing used to train. Medical education research describes these risks as deskilling, never-skilling and mis-skilling [1, 3, 5, 6, 10].

The second concerns the tool itself. Safeguards against unintended use are often missing, for example when the scribe crosses the line from a documentation tool to a decision support tool by generating diagnoses, formulations or treatment plans.

Hard evidence on deskilling, never-skilling and mis-skilling risks is still thin and comes neither from scribes nor from psychiatry. The four recommendations below address the question that follows: how can psychiatric services introduce AI ambient scribes safely, protecting quality of care, patient safety and clinical competence? They cover clinical workflow and governance; data protection would need an article of its own.

RECOMMENDATION 1

Clinical judgement must precede the AI draft

The most effective safeguard is the order of work, and the tool itself should enforce it. The clinician's own observations and judgement must exist before the AI draft is generated [8, 9, 12, 14].

The new workflow

ClinicianPatient

The scribe listens

The scribe on the table records the conversation and the dictation. The clinician takes notes alongside.

Steps 2 and 3 are built into the tool. The scribe waits for the clinician's dictation before drafting, so the review never takes place within a frame set by the tool.

What the tool must do

Four rules for the scribe

To be agreed with the vendor and built into the prompt and the product before go-live.

1

Wait for the clinician

After the conversation the scribe pauses. It drafts only once the clinician has dictated observations, formulation and plan.

2

Never infer psychopathology

The tool draws no conclusions about the mental state from the conversation. Missing findings appear as placeholders marked "missing" or "not assessed".

3

Document only what was said

Explicit statements may be recorded. If the patient says they wake at four and cannot fall asleep again, the note may record early morning awakening, but not a full mental state examination.

4

Never write formulation or plan alone

Assessment, formulation and plan come only from the clinician's dictation.

Implementation steps

Before the first user goes live

  • Train every user on the six steps, including experienced specialists. The workflow is the safeguard, so it cannot be optional.
  • Confirm with the vendor that the scribe pauses after the conversation and follows the four rules above.

RECOMMENDATION 2

Reviewing the draft follows a system

Reviewing a fluent draft is a distinct skill: clinicians should first learn to write, then learn to review, and only then use AI in daily practice [11].

Independent writingcore skill
Review trainingsecond core skill
Use AIin daily practice

The PDSQI-9 [4] is a validated instrument for rating AI summaries of the medical record against a written source. An ambient scribe has no written source: the reference is the consultation itself. I therefore adapted it for psychiatry. The clinician who led the consultation reviews the note against their memory of the conversation.

PDSQI-8 · eight criteria

  1. 1 · AccurateIs the note accurate?
  2. 2 · ThoroughIs the note complete, without omissions?
  3. 3 · UsefulIs the note useful?
  4. 4 · OrganisedIs the note well structured?
  5. 5 · ComprehensibleIs the note clear and easy to understand?
  6. 6 · SuccinctIs the note concise and economical?
  7. 7 · SynthesisedDoes the note condense the consultation?
  8. 8 · Stigmatising languageDoes the note contain stigmatising language?

Criteria 1–7 are rated from 1 (not at all) to 5 (fully), criterion 8 with yes or no. The full adapted PDSQI-8, with explanations, examples and rating scale, is available on request. It is my working adaptation and is not itself validated; in my trial it is collected as a descriptive outcome.

RECOMMENDATION 3

Scribe access for trainees is earned, not granted

Access depends on demonstrated competence, not on a date. New residents write their notes themselves until a supervisor confirms that they can. Others have arrived at the same idea with different thresholds, none of them based on data [2, 6, 7].

Stage 0

New residents and trainee psychologists

Usually the first year of training

Scribe: not for primary documentation
Self-written: the entire note

Assessment

A supervisor confirms that self-written admission notes are good enough

Instrument: PDQI-9 Psych

Stage 1

Advanced trainees and specialists

Scribe: full drafting support
Self-written: the clinical layer of every note

From my research · PDQI-9 Psych

For the assessment I adapted the validated PDQI-9 [13] to psychiatric admission notes, with psychiatric anchors and four safety-critical elements.

Implementation steps

Gate access by competence

  • Record each trainee's stage and set up tool access to match.
  • Let supervisors decide when access is granted, based on their rating of self-written admission notes.

RECOMMENDATION 4

Supervision targets clinical reasoning, not the note

Once a scribe writes part of the note, the note no longer shows reliably what a trainee thinks [8]. Supervisors give feedback on observations and reasoning directly. Four formats work together.

1

Bedside teaching

A supervisor observes the interview. In Switzerland, the SIWF already asks for at least four workplace-based assessments (Mini-CEX or DOPS) per year.

2

Train the supervisors

Supervisors learn the workflow, typical scribe errors and the PDSQI-8 themselves, so they can tell what a trainee thought and what the AI generated.

3

Use AI-generated feedback on interviewing

A suitably prompted AI tool reads the transcript and shows which questions were asked and which were missed, such as the follow-up questions after a positive suicide screen. This needs documented consent from the trainee, with a real option to decline.

4

Oral case presentations with the Assessment of Reasoning Tool

Syndrome, key symptoms, differential diagnoses, risk and plan, rated in a structured way.

For oral case presentations, the Assessment of Reasoning Tool rates five domains of clinical reasoning, from data gathering to metacognition [15].

About the author

Dr. med. Maja Dobelsek

Psychiatry resident, AI researcher and PhD candidate at the University of St. Gallen. I research AI ambient scribes in psychiatric care, with one aim: AI should strengthen the human contact in psychiatry and keep patients safe.

LinkedIn · ORCID

Behind this article

  • Principal investigator of a clinical trial of an ambient scribe in psychiatric care
  • Developed the PDSQI-8 and PDQI-9 Psych adaptations and the 64-item content coverage checklist
  • Member of a hospital digital transformation team and of the AlpineAI pilot project with an ambient scribe
  • Daily user of an ambient scribe for psychiatric admission reports

Organisations introducing ambient scribes in psychiatric or mental health services are welcome to get in touch to exchange experience on safeguards and implementation.

References

  1. Abdulnour RE, Gin B, Boscardin CK. Educational strategies for clinical supervision of artificial intelligence use. N Engl J Med. 2025;393(8):786–797. doi:10.1056/NEJMra2503232
  2. Abernethy J et al. Integrating AI scribes into medical education: guardrails for preserving clinical reasoning. J Gen Intern Med. 2026;41(9):2598–2602. doi:10.1007/s11606-025-10149-w
  3. Budzyń K et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025;10(10):896–903. doi:10.1016/S2468-1253(25)00133-5
  4. Croxford E et al. Development and validation of the provider documentation summarization quality instrument for large language models. J Am Med Inform Assoc. 2025;32(6):1050–1060. doi:10.1093/jamia/ocaf068
  5. Heudel PE et al. Artificial intelligence in medicine: a scoping review of the risk of deskilling and loss of expertise among physicians. ESMO Real World Data Digit Oncol. 2026;12:100693. doi:10.1016/j.esmorw.2026.100693
  6. Ke Y et al. AI-induced never-skilling in medical education. Nat Med. 2026;32(6):1997–2006. doi:10.1038/s41591-026-04438-y
  7. Khan M et al. Artificial intelligence in psychiatric graduate medical education. Acad Psychiatry. Published online July 28, 2026. doi:10.1007/s40596-026-02406-9
  8. McGaughey S et al. When the scribe does the reasoning: ambient artificial intelligence, inference impersonation, and the development of trainees’ clinical judgment. Acad Med. Published online September 15, 2026. doi:10.1093/acamed/wvag287
  9. Rucker P, Miller M, Armstrong D. How Cigna saves millions by having its doctors reject claims without reading them. ProPublica and The Capitol Forum. March 25, 2023. propublica.org
  10. Qazi IA et al. Automation bias in large language model–assisted diagnostic reasoning among physicians trained in AI literacy — a randomized clinical trial. NEJM AI. 2026;3(5). doi:10.1056/AIoa2501001
  11. Raka Y, Van Deusen R, Wiberg H. From authoring to editing: a competency framework for trainee use of ambient AI clinical documentation. npj Digit Med. 2026;9:785. doi:10.1038/s41746-026-03351-5
  12. Roth AS, Ayers NB. AI scribe functions in psychiatric practice: clinical oversight, consent, and regulation. Lancet Psychiatry. 2026;13(9):804. doi:10.1016/S2215-0366(26)00201-4
  13. Stetson PD et al. Assessing electronic note quality using the Physician Documentation Quality Instrument (PDQI-9). Appl Clin Inform. 2012;3(2):164–174. doi:10.4338/ACI-2011-11-RA-0070
  14. Talwalkar JS et al. Impact of an ambient AI scribe on medical student objective structured clinical examination notes: nonrandomized clinical trial. JMIR Med Educ. 2026;12:e88264. doi:10.2196/88264
  15. Thammasitboon S et al. The Assessment of Reasoning Tool (ART): structuring the conversation between teachers and learners. Diagnosis. 2018;5(4):197–203. doi:10.1515/dx-2018-0052

Views are my own and do not represent my employer.