AI in practice and pharmacy: three workflows that give writing time back

Starting point: the paperwork grows faster than the consultation hour

Doctors in ambulatory practice spend an average of 54 minutes per working day on requirements from authorities and insurers alone, and 87% of those in ambulatory care name insurers as the heaviest burden.¹ That is close to an hour arising every day in a place where nobody is treated: at the desk, usually after the last consultation.

At the same time, the hours are getting shorter. General practitioners today have 6 hours per week less available for treatment than 20 years ago, and by 2035, 40% of them have to be replaced.² Those who stay take on more patients, not fewer forms.

The federal government is now measuring the effort itself: since May 2026 the Federal Office of Public Health (BAG) has been collecting data from general practitioners, first interim results were expected by the end of August 2026 and the final report is due in January 2027.³ So the debate about relief is under way. Today’s post is waiting all the same.

In the pharmacy the arithmetic differs, the result is similar: since 1 January 2020, Article 30 of the Therapeutic Products Act has required a quality assurance system.⁴ Around 1,800 pharmacies with an average of 12.7 people and 9 full-time equivalents therefore each run their own quality management system (QMS). That includes work instructions, training records and an annual review.⁵ In home care, around 400 non-profit Spitex organisations with more than 40,000 employees serve about three quarters of all Spitex clients.⁶ Every one of them writes progress reports to insurers and authorities.

That sets the frame for this article, and it is a narrow one. An AI colleague makes no diagnosis, performs no triage, recommends no therapy and no dose, checks no interactions and replaces no medication check. Such functions would make it a medical device, with everything that conformity assessment brings with it. Swissmedic states where the line runs: software is a medical device as soon as it serves a medical purpose for the benefit of an individual person and does more with the data than store, archive, search, communicate or losslessly compress it.¹⁰ This is about administration, documentation, the knowledge base and translation. The medication check stays in the pharmacy’s specialist system; the AI colleague writes the text around it.

Use case 1: reports and insurance enquiries as a draft

An enquiry is in the inbox in the morning: cost approval for a therapy, with a deadline. The practice releases the enquiry and the matching extracts from the documentation to the AI colleague. Back comes a structured report draft in your own report format, with the details in the places where the insurer expects them, and with a source reference for every detail taken over. The doctor reads, deletes, adds the assessment and signs.

The same pattern carries the other reports: referral, discharge, disability insurance enquiry, and in home care the progress report to the insurer and the report to the child and adult protection authority (KESB). The accompanying text for the medical instruction or medical order under Article 8 of the health insurance benefits ordinance (KLV) is drafted as well. The needs assessment under Article 8a KLV and the care level classification stay in the specialist system and with the professional.⁹ The draft organises what is already documented. It invents nothing on top. Where a detail is missing, a marked gap remains instead of a phrase that merely sounds plausible.

This is exactly where the 54 minutes from the starting point sit.¹ What the draft deliberately does not contain is the clinical assessment: fitness for work, prognosis and indication are written by the professional. Nothing goes out automatically, every report waits for a signature.

Use case 2: the team’s knowledge becomes queryable

The medical practice assistant has a tariff question, is looking for the internal rule on a late cancellation or the right chapter in the practice manual. Today she interrupts the consultation for that. With a knowledge base built from the practice manual, the internal procedures and the billing documents of the business, she gets the answer with its source, checkable in seconds.

In the pharmacy the quality manual becomes queryable: “Where is our cold chain rule written down?” And the writing work for the quality management system is drafted from the existing documents, from the work instruction through the training record to the annual review. The medication check stays in the specialist system, where it belongs.

In home care it is the hygiene and emergency concepts and the onboarding of new employees, who ask in their own language and get answers from the same documents. Permissions apply down to individual tools: whoever may not see a folder does not see it through the AI colleague either, and every access appears in the audit log.

Use case 3: patient and customer information in several languages

A fact sheet on preparing for an examination exists in German, cleared by a professional. From it come the French, Italian and English versions, plus the languages actually spoken in the catchment area, and a plain-language version for everyone who is shut out by officialese.

The same applies to intake instructions at the counter and to aftercare instructions following a procedure. Before being handed over, every version goes through professional clearance, just as the German original did. What becomes noticeable afterwards: fewer calls back and fewer misunderstandings at the counter.

The line is clear here too. The AI colleague transfers and simplifies a cleared text; it adds no advice that is not in the original, and it leaves out no warnings. Quantities, timings and dosage figures are checked against the original before clearance. It does not replace the consultation.

Connectors: what can be linked to your industry software

An AI colleague is only as useful as what it is allowed to reach. A connector is nothing more than the link to another program. MCP is an open standard for such links. Through it, programs with a documented interface can be addressed. Each individual tool has its own permission, and every call is logged. In healthcare the starting position is particular: much of the field is closed. This is what an AI connection looks like for common software in practices, pharmacies and home care, checked in August 2026:

The honest summary: this field is largely closed. That is precisely why system connections are not at the centre of this article. Report drafts, the knowledge base and fact sheets work with documents and run without any link to the practice software. Where the specialist system speaks the FHIR data standard, one route runs through generic MCP servers. Generic means: they are tied to no single vendor, and they exist as open source.⁸ For service billing, the Forum Datenaustausch standard invoice 5.0 has been usable in production since 1 January 2026; the previous version 4.5 is retired on 30 June 2027. It is an open file format built on XML that can be read and written as a document.⁷ How a connection works in principle is described on the integrations page.

Data protection and professional secrecy: the strictest section

Article 321 of the Swiss Criminal Code names doctors and pharmacists explicitly. It covers their auxiliary staff as well. Under the revised Swiss Data Protection Act, health data counts as sensitive personal data. In the EU, Article 9 of the General Data Protection Regulation places it among the special categories of personal data: processing it is prohibited in principle and permitted only under the exceptions listed there, such as explicit consent or purposes of health care.¹¹ Taken together this means: a tool that processes such data is not introduced on the side. This article is not legal advice.

The first step is therefore the easiest: two of the three use cases in this article need no patient data at all. The knowledge base lives on the practice manual, the quality manual and internal procedures; the fact sheets are your own texts. You can start there without a single file leaving the business.

Where patient data does belong, namely in the report draft, the usual measures apply in strict form. In managed operation the instance runs in the ISO 27001 certified data centres of Green in Lupfig, or on your own server; the certification is held by Green, and one of our own is in progress. There is a data processing agreement and a public subprocessor list. Customer data does not train any models. The models from OpenAI (ChatGPT) and Anthropic (Claude) run in the AWS data centre in Frankfurt; AWS is a US company, so the US CLOUD Act applies there too, and we say so openly. With health data in particular, open frontier models such as GLM, Kimi and DeepSeek are therefore available from European providers on servers in Europe; the CLOUD Act does not apply there. Permissions reach down to individual tools, and the audit log can be exported.

What we do not claim: that patient data is anonymised in the process. A progress report remains personal data even without a name, and anyone promising otherwise is promising too much. The answer lies in where the processing happens and in access control, not in a text filter.

For Spitex organisations and institutions there is more: data of people lacking capacity is processed. Contact with the child and adult protection authority is part of the work as well. Deletion policies and access logging have to be set more strictly there than in a single practice. Care level assessments under BESA or RAI, the recognised instruments for determining care needs, are untouched by all this: they arise in the specialist system and are answered for there. The evidence on operations is available in the Trust Center.

Frequently asked questions from practices, pharmacies and home care

Sources

All figures in this article come from the following sources, last checked in August 2026.

1. gfs.bern on behalf of the FMH, survey on the medical working environment 2025 (N = 1,532, of which 330 in ambulatory care): cockpit.gfsbern.ch
2. mfe Swiss Family Docs together with the University Centre for Primary Health Care of the University of Basel, workforce study 2025 (published 2 October 2025): hausaerzteschweiz.ch
3. Federal Office of Public Health, administrative effort in healthcare, survey among general practitioners running since May 2026: bag.admin.ch
4. pharmaSuisse, quality in the pharmacy, quality assurance system under Article 30 of the Therapeutic Products Act since 1 January 2020: pharmasuisse.org
5. pharmaSuisse, dossier on the shortage of skilled staff, around 1,800 pharmacies with an average of 12.7 people and 9 full-time equivalents: pharmasuisse.org
6. Spitex Verband Schweiz, around 400 non-profit organisations with more than 40,000 employees: spitex.ch
7. Forum Datenaustausch, XML standards for service billing: forum-datenaustausch.ch. Invoice 5.0 usable in production from 1 January 2026, end of life of version 4.5 on 30 June 2027, per the CSS service-provider information: css.ch
8. WSO2, FHIR MCP Server, a generic open-source connection to FHIR systems, Apache 2.0 licence: github.com
9. Health Care Benefits Ordinance (KLV, SR 832.112.31), Article 8 “Medical order or medical instruction” and Article 8a “Needs assessment”: fedlex.admin.ch
10. Swissmedic, guidance document on medical device software BW630_30_007, version 3.0, valid from 21 April 2026, section 4 on qualification: swissmedic.ch
11. Legal bases: Article 321 of the Swiss Criminal Code (SR 311.0) on professional secrecy, fedlex.admin.ch; Article 5 of the Swiss Data Protection Act (SR 235.1) on sensitive personal data, fedlex.admin.ch; Article 9 of the General Data Protection Regulation (EU) 2016/679, eur-lex.europa.eu
12. FDHA Ordinance on the Electronic Patient Record (EPDV-EDI, SR 816.111), Annex 5 with the CH EPR FHIR (R4) Implementation Guide, version of 1 June 2026: fedlex.data.admin.ch