The problem
Preparing anonymised case summaries for monthly supervision means rewriting notes by hand to strip out names, places and workplaces. At 6 cases a month and about 30 minutes each, that's 3 hours a month. Cloud AI tools aren't suitable, because identifiable therapy notes are special-category health data.
What the AI does
Ollama is a free app that runs open AI models on your own Mac, Windows or Linux computer, and Meta's Llama 3.1 8B model is a 4.9 GB download under a royalty-free license. Once set up, you paste an extract of your own notes into the local chat and ask it to replace names, places, workplaces and dates with placeholders and summarize it in your supervision template. The text is processed on your computer, not sent to an AI provider. You read every line, remove anything identifying it missed and add your own clinical reflection.
A worked example
A counselor in private practice in York sees about 20 clients a week and takes 6 cases to monthly supervision. They installed Ollama and Llama 3.1 8B on an encrypted laptop and use it only for drafting supervision summaries.
Before
About 30 minutes per anonymised summary: 3 hours a month
After
About 12 minutes per summary checking and editing: 1.2 hours a month
What it's worth
About $120£90€107A$173C$169 a month in time saved, at no cost on a computer you already have. Summaries drop from 3 hours a month to about 1.2, and 1.8 hours at an assumed $67£50€60A$96C$94 an hour is $120£90€107A$173C$169.
How to set it up
- Turn on full-disk encryption first: FileVault on a Mac, or BitLocker or device encryption on Windows.
- Download Ollama from ollama.com and install it, then download the Llama 3.1 8B model (the 4.9 GB one; Ollama's page lists the command 'ollama run llama3.1').
- Check you are using that downloaded local model, not one of Ollama's cloud models, and don't sign in to cloud features.
- Paste one note extract and ask: 'Replace every name, place, workplace and date with a placeholder, then summarize in these headings: presenting issue, themes, my questions for supervision.'
- Read every line, remove anything identifying it missed, add your own reflection, then delete the chat.
What goes wrong
- A small local model can miss identifiers or invent details. You stay responsible for confidentiality under UK GDPR and your professional body's ethical framework, such as the BACP's.
- It doesn't make clinical judgments. Nothing it writes goes into the client record, and decisions stay with you and your supervisor.
- Ollama also offers cloud models and paid cloud plans. Cloud prompts leave your computer, so use only a downloaded local model for this.
- Speed depends on the computer: Ollama says large models are slow without a strong GPU. If the 8B model is too slow, the job isn't worth forcing; this is a technical setup.