Local AI for neuropsychologists: cloud-free reports
In short: local AI allows a neuropsychologist to automate the writing of their reports, the transcription of their interviews, and the calculation of their test scores, without any patient data leaving their computer. Unlike online subscription solutions, all processing runs on a workstation installed in the practice. We designed and delivered such a solution in 2026: dedicated hardware configuration, locally installed AI models, and a custom-developed reporting application built around the practitioner's actual workflow.
Report writing is the most universally shared point of friction for neuropsychologists. A full assessment represents several hours of administration, followed by an entire evening of writing at home. The question is no longer whether AI can help: it can. The real question, in a profession that handles health data, is where this data is processed.
This article explains concretely what local AI can do for a neuropsychology practice, what it absolutely should not do, and details a real project we carried out from start to finish.
Why the cloud is a problem for a neuropsychology practice
A neuropsychological assessment report contains a patient's identity, often a minor, their date of birth, their educational background, their medical and family history, and their psychometric test results. This is health data within the meaning of the GDPR, in its most sensitive category.
Using an online AI service to write these documents raises three distinct questions:
- Transfer. Data leaves the practice and travels to a third-party server. It is then necessary to ensure the hosting location, HDS (Health Data Host) certification when required, and the subcontracting chain.
- Dependence. A monthly subscription means you never own the tool. The price can change, so can the conditions, and termination of the service ends its use.
- Fixed scope. SaaS platforms offer a finite number of report templates. If your scoring method or framework doesn't match exactly, you adapt your practice to the tool rather than the other way around.
A local installation answers all three: data does not leave the machine, the hardware and software belong to you permanently, and the application is built around your existing framework.
Important clarification: local data processing removes the issue of transfer to a third party, but does not exempt the practitioner from their obligations as a data controller (register, patient information, workstation security, consent to interview recording). Local AI significantly simplifies compliance; it does not replace it.
What local AI can do, concretely
On a properly sized workstation, current open-source AI models cover four needs directly useful for an assessment:
1. Transcribe an anamnesis interview
A voice transcription model (Whisper) runs locally and converts the recording of the interview with the parents into text. No audio file is sent anywhere. On a recent graphics card, one hour of interview transcribes in a few minutes.
2. Write the anamnesis from this transcription
A local language model takes the raw text—with its hesitations, repetitions, unfinished sentences—and formats it into reported speech, in professional written French. The practitioner reviews and corrects: the goal is to eliminate the blank page, not the review.
3. Read hand-filled scoring sheets
A vision model (AI-powered OCR) transcribes images of scoring sheets, including scanned tables or those filled with a stylus on screen, into structured, usable data.
4. Query your own documents
A RAG (Retrieval Augmented Generation) system indexes the practice's documents and allows them to be queried in natural language, always without a connection.
What AI must not do in an assessment
This is the point we consider most important in the project, and what distinguishes a serious solution from a seductive demonstration.
A language model must never calculate or decide a score. Language models are probabilistic: they produce what looks like a plausible answer. For a z-score or clinical classification, this approximation is unacceptable.
In the solution we delivered, the separation is strict:
| Task | Performed by | Why |
|---|---|---|
| Z-score calculation | Deterministic code | A division. Reproducible, verifiable by hand. |
| Classification (normal / fragile / pathological) | Deterministic code | Comparison to a fixed threshold. |
| Table formatting | Deterministic code | No interpretation possible. |
| Writing narrative sections | Local language model | Reformulates already calculated results, without re-discussing them. |
| Diagnosis, clinical interpretation | The practitioner, exclusively | Outside the scope of the tool, by design. |
In other words: the machine calculates what can be calculated, puts what has been calculated into sentences, and stops there. Interpretation remains entirely the responsibility of the professional.
Concrete case: a custom AI workstation for a neuropsychology practice
In 2026, a private practice neuropsychologist contacted us with a specific need: to automate the production of her cognitive assessment reports, without uploading any patient data online. Here's how the project unfolded.
The starting point
Her workflow relied on scoring sheets filled with a stylus, Excel files whose content was entirely composed of images, PDF conversion tables, and a personal report template refined over the years. No off-the-shelf platform matched this organization.
Hardware configuration
| Processor | AMD Ryzen 5 7500F |
| Graphics Card | NVIDIA GeForce RTX 5060 Ti 16 GB |
| RAM | 32 GB DDR5 6000 MHz |
| Storage | 1 TB NVMe SSD |
| Peripheral | Conference micro-speaker, capture up to 4.5 m |
The 16 GB of video memory is the key differentiator: it comfortably runs a 12 billion parameter language model and a vision model, without relying on the processor.
Installed software layer
- Ollama, with two pre-installed general-purpose language models
- A vision model for reading scoring sheets
- Whisper for voice transcription
- A document RAG system
- Secure remote access for technical support
The custom-developed application
This is the heart of the project. Rather than imposing a generic tool, we developed an application that precisely replicated the practitioner's report framework, organized into five tabs that follow her actual workflow:
- Interview — recording via microphone or import of an existing audio file, followed by automatic transcription and anamnesis drafting.
- Patient File — header fields (name, age, class, reason for consultation) are pre-filled from what was said during the interview, and visually flagged for verification.
- Scores — entry of raw scores noted on scoring sheets, with immediate calculation of z-score and status, and keyboard navigation to enter an entire sheet without the mouse.
- Report Sections — automatic drafting of intellectual efficiency, attentional capacities, summary, and conclusion, based on the already calculated table.
- Raw Transcription — what the microphone actually captured, consultable in case of doubt about a reformulation.
The export produces a Word or OpenDocument file retaining her complete layout, including logo and contact details, automatically classified by year and month.
A telling detail
When validating our calculation module against a real, already submitted report, we recalculated twenty-seven lines of results. Twenty-six matched exactly. The twenty-seventh had an inverted sign — a manual entry error, invisible upon review, in a document already transmitted to the referring physician.
This is the most accurate argument in favor of this type of tool: it is not about replacing the practitioner's expertise, but about eliminating mechanical errors that fatigue makes inevitable after several hours of assessment.
How much time does this save?
The gain is on three fronts: the blank page of the anamnesis, the manual copying of scores and calculation of z-scores, and the final document's layout. In practice, a practitioner who used to spend an evening on a report now moves to a review and correction — the observed order of magnitude is a 50% reduction in writing time, varying with the complexity of the assessment.
The often underestimated point: the ability to perform two assessments in the same day without having to take the writing of the first one home.
Frequently Asked Questions
Is local AI truly GDPR compliant?
Local processing eliminates the transfer of data to a third party, which removes the issue of hosting and subcontracting. However, the practitioner remains the data controller: they must maintain their register, inform their patients, secure their workstation, and obtain consent before recording an interview. Local AI greatly simplifies compliance; it does not replace it.
Do I need an internet connection to use the solution?
No. Once installed, everything works offline. A connection is only useful for updates and remote support, if you wish.
Is a local model less powerful than an online model?
For complex reasoning tasks, proprietary online models still have an advantage. For formatting dictated text and structured report writing from provided data, a local 12-billion-parameter model is perfectly suited. The relevant criterion is not raw power, but suitability for the task.
Does the AI write the diagnosis?
No, and this is a design choice. The application is explicitly constrained to describe already calculated results, without formulating clinical hypotheses, diagnoses, or recommendations. Interpretation remains entirely the responsibility of the practitioner.
Can the solution be adapted to other tests or professions?
Yes. The logic — data entry, deterministic calculation, assisted drafting, export to your template — can be applied to other test batteries and other regulated professions: speech therapy, clinical psychology, medicine, expert appraisal.
What happens in case of a technical problem?
Remote access software is installed, which can only be activated when you decide. The machine also benefits from our 2-year warranty.
For whom does this approach make sense?
A local AI workstation becomes relevant when three conditions are met: you handle data that you don't want to transfer online, your workflow is too specific for a generic platform, and the volume of documents produced justifies automation.
This is typically the case for neuropsychologists, speech therapists, clinical psychologists, private practice doctors, lawyers, and accountants — all professions where professional secrecy and document production intersect.
Let's discuss your case
Every practice has its own workflow. We study your current organization, then honestly tell you what can and cannot be automated.
Discover our local AI workstations · Request a personalized study
