NOVEMBER 16, 2026

Training AI Agents on Medical Documents: What to Check First

Before a clinic or hospital commits to an AI vendor for document automation, here are the questions every procurement team should ask.

Omer Shalom

Posted By Omer Shalom

4 Minutes read


Clinics and hospitals weighing AI for healthcare document automation often run into the same wall: a vendor says they can train an AI on the clinic’s records, but the word means different things to different vendors, and the gap matters for compliance, security, and whether the system will work at all in Hebrew medical text.

Let's Talk About Your Project

What “training” actually means

Vendors use the word “training” to describe at least three different technical approaches. The first is fine-tuning: adjusting a model’s weights on a labelled dataset so it performs better on a specific task, requiring your data to leave your environment for processing. The second is retrieval-augmented generation (RAG): connecting a model to a document store so it pulls context at query time, without changing the model itself. The third is prompt engineering with examples, which some vendors call training even though nothing is learned from your data. Each approach has different infrastructure requirements, different data-handling risks, and different maintenance obligations after go-live.

The compliance layer you cannot skip

Israel’s Privacy Protection Authority treats medical records as sensitive personal data under the Privacy Protection Law, 5741-1981. Any processing of patient data by a third-party vendor, including training or indexing, requires a data-processing agreement that meets the Authority’s requirements. This applies whether the vendor runs the system on-premises or in a remote cloud environment. The Ministry of Health’s requirements for clinical software systems are published at health.gov.il.

Interoperability: ask about HL7 and FHIR

The international standard for exchanging healthcare information is HL7 FHIR (Fast Healthcare Interoperability Resources), and a vendor that cannot read or produce FHIR resources will create a data silo that is expensive to untangle later. Before signing, ask which FHIR resource types the vendor’s output supports and whether it can be ingested by your existing patient management system without a custom middleware layer.

The Hebrew medical text problem

Most general-purpose large language models were trained predominantly on English text. Hebrew medical documentation adds layers of difficulty: right-to-left rendering, inconsistent transliteration of drug names, mixing of Hebrew and English within the same record, and the use of medical shorthand that differs between institutions. A vendor that has not tested on Hebrew clinical text from Israeli hospitals or clinics is likely to deliver poor extraction quality on fields like diagnosis codes, medication lists, and referral summaries.

Five questions for every vendor

  • Where does the data go? On-premises, in a local data centre, or in a cloud region outside Israel? Moving sensitive personal data abroad requires either adequate protection agreements or explicit patient consent under Israeli law.
  • What is anonymised before training, and how? Ask for the specific anonymisation method, not just an assurance. Pseudonymisation is not anonymisation under the Privacy Protection Law.
  • Who owns the trained model? Some vendors retain the right to use your data to improve a shared model. That can expose patient data to other clients’ queries.
  • What happens on model update? If the vendor updates the base model, does the customisation need to be rebuilt? Who bears the cost and the risk?
  • How is performance measured on Hebrew text? Ask for precision and recall figures on a test set drawn from Hebrew clinical records, not English benchmarks.

Pricing and scope

Unclear scope is the biggest source of cost overruns in clinical AI projects. Palmidos works at a fixed price agreed before the project starts.

What a compliant engagement looks like

A well-structured engagement for a clinic or hospital starts with a data audit: which document types, which fields, which systems produce them. The vendor defines the task precisely: extraction, summarisation, classification, or question-answering, before selecting a technical approach. Data is anonymised before leaving the clinical environment, the vendor signs a data-processing agreement, and a working prototype is delivered on a representative sample before full deployment.

More articles that may interest you

Tell us what you need

One of our engineers will get back to you within 24 hours.