Skip to content

The line lives in the code

Most AI in care draws its line in a disclaimer: the user stays responsible. Our own answers of October 2025 did the same. The new architecture fixes Inora’s safety rules in its core, when it is built. Why the line belongs there, and how we count the errors at it.

Written by

Essay7 min read

One sentence from the section on safety in the system architecture of Inora 2.0, which I shared today, goes against something I wrote myself five months ago:

“Safety in Inora 2.0 is not a prompt concern. It is an architectural concern.”

Safety is not a matter of what you tell a language model. It is a matter of how you build the system.

The line in a disclaimer

Most AI in care draws its line in a disclaimer. The tool supports, it replaces no professional judgement, and the user stays responsible for what they do with an answer. The sentence is in the terms of use, sometimes in small print under every answer. And so the line is drawn, on paper.

We did the same. On 1 October 2025 I answered a care organisation’s questions about privacy and the responsible use of AI. About the user agreement I wrote three things. Inora supports and replaces no professional judgement. Users remain responsible for their own decisions. And however reliable the sources, users check every answer themselves.

Every sentence is true. The professional decides, and that does not change. But look at what the three sentences do together. They put the whole line on the user. What Inora would and would not do was not in Inora. It was in an agreement the user had accepted once.

What a disclaimer asks

Take the moment when the line matters. It is three in the morning. At a care location there is often one nurse in the building at that hour; if they are in doubt, they call the person on call or decide alone. The nurse types a question with someone in it, someone in front of them. You can ask a language model anything, including whether it is time to call the physician, or how much of a medicine someone may have. A model held only by an instruction answers that in sentences that sound sure, without their organisation’s agreements and without anyone having shown that it holds.

Then the disclaimer says: check it yourself. It asks for the most attention at the moment there is least of it, from the person who is most alone. A line that rests on one professional’s attention at three in the morning is not a line. It is a hope.

A line in the model’s instruction is not much better. An instruction can be talked away: with a cleverly worded question, with insistence, with a setting someone else switched on. What is in a prompt holds until someone asks for something else.

Whom the law points to

The disclaimer shifts responsibility to the professional. The law does not. The Dutch Health and Youth Care Inspectorate holds the care provider legally responsible for introducing and using generative AI with care, and the board answers for it. The inspectorate considers it important that a care provider knows whether a new application with generative AI is a medical device. Under the EU AI Act, Axiomatic is Inora’s provider, and the care organisation that works with it is its deployer.

And the EU Medical Device Regulation judges whether software is a medical device by everything its maker says about it: the label, the instructions for use, the sales material and the statements. A disclaimer is one sentence among all the others. It moves nothing.

So the organisation answers for careful use, and we answer for what we build. Neither can do that with a sentence in the terms. Only with what the system does.

What the new architecture fixes

In the new architecture, Inora’s safety rules sit in its core. They are fixed when Inora is built, not loaded while it runs. No organisation setting, no administrator and no instruction in a question can switch one off.

The architecture enforces them in three layers. The rules sit in the core. Every step, before it runs, is checked by code that does the same thing every time, not by a language model that might see it differently. And every check is logged, including the checks that stop nothing. That puts a rule layer in front of the model that splits every question into what a document answers and what only a professional decides.

Where the line is

Inora answers about sources and about what the care worker wrote, never about a client. It does not assess a client, decide urgency, calculate a dose or a score, or advise on a treatment. Past that line, software is a medical device. Inora is neither a medical device nor a high-risk AI system, and we build nothing that would make it either. That is also why it assesses no member of staff: the AI Act counts AI that evaluates the performance or behaviour of staff as high-risk. The decision stays with the person who answers for it.

Take a question like this: “A client has been restless since this evening and has a temperature. Should I call the physician now?” There are two questions in it. One part is in a document: what the organisation’s protocol says about restlessness at night, which observations the escalation agreement asks for, who can be reached at night, what you record. The other part is a decision about this client: is it urgent now? That part is in no document, and it is not Inora’s to make.

A system with only a disclaimer does one of two things. It answers the whole question, judgement included. Or it refuses the whole question, because a client is in it. Inora does neither. It answers the part a document answers: the escalation agreement word for word, with the passage alongside, and the observations it asks for and which are still missing. And it says who decides the rest, under the organisation’s own protocol: whether you call now is your decision. Inora never refuses a whole question because a client is in it.

Two errors, both counted

A line can fail in two directions. Wrongly answered: a judgement about a client slips through. Wrongly refused: a question the sources answer gets a no. It is tempting to count only the first, because only the first feels dangerous. A system that refuses everything with a client in it is flawless on that count, and helps no one at night.

This winter we heard it from the floor, in both directions. Care workers trusted an answer when they could check where it came from. Inora lost trust when it gave way as soon as someone contradicted it, and when it turned a question down too readily. Anyone who hears no too often asks somewhere else, of something with no line at all.

So from this architecture on, every new version of Inora first goes through at least fifty real questions from the night shift, and is counted at the line in both directions: wrongly answered and wrongly refused. Both counts go to the organisation’s quality team. Not an average on a slide, but two numbers for each version, which they can hold us to.

Who answers for what

That puts responsibility where it belongs, in three parts. The professional decides about the client. The organisation decides which sources lead and answers for careful use. And we answer for what Inora is for and where it stops. As managing director, I answer for that personally. In November 2025 I had Inora’s position under the AI Act and the Medical Device Regulation reviewed by an outside adviser. Whether Inora is a medical device is a question that should start with us, not with a quality team that would otherwise have to work it out alone.

Inora is built to make care professionals stronger, not to replace them. A tool that takes over the decision makes nobody stronger. A tool that shows what an answer rests on, and leaves the decision with the professional, does.

Back to October

The sentences of October were not untrue. They were in the wrong place. The professional still decides, and they stay responsible for what they decide. But they no longer have to guard, on their own, the line of what Inora does. It is in the code, at three in the morning too.

Postscript, 2 November 2026

Inora is not a medical device. Inora supports triage: the professional triages, scores and decides. Inora puts the organisation’s own protocol beside the professional, flags what is missing and records the source and version of every answer; it does not diagnose or decide on treatment. Its intended use keeps it outside the medical device rules (The professional decides).

Sources

Do you work in care and use Inora? Sign-in and help go through your own organisation: your team’s project lead and ambassadors can help you.

Talk through what your team needs next.