# Shadow AI is already on the floor

15 April 2026 · Essay · 8 min read

By [Axiomatic](https://axiomatic.digital/en/about), Maker of Inora

Just over half of the 890 care workers Nivel surveyed used AI at work in the second half of 2025. Two in three of them did so on their own initiative. The question for a board is not whether AI comes in but which one, and whether it can answer for that choice. A ban only takes away the view.

Last Thursday, 9 April, Nivel published a survey of 890 nurses, senior care assistants, support workers and other care workers about AI in their work. Every board should read two sentences from its summary.

“Just over half of the respondents used AI in care in the second half of 2025.”

“For two thirds of AI users, the initiative to start using AI lay with themselves.”

Read the second sentence again. Of the respondents who used AI, two in three started on their own. Not because their organisation chose a tool and brought it in, but because they had work to do and something was within reach.

That is shadow AI: AI used in the work without the organisation having chosen it, knowing it or being able to see it. It is no longer on the margins.

## What else it says

The rest of the summary completes the picture.

- AI is used most for looking up information and for reporting in the electronic client record.
- Half the respondents miss a clear AI policy for the whole organisation. Three in ten say their organisation has ethical guidelines for AI.
- Nearly nine in ten have never had any training in AI. More than two thirds feel insufficiently prepared.
- Two thirds consider themselves responsible when AI makes a mistake.

Put those lines together and you see a floor that uses AI for information and for reporting, often without a policy, mostly without training, and that takes the responsibility on itself.

That is no criticism of the floor. It is the opposite. If you have a question at three in the morning and no colleague to ask, you take what is there. Care workers have shown they are willing. They did not wait for anyone to ask them.

## Who answers for it

The two thirds who hold themselves responsible carry something the inspectorate places on someone else too. On 10 February 2025 the Health and Youth Care Inspectorate (IGJ) [wrote](https://www.igj.nl/documenten/2025/02/10/igj-roept-zorgaanbieders-op-ga-zorgvuldig-om-met-invoering-van-generatieve-ai-toepassingen): “Care providers who use generative AI in the provision of care are legally responsible for its careful implementation and use.” And under the Governancecode Zorg, the board is responsible for the general course of affairs.

The same publication says something about the kind of model. One of its headings reads: “‘General’ LLMs are mostly not intended for use in care.” The inspectorate advises a care institution to use a language model only when the supplier states that it is fit for the intended purpose. And acquiring and implementing one should follow a clearly described process: the risks mapped, the users trained, the system tested.

Put that beside Nivel’s report. Much of the AI on the floor was not chosen by the organisation. There was no process, no supplier stating what it is fit for, no training. The board answers for it all the same. Including the AI nobody bought.

## A ban does not remove the question

I understand the first reflex of a board reading this: forbid it. A rule in the staff regulations, a block on the network, an email from IT.

I think that is the worst answer. A ban does not remove the question. The senior care assistant who does not know at night how something is done in their organisation still has that question. They just no longer ask it on the ward computer. They ask it on their own phone, of a general model, outside the network and outside every agreement. That model does not know their organisation’s protocol and answers anyway, and it sounds just as sure when it is wrong.

A ban only takes away the view. The risk stays. The board no longer knows where it is.

So the question for a board is not whether AI comes in. It is in. The question is which AI, and whether the board can answer for that choice: to the inspectorate, to the works council, and to the employee who uses it.

## What a board can answer for

My answer is one chosen assistant, for everyone in the organisation. Not a trial here and a trial there, and not only for the care teams while everyone else works it out alone. One assistant in the start menu, about which the board has five things in writing before the first question is asked. I think every supplier should deliver those five. For Inora, this is how we deliver them.

**It is in the register.** The organisation receives Inora’s entry for its own register of AI systems, ready to file: the provider, the intended purpose, the risk class with its reasons, which data is processed where, and who answers for it. Under the AI Act, Axiomatic is the provider and the care organisation the deployer. The organisation also receives in writing, function by function, what Inora is for, what it is not for, and why it is not a medical device. That is exactly what the inspectorate asks a care provider to know: what the supplier says its application is fit for, and whether it is a medical device.

**Processing stays in Europe.** All processing for Inora stays in the European Union.

**Nothing is used for training.** No one trains a model on what staff put in: not on documents, questions, answers or reports, and not on conversations. Not us, and not any model provider.

**The councils know before the first question.** The works council must consent to rules on processing the personal data of staff; Article 27 of the Dutch Works Councils Act says so. So it receives a model consent request: what Inora processes and keeps, what the team lead sees per team and the board sees summed across all teams in the monthly report, that Axiomatic reads a conversation only when the organisation asks in writing, and that no report names or assesses an employee. The nursing advisory council receives one page on what Inora does and does not do, and how it declines what only the professional decides. The client council receives a brief on what changes for clients.

**Every team knows what it is for.** Nearly nine in ten of Nivel’s respondents have never had any training in AI, and the AI Act asks that people who work with AI know enough about it. So every team starts with a round on what Inora is for and what it is not: support, never a decision. Only then does it ask its first question.

None of these five is complicated. Together they make the difference between AI a board has to guess about and AI it can sign for.

## It has to beat the phone

One condition is in no register. The chosen assistant competes with the phone in the employee’s pocket. If it is slower, harder to find, or someone reads along, the phone wins.

On 6 September 2025, two days after the first care organisation started with Inora, I wrote that Inora had to be at least as smooth and fast as a general model, and that it annoyed me when it was not. Back then it was about a logo that flickered with every new conversation. Now I know why it matters. An assistant that loses to the phone has not replaced shadow AI. It has only opened a second window.

Three things decide it.

*Access.* At that first organisation, Inora was in the start menu of every workstation on 11 February, twelve days after it was asked for. For everyone in the organisation, not only for the care teams, as the AI that staff are allowed to use. In March three times as many people used Inora there as in December.

*The answer.* A chosen assistant has to know better than the phone. Inora answers from the organisation’s own protocols, work agreements and quality handbook, with the source alongside, day and night, and says so when the answer is not there. A general model on a private phone cannot do that, however fluently it writes.

*No one reads along.* On a private phone, no one from work reads along. At work it has to be the same. In Inora, no one in the organisation reads a conversation, managers and the board included, and reports on use cover teams of at least five people, never an employee. A member of staff who suspects their manager is reading along reaches for the phone again.

## What the floor has already decided

The floor has chosen: AI, now, with whatever is within reach. A board can no longer reverse that choice. It can still choose which AI it will be, and make sure it can sign for it.

---

*Postscript, 10 August 2026.* Four months later, an AI coordinator in long-term care put it this way in [ICT&health](https://www.icthealth.nl/magazine/editie-4-2026/organisaties-in-de-langdurige-zorg-krijgen-hulp-bij-ai-gebruik): “If staff do not use it professionally, they will use it privately. And if not them, then the clients and informal carers will. This is not going to blow over.”

## Sources

- Nivel, [Kunstmatige intelligentie (AI) in de zorg: de meningen van verpleegkundigen, verzorgenden, verpleegkundig specialisten, begeleiders en praktijkondersteuners](https://www.nivel.nl/nl/publicatie/kunstmatige-intelligentie-ai-de-zorg-de-meningen-van-verpleegkundigen-verzorgenden), 9 April 2026.
- Health and Youth Care Inspectorate (IGJ), [IGJ roept zorgaanbieders op: ga zorgvuldig om met invoering van generatieve AI-toepassingen](https://www.igj.nl/documenten/2025/02/10/igj-roept-zorgaanbieders-op-ga-zorgvuldig-om-met-invoering-van-generatieve-ai-toepassingen), its call to care providers on generative AI, 10 February 2025.
- ICT&health, [Organisaties in de langdurige zorg krijgen hulp bij AI-gebruik](https://www.icthealth.nl/magazine/editie-4-2026/organisaties-in-de-langdurige-zorg-krijgen-hulp-bij-ai-gebruik), on help with AI in long-term care, issue 4, 2026.
- [Governancecode Zorg 2022](https://www.governancecodezorg.nl/), the governance code for care.
- [Wet op de ondernemingsraden](https://wetten.overheid.nl/BWBR0002747/), the Works Councils Act, article 27.
- [Regulation (EU) 2024/1689, the AI Act](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689), EUR-Lex.

*Quotations from Dutch sources are translated.*

## 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.

## Read on

- [A language model does not reason. It supports yours.](https://axiomatic.digital/en/updates/a-language-model-does-not-reason.md): A language model does not reason: it picks the most likely next text, astonishingly well, and does not know what it does not know. The professional does the reasoning. Inora supports it, clinical reasoning included, with the organisation’s own knowledge and with the source shown. The professional decides.
- [The line lives in the code](https://axiomatic.digital/en/updates/the-line-lives-in-the-code.md): 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.
- [Our answers on privacy](https://axiomatic.digital/en/updates/our-answers-on-privacy.md): A care organisation asked how Inora handles its data. This is my answer of 1 October 2025: no training on conversations, no name attached to a message, storage in Europe, and a library that grows only after experts check it. What has changed since is dated beneath it.
