TrompTechLast updated: 1 October 2026. This text has been checked against the software itself.
TrompTech builds software that contains AI and also uses AI in its own work. Here you can read where you come across AI, what data a language model gets to see, where a person decides and when something happens automatically.
1. On this website
- The chat at the bottom right is not AI. It chooses from fixed answers and does not send your questions anywhere. If you leave a message in the chat, it reaches Jeremy as an enquiry, just like the enquiry form.
- An enquiry is not assessed or answered by AI. No automatic reply is sent: Jeremy reads your enquiry and responds himself.
- Afterwards, in our own workspace, AI can summarise your details or suggest a next step, as help for Jeremy. This is done with Claude (Anthropic), ChatGPT and Codex (OpenAI) or Gemini (Google). Your email address and phone number are not included in a summary or advice.
- Copy. We sometimes use AI when writing copy for this site and our messages. Nothing goes online without Jeremy having read and approved it; he is responsible for the content.
- Images. We do not use AI-generated images of real people, places or events.
2. AI in TrompTech Layer
Layer uses language models to summarise, prepare and answer questions: Claude from Anthropic, and until the switch described below also ChatGPT from OpenAI, configurable per agent. On the EU route, Gemini from Google is added for light work such as sorting and summarising. Customers use them through TrompTech, included in the subscription. Never through a personal subscription. A provider key of your own is only possible if TrompTech enables it for the workspace.
Where it runs: we are moving customers’ AI to processing in the EU, via Google Cloud (Vertex AI) or AWS (Amazon Bedrock), in an EU region. We notify customers of this switch 30 days in advance. If the EU route is ever unavailable, the AI does not run: nothing is then sent to a language model outside the EU. Until the switch, the language model runs at Anthropic or OpenAI, in the United States, under the arrangements on Sub-processors. You can see where your workspace’s AI currently runs in Layer under Settings, AI. What we do to send as little as possible:
| Subject | How it works |
|---|---|
| By default only the header and the first 600 characters, without earlier messages in the thread. The full text only with a stated reason, and at most five emails at a time. | |
| Customer summaries | Name, company, amounts and notes are included. Email addresses and phone numbers are not, even if they appear in a note. The data is not otherwise anonymised. A summary is made at most once a week, when someone opens the contact. |
| Staff via integrations | No contact details and no notes. |
| Passwords and keys | Never reach the language model. |
| Speech to text | With speech recognition on your own computer or server (Whisper), the recording goes nowhere. If you choose the browser’s speech recognition, Google (Chrome) or Apple (Safari) processes the recording. |
| Read aloud | On your own computer or server. |
| Retention | AI conversations and completed proposals are kept for 12 months, then deleted. If you delete a contact, the administrator chooses whether the AI conversations and proposals about that contact are deleted too. |
A person decides, unless you set it up otherwise
Everything that leaves the workspace (an email, a change in another system, a social media post, a reply to a review) starts as a proposal. A person approves it, edits it or rejects it. Only then does the server carry it out. This is how every type of task is set when you start.
A type of task only becomes automatic once it has proven itself and an administrator agrees to it. The software suggests it; it never switches it on by itself. When it suggests it depends on what the task does:
- Internal and reversible (creating a task, recording a contact moment or note, proposing hours, and changing the status of a project, lead or ticket, except marking a lead as won or lost): after at least 10 approvals without changes, over at least 7 days, and at least 80% of all decisions without changes. What then happens automatically can be reversed within 30 days with Undo, as long as nothing about it has changed in the meantime.
- Everything else that may happen automatically, including what goes outside (such as an email the AI wrote, sending an invoice, a first payment reminder, or three small changes in Google Ads that cost no money and can be reverted there with one click): only after at least 30 decisions over at least 4 weeks, and with a safe margin at least 80% approved without changes: with 30 decisions that is 29. A sent email cannot be recalled.
That 80% is the default; an administrator can set it per type between 50% and 99%.
Even then, 1 in 5 proposals still goes to a person, an automatic proposal is first shown as ready for 5 minutes, at most 25 go out by themselves per day, a type returns to approval after a complaint or a drop, and an emergency stop puts everything back on approval.
Never automatic: payments, formal reminders and other decisions about money (such as a new monthly amount or a final invoice), sending quotes, changes through an integration with another system (apart from those three small changes in Google Ads), HR matters, searching the internet and replies to reviews. And if a conversation has read text from outside, such as an incoming email or a web page, every action after that requires a person.
An email written by the AI that goes out without a person gets this line at the bottom: “Dit bericht is automatisch opgesteld met hulp van AI. Antwoorden lezen we zelf.” Standard emails that the software composes itself, such as an invoice or a reminder, are not AI and do not get that line.
Guidance for whoever approves
The first time someone approves proposals, a short explanation of what matters opens automatically. This is how we work on the AI literacy that Article 4 of the AI Act requires.
AI makes mistakes
Language models sometimes make mistakes. That is why everything starts as a proposal, something only becomes automatic after proven trust, and even then a sample keeps going to a person.
3. What we do not do
We do not use AI to:
- recognise people by their face, voice or another physical characteristic;
- recognise the emotions of employees or customers;
- assess or rank people for work, such as in recruitment or staff scheduling;
- mislead anyone or put them under pressure.
4. Prefer no AI?
Email info@tromptech.nl or call +31 6 17720360. You will speak to Jeremy.
For Layer customers: every type of task is set to approval by default. Something only becomes automatic if an administrator switches it on.
5. Questions or complaints
Email info@tromptech.nl. If you disagree with how we handle your data, you can also file a complaint with the Autoriteit Persoonsgegevens (the Dutch Data Protection Authority).
See also: Privacy statement, Sub-processors and Documents & agreements.