How Our AI Works – AI Literacy Guide

Onboarding and training guide for hotel staff (users) and the internal Discontinue team

Last updated: 7 July 2026 · Deutsche Version

This guide explains in clear language how the artificial intelligence (AI) in Discontinue works, what it can do, what it deliberately does not do, and what you should pay attention to when working with it. It is part of our measures for AI literacy under Art. 4 of the AI Act (EU 2024/1689).

1. Purpose & scope (Art. 4 AI Act)

Art. 4 of the AI Act requires that providers and deployers of AI systems ensure a sufficient level of AI literacy among all persons who work with the system – that is, among their own staff and among the users. This obligation has applied since 2 February 2025.

This guide fulfils this obligation for Discontinue and is aimed at two groups:

• Hotel staff who use Discontinue (creating reports and agents, receiving messages, granting approvals).
• The internal Discontinue team, which operates and develops the platform.

The aim is for you to be able to correctly assess AI results, use them sensibly, and recognise and report errors.

2. Basic understanding: what is the AI in Discontinue – and what is it not?

Discontinue is an AI-assisted operations platform for hotels that sits on top of the customer-connected source systems (e.g. a property management system). The AI is an assistant that helps you turn your hotel data into useful reports and workflows more quickly.

Think of the AI as a very diligent assistant that:

• reads your data from the connected systems (read-only by default, via granular, configurable tools – each agent only with what its task requires),
• structures reports from it and phrases them clearly,
• composes messages that your agents send,
• helps you set up reports and agents in natural language.

The AI is explicitly not an autonomous system that decides on its own about your operation, your guests or your employees. It is a tool in your hand: we provide the tools, you build with them what you need.

3. What the AI does – and what it does not

Key takeaway: The AI structures and phrases. It does not decide, it only has the permissions you give it per agent, and it writes only with your consent — individually via the approval queue, or generally via your explicit per-agent opt-in.

4. "YOU build your reports and agents – the AI helps you do it"

This is the most important idea of this guide. In Discontinue, you are the designer:

• You decide which report or agent you need.
• You describe it in natural language – the AI translates your description into a configuration.
• You determine when it runs, which data it uses and via which channel it delivers (email, chat/messaging services you have connected, file download).
• You approve or reject write operations.

This has an important consequence for responsibility: what you build for yourself (e.g. a particular report or an agent) is your responsibility as the operator of your own data – as with any automation platform. Discontinue provides the tool; you design the specific application.

Counterpart: Templates that Discontinue curates itself in the marketplace are our responsibility. That is why we deliberately offer no templates that evaluate, rank or "score" guests.

5. Why the data values are correct – but the wording should be verified

A common misconception: "If AI is involved, the figures might be made up." At Discontinue this is not the case:

• The data values come from the connected source system (your live data) and are calculated deterministically in Python for metrics such as ADR, RevPAR or occupancy – according to fixed formulas, not estimated by the AI.
• The AI element lies in the structuring and wording – that is, in how the figures are arranged, summarised and put into words.

What this means for you:

• The underlying figures are reliable (connected source system + fixed calculation).
• Nevertheless, the AI can make mistakes in structuring and wording – for example, misclassify a figure, describe a relationship in a misleading way, or phrase a note inappropriately.

Checklist – verifying critical figures:

• Does the report's statement match the figures you expect?
• Are the period and property correctly named?
• Does a figure look unusual? Then cross-check against the connected source system before basing a decision on it.
• Is the wording unambiguous, or can it be misunderstood?

6. Dealing with possible errors – and how to report them

AI systems can be wrong. That is normal and no cause for concern, as long as you know it and handle it correctly.

If you notice something:

1. Do not rely on the wording alone for important decisions – cross-check the critical figure in the connected source system.
2. Briefly document what you noticed (which report/agent, which run, what was striking).
3. Report it to us: by email to office@discontinue.dev. We follow up on every notice.

You always recognise AI content by the marking – e.g. by the report footer "This report was generated automatically … please verify critical figures" or by the note "AI-generated" on agent messages.

7. What run traces are – and how you read them

For every agent run, Discontinue records a complete run trace. This is a traceable record of what happened in that run. For this we use Langfuse (Langfuse Cloud, EU region); processing takes place in the EU.

A run trace shows, among other things:

• the trigger (what initiated the run),
• a snapshot of the system prompt,
• the tool calls with their parameters and results (which data was read),
• the model used and the consumption (shown in euros),
• the deliveries (to whom/via which channel),
• any errors.

How to read a trace:

1. Select the relevant agent run.
2. Look at the trigger and the data read – this explains what the run is based on.
3. Follow the tool calls from top to bottom – this shows how the result came about.
4. Check the delivery and any error messages.

This way you can always understand how a result came about. Run traces are retained for 12 months, then anonymised or deleted.

8. Understanding human-in-the-loop / approval queue

Discontinue never alters data in your connected systems without your consent. You define what an agent may do via its permissions — every scope and every tool is set separately per agent. In addition, by default every write operation goes through an approval queue:

Proposal by the AI → Approval queue → Approval by an authorised human → Execution

• The AI can propose a write operation (e.g. a note, a folio adjustment, a maintenance ticket).
• It is only executed once an authorised human designated by you approves it.
• Every step is logged.

We call this human-in-the-loop: the human retains control over all changes. Reading operations (retrieving data, creating reports) run without approval; write operations by default never run without individual human approval. The only exception: you deliberately activate the opt-in for an individual agent — then that specific agent executes write operations without individual approval, still only within the permissions you have set separately. You make and confirm this decision yourself per agent, it is revocable at any time, and all operations remain fully logged.

9. Best practices & responsible use

Do this:

• Use the AI for structure, overview and wording – it saves you time.
• Verify critical figures before basing decisions on them.
• Handle approvals in the approval queue deliberately and with scrutiny – they are your safety net.
• Report anomalies so that we can improve the platform.

Don't do this:

• Do not build any evaluation, ranking or "scoring" of guests or employees.
• Do not pursue automated decisions about individual persons (this contradicts Art. 22 GDPR and our platform guardrail).
• Do not rely blindly on a wording without knowing the figure behind it.
• Do not process any special categories of personal data (Art. 9 GDPR) via the platform.

Why "no guest scoring"? Such applications can fall into the high-risk area of the AI Act and interfere with the rights of individual persons. Discontinue is deliberately designed as a Limited Risk system – without profiling, without scoring, without decisions about individuals. Please adhere to this.

10. Notes for the internal team training

This section is aimed at the Discontinue team and supplements the customer documentation. It also ensures AI literacy under Art. 4 of the AI Act internally.

Training content (minimum scope):

• Functioning, limits and typical sources of error of the AI system used.
• Discontinue's role as provider of the AI system; delineation from the provider of the GPAI model (Anthropic) and the deployer (hotelier).
• Transparency obligations under Art. 50 (AI labels, report disclaimer, transparency page).
• Monitoring and logging obligations (run traces in Langfuse, 12-month retention, planned weekly trace review and alerting).
• Platform guardrail: no marketplace templates with rankings/scoring/evaluations.
• Data protection basics: roles (controller/processor), no-training guarantee and limited retention (max. 30 days), third-country transfer (EU SCCs + TIA).

Evidence & documentation of the training:

• Participants, date and content of each training session are recorded.
• New team members are trained during onboarding.
• A refresher takes place on an ad-hoc basis (e.g. for a new model provider or a significant functional extension) and as part of the regular compliance review.
• The training documentation is retained together with the AI Act Compliance Dossier.

11. Contact

discontinue.dev MAS GmbH Bruno-Marek-Allee 5, 1020 Vienna, Austria Managing Director: Stefan Starflinger Internally responsible for compliance: Adrian Schmidt (Business & Compliance)

• General enquiries / error reports: office@discontinue.dev
• Data protection: datenschutz@discontinue.dev

Related documents: AI transparency page · Privacy Policy · AI Act Compliance Dossier (internal) · Data Processing Agreement (DPA)

discontinue.dev MAS GmbH · Last updated: 7 July 2026