AI in the Reporting Office: Confidentiality First

How it works today in most reporting offices

The channel is up. The reporting form is published, the QR code is on the notice board, responsibilities are settled. Now begins the part rarely discussed before the rollout: handling the reports that actually arrive.

In mid-sized organizations the internal reporting office is staffed by one person, rarely two. Usually it is the compliance officer in a double role with legal, HR or internal audit. Reports are a secondary process there: they arrive unannounced, rarely, and not as a completed form but as text. Two screens of prose without paragraphs, partly anonymous, partly in broken German or a language nobody in the building speaks. Plus three attachments: a photographed shift schedule, a screenshot from a chat history, a PDF without a meaningful file name.

By the time someone has understood what the matter is about – which period, which department, which allegation, which person is affected and which merely mentioned – half a working day has gone. Then the drafting begins. The acknowledgment of receipt is written by hand, because the case is precisely not standard. Risks are noted in keywords, measures described in a paragraph, and the feedback after three months again produced by hand, often from memory and the email trail.

Meanwhile the clock runs: seven days until the acknowledgment of receipt, three months until the feedback. Both deadlines start on receipt, not on the day the responsible person gets back from vacation.

The shortcut nobody documents

Here a behavior arises that no procedure description mentions. To save time, someone selects the facts, pastes them into a public chatbot in the browser and asks for a summary, a risk assessment or a draft text. The result is usable. The process is not.

Transmitted is everything on the clipboard: the complete set of facts, the named employees and managers and, in the worst case, the contact details of the reporting person. Copy and paste knows nothing about data minimization. Where the content went, under which account, with which provider and for which purpose, nobody can say afterwards. No log exists, because it took place outside every system.

The moment it becomes visible

It becomes visible in three situations: when management or an auditor asks which tools reports are handled with and where content has been transmitted; when a reporting person wants to know who has read their case; or when a deadline has been missed and someone has to reconstruct what the handling was waiting on. The question then is not "Do you use AI?" but "Who transmitted which content to whom and when – and where is that documented?"

The target process in six steps

The difference between permissible support and an incident lies not in the model but in the process. These six steps describe it.

1. Switch AI on deliberately instead of using it casually. The feature is activated per team, exclusively by administrators, and activation requires explicit confirmation of consent to transmitting data to the chosen provider. On top of that comes a personal approval per user. Only with both in place do AI elements appear in a report at all.

2. Determine the provider yourself. The access comes from the customer: your own API key, stored encrypted, shown only masked in the interface and tested against the configured model on saving. Alongside the supported providers, any OpenAI-compatible endpoint can be entered, including a self-hosted one. Then the content does not leave your own infrastructure.

3. See what will be transmitted before the run. The start panel names the configured provider and points out the transmission before the analysis begins. Transmitted are the facts and messages of the report plus the optional additional notes. The name and email address of the reporting person are not transmitted.

4. Structure the handling rather than speed it up. Four functions apply: a summary as its own section in the report, a risk analysis with likelihood of occurrence and severity of impact, proposed measures including ownership for implementation, and a message draft for the correspondence. The result is not a faster decision but a uniform basis for your own.

5. Review every suggestion individually. Suggestions are never adopted automatically. Adoption runs through a selection list, each suggestion can be deselected, and all fields remain editable before saving. Adopted entries are then ordinary risks and measures, logged like manually recorded ones.

6. Keep evidence. Every call lands in the AI activity log, successful or failed, with time, status, user, function, model, provider, duration, attempts and token consumption. Request and response content is stored encrypted and cleared automatically after seven days; the call metadata is retained.

What AI in the reporting office does not do

It does not decide. There is no automatic change of status, no deadline set or postponed and no decision about a report. Nor is there a legal assessment or an assessment under whistleblower protection law: whether a set of facts falls within the scope of the German Whistleblower Protection Act is for the reporting office to decide. Nothing is preset – without activation and approval the feature does not exist in the interface. preeco | whistleblower is a platform of its own and, beyond that, writes no system log files, so operating the application allows no inference about reporting persons.

Before and after in direct comparison

What the switch means in practice

The effort lies not in the technology but in two decisions taken before activation. First: which provider do we work with, and is there a contract including a data processing agreement? Second: who in the reporting office may use the feature, and who may not? Both are management, not IT questions. preeco supplies neither model nor access and is not a contracting party of the AI provider; reviewing the provider remains with the organization.

Anyone unwilling to answer both questions leaves the feature switched off. That is a workable state: on delivery it is deactivated, and handling reports works without it.

Three mistakes that make use in the reporting office indefensible

Treating activation as a technical detail. Confirming consent to the data transmission is not a formality you click away. It is the point where it is settled that report content goes to a third party – with everything that hangs on that contractually and in a third-country transfer.

Taking the suggestion for the result. A proposed likelihood of occurrence is a hypothesis, not an assessment. Anyone adopting suggestions unchecked moves the assessment to a place that does not know the case – and ends up documenting a judgment nobody can justify.

Filling additional notes with personal data. The field for additional notes serves the analysis perspective, not the case file. Anyone entering names there cancels out the data minimization the feature observes elsewhere.

How to tell that it is time

• Case content in your reporting office has already been copied into a public AI tool once.
• You cannot prove which tools a report was handled with.
• The acknowledgment of receipt is regularly delayed because reading and sorting takes time.
• Risks and measures are named differently in every case, and analysis across cases does not work.
• You have an AI provider in the organization but no rule on where it may be used and where not.