A cartel report in four languages: an internal investigation run with Claira

The client makes industrial valves in Germany and sells them through subsidiaries in Italy and Poland. A sales manager at the Italian subsidiary uses the group's internal reporting channel to say that colleagues have been discussing prices with two competitors on the margins of trade association meetings. Under the Hinweisgeberschutzgesetz, Germany's whistleblower protection act, the reporting office must acknowledge the report within seven days and give the reporter feedback within three months.

We are outside counsel, instructed by the management board, and the board needs to know quickly whether the report holds up. The Bundeskartellamt grants full immunity from fines only to the first cartel member to reveal a cartel it did not know about, and a marker can hold that place while an application is completed.

The record arrives in four languages. Head office writes in German, the subsidiaries in Italian and Polish, and the group uses English between them.

IT collects the mailboxes, Teams chats and voice messages of the sales team, the managing directors and the in-house lawyers from 2019, along with the association's agendas and minutes and scans of the sales managers' notebooks. Once the data is in Nuix Discover, the team wants the shape of the set before anyone reads a document.

Early case assessment runs from the Claira menu under More, with a click on Run assessment. It reads metadata only, writes nothing to Discover and covers the whole case: custodians, date range, file types, the languages present, senders and recipients from the People fields, and a timeline of spikes and gaps.

The language count settles staffing by showing how much of the set needs an Italian or Polish reader. If a custodian's mail thins out around association meetings, that becomes a question for the interviews. The assessment downloads as a PDF for the investigation file.

The team starts a Case Context from the Internal investigation template, which creates an Investigation Brief with sections for the people involved, the allegations and background, the investigation questions and the confidentiality rules. They select five key documents in Discover, here the report, two meeting agendas, a price-increase letter and the organisation chart, and click Generate from selection. Claira reads the extracted text and, where there is one, the native file, so a scanned agenda can feed the brief.

Claira then lists up to five questions the documents leave open, and Answer with Agent opens a chat to settle them. The answers that matter here are the competitors' names, the attendees at each meeting, and the nicknames and abbreviations used in Italian and Polish mail. The confidentiality rules record that the reporter's identity stays out of the review. A partner approves the brief, Claira records who approved it and when, and prompts are then built from the approved version.

The review prompt starts from the cartel screening workflow. It asks for a risk call from High to None, the conduct types, the parties involved, a short justification and supporting excerpts, and it reads for hedged language and for moves off the written record, such as a suggestion to call instead of writing. The team adds two fields from the translation workflow, the primary language and a two-sentence English summary. Claira detects the language itself, so nobody sorts the set by language first. The prompt is tried in single review on a document in each language.

A bulk scan then runs it across the collection and writes each answer to its own field in Discover. In Auto mode Claira picks the mode for each document, so the notebooks are read as images and the voice messages as audio. The team writes the prompt to each document's Prompt Log, sends the scan to the background and leaves a comment on the task saying what it was for.

The team samples by language and by risk level, starting with the Medium calls, where subtler signals tend to sit. A reviewer fluent in Polish reads Polish documents against their English summaries, and an Italian speaker does the same for Italian. The summaries speed up triage, but they are not certified translations, and models handle some languages more reliably than others, so each language gets its own sample.

Where reviewers have already coded documents, the QC review workflow applies the relevance criterion to their work so that disagreements with Claira rise to the top. A disagreement often traces back to something the brief left out, such as a competitor's nickname or an Italian word for a price list. The team corrects the Investigation Brief, revises the prompt and reruns it on the documents affected. Each version of the prompt is kept, and a bulk scan that did not run the version marked final carries a note saying so.

With the calls checked, the team builds an Insight from the bulk task in Timeline format. The prompt asks for a table of contacts between the client's staff and competitors, association meetings and price announcements, with columns for the date, the people, the language of the source and what was said, and a source document on each row. Claira reads the field values as they stand when the Insight is generated, so reviewers' corrections are included. It checks each citation against the documents the Insight read and removes any that fail, and the team still opens the key sources to confirm what they say.

The timeline shows events by month and filters by person. Set beside the price announcements, it shows whether competitor contact clusters before each increase. For the interviews, the team runs the fact extraction prompt on each key custodian, with their aliases, to build a dated list of the actions the documents attribute to that person.

Two decisions shape how the team treats its own material. In Akzo Nobel, the Court of Justice confirmed in 2010 that in EU competition investigations, internal communications with in-house lawyers are not covered by legal professional privilege. In 2018 the Bundesverfassungsgericht rejected complaints against the search of a law firm's Munich office and the seizure of documents from an internal investigation it had run for a client. The in-house legal team appears throughout this collection, and the investigation's own notes may not stay protected.

The privilege review workflow gives a first pass once its prompt names EU and German law in place of Canadian law and the Investigation Brief names the in-house and outside lawyers. Its calls show the team where to look. Each privilege decision, the handling of interview notes and the choice of what to put in writing stay with counsel.

The largest decision belongs to the partners and the board. On the strength of what the review has found, they decide whether to place a marker for leniency, and with which authority.

The board receives a short report with the chronology attached. Insights download as PDF, Word or Excel, each with a list of the source documents and their Document IDs, so any row can be traced back into Discover. The Bulk Task History Report gives a PDF record of each task run on the case, with its full prompt, the model and the comments left on it, for anyone who later asks how the review was done.

An authority needs more. If the board decides to apply, the documents relied on go to qualified translators, since a submission calls for human translation and the AI summaries served triage. The reporter is owed feedback within three months of the acknowledgement, and the team agrees with the board how much it can say without affecting the investigation or the rights of the people named in the report.