AI Disclosure

How Druma uses AI, what data the AI sees, and what controls you have over it.

Effective 13 November 2026 · Version 2026-11-13

This document is available in English only. The English version is the legally binding version.

Overview

Druma uses AI to read documents and emails into records. A person confirms each record; a CMR reading goes to a person only when it is uncertain or does not match the order. Druma also uses AI to translate messages, transcribe driver voice notes, answer questions and write short comments on figures. Where a company has an AI assistant, the assistant chases the documents an order still needs after delivery. Druma does not use AI to make legal decisions or regulatory filings, or to take a decision that has a legal or similarly significant effect on a person.

Where AI is used

  • Smart Import reads trucks, drivers, clients and insurance from your existing files. You check the result in a review table and confirm before anything is saved.
  • Inbound order email: AI reads the lane, dates and goods of a transport request emailed to your Druma address into a draft order, which stays a draft until a planner accepts it. Scanned transport orders and tender files are read the same way. For a sender that writes often, AI also writes a reading template from that sender's earlier documents, and a planner's corrections make it write a better one.
  • Inbound carrier and supplier invoices: AI reads the amount, VAT, dates and references. The result waits for a person's approval before it reaches the books.
  • CMR reading: where Druma has switched it on for a company, Mistral's OCR model reads a photographed or scanned paper CMR, and Druma's own code compares the reading with the order.
  • Message translation lets a driver and a planner who use different languages each read the other in their own language.
  • Driver voice notes are transcribed into text.
  • Ask Druma answers questions about how to use Druma from our documentation. If a company administrator switches on its data assistant, it also answers questions about your own orders, fleet, finance, driver hours and order notes, under the asking user's own permissions. Order notes and other free text can contain names, and these go to the AI provider to compose the answer.
  • The setup guide has an assistant that explains a step and proposes values from what you tell it or from a document you upload. You confirm each value before it is saved.
  • AI Reports turns a question typed in plain language into a report and writes a few comments on the result. Some dashboards show a short AI-written comment on their figures.
  • The client portal has an assistant that answers a client's questions about its shipments.
  • Druma AI and the AI assistants have their own sections below.

Models and providers

Druma's own AI features run on models from Google, Anthropic and Mistral AI. As of 13 November 2026 Druma is moving off Google Gemini: order and document reading and reading templates go to Mistral, the rest to Anthropic. This section names each feature's provider on that date and will be updated as each one moves.

Google Gemini on Vertex AI reads documents for Smart Import, inbound orders, invoices, tenders and the setup guide, writes the reading templates, and runs the Ask Druma data assistant, the client-portal assistant, voice-note transcription and the comments on dashboard figures. Druma sends these requests to Vertex AI's EU multi-region endpoint.

Google's terms commit that data sent to that endpoint is stored and processed in EU member states, and that Google does not train or fine-tune models on customer data without Druma's permission. Google affiliates, including Google LLC in the United States, may in limited cases access data to run the service; Google relies on the EU-US Data Privacy Framework and Standard Contractual Clauses for that. Prompts that Google's safety classifiers flag may be logged for up to 90 days in the same region to investigate abuse.

Google Cloud Translation translates messages. Google states the text is used only to translate it, and it may be processed outside the EU.

Mistral AI (Mistral AI SAS, France) runs the CMR reading described above. When Druma moves order reading to Mistral, it also extracts the order data from transport orders received by e-mail, the Outlook add-in, the document scan and the client API, and writes the reading rules for each sender; until then Google Gemini does that step. Preload notices, bulk onboarding imports and the check of a loading confirmation against its order stay with Google Gemini for now. Mistral states that data sent through its API on Druma's paid plan is not used to train its models. Its privacy policy says it keeps API inputs and outputs as long as needed to produce the output and then for 30 rolling days to monitor abuse, unless zero data retention is activated.

Anthropic Claude runs, through Anthropic's API in the United States:

  • The Ask Druma help chat and the setup-guide chat, on Claude Haiku 5.5, with Gemini as the fallback when Claude is unavailable. They send your question, the earlier turns and the help article; the setup-guide chat also sends where your company stands in its setup.
  • AI Reports, on Claude Haiku 5.5 by default; Druma may switch it to Claude Sonnet 5.5. It sends the question and an extract of the result you may already see: at most 20 rows per chart or table, with names of clients, carriers, suppliers, countries or trucks and their figures. Driver names are replaced by neutral labels first.
  • The AI assistants, on Claude Haiku 5.5, as described in AI assistants.

Under Anthropic's published policy, it deletes API inputs and outputs within 30 days, unless the law requires longer; requests flagged under its usage policy may be kept up to 2 years. Anthropic's commercial terms do not allow it to train its models on them. The transfer safeguard for each provider is in our sub-processor list.

Druma's own features use no OpenAI or Meta model. Without Druma AI, nothing goes to OpenAI or the public Gemini API, and Anthropic receives only what the features above send. With Druma AI, the chosen provider receives what that section describes.

Druma AI (bring your own AI)

Druma AI is an optional chat in which a user asks questions about the company's operational data. It can also propose a change, which runs only after the user confirms it. It is off by default. A company administrator switches it on and either connects the company's own API key for OpenAI, Anthropic or Google Gemini, or uses a key Druma provides. Before saving its own key, the administrator records an acknowledgement of the data flow described here.

On the company's own key, the prompts and the records Druma AI reads for the user, within that user's permissions, go directly to the chosen provider: api.openai.com, api.anthropic.com or Google's public Gemini API. That provider's terms and its data-processing agreement with your company apply, and none of these APIs is set to an EU endpoint. A free-tier Gemini key lets Google use what is sent to improve its products. Google's terms allow only paid Gemini API keys for an application used in the EEA, Switzerland or the UK, so connect a paid key.

On Druma's key, a conversation goes to that key's provider under Druma's account, within a monthly spending cap Druma sets for the company. The company chooses which key a conversation runs on.

A saved key is encrypted at rest in a managed vault, used only on Druma's servers, and never shown again.

AI assistants (Druma Agents)

An AI assistant takes over part of an office task for a company. Druma sets up an assistant when a company asks for one, and a company administrator switches it and its skills on. The first skill is the Document chaser. After an order is delivered, it asks the driver in the driver app, or the subcontracted carrier by email, for the documents the order still needs before it can be invoiced, such as the CMR of a cross-border load or a proof of delivery. It records the answer as the reason a document is missing.

An assistant sends messages only to the company's drivers and subcontracted carriers, never to its clients. Every message says it comes from an AI assistant, and a name the company gives its assistant does not replace that label. A carrier who replies to the email reaches the company's planners; the assistant does not read the reply.

An assistant reads only what its own restricted role allows. For the Document chaser that is the order number, the route countries, the delivery time, which documents are on file and who drives or carries the load. Its role has no access to driver performance or eco-driving scores. A tapped answer such as The client kept it is recorded without AI. A driver's free-text answer goes to the model with today's date and the delivery date, and comes back as one reason from a fixed list, a date if the driver named one, and a translation for the planner. No order number or name is added, but the driver's text may itself contain personal data. What a driver tells an assistant is never used to evaluate or score the driver; the driver scorecard does not read it.

Each step runs in shadow, draft or automatic mode, up to the highest mode Druma allows for it; in automatic mode a step runs without a person approving each action. The company sees every action, can pause an assistant or take over an item, and can undo an action where Druma offers it. The modes, the records Druma keeps and how long are set out in the Druma Agents Terms.

Conversation history

Ask Druma has two tabs. The help tab is not saved: the conversation lives only in your browser tab and is gone when you close the panel or reload the page. Druma writes nothing to its database. When the help chat runs on Claude or its Gemini fallback, the provider's retention in Models and providers applies. If you click "Talk to a human", the help conversation up to that point is attached to the support conversation. The button opens Live Support, where a person at Druma answers, and that conversation is deleted 90 days after its last activity.

The assistant tab is saved for you. Conversations with the data assistant and with Druma AI are kept so you can reopen them and ask a follow-up. Only the user who created a saved conversation can read it; colleagues and company administrators cannot. A company that needs oversight has the audit log, which records that a user ran a data query and when, but not the text of the question.

A saved conversation is deleted 90 days after its last message, and you can delete one yourself at any time. Saved conversations are part of a personal-data export and are removed when your user account or the company account is deleted. A saved answer shows records as they were when you asked, and grants no new access: the assistant queries your data under your own permissions every time.

Training and data use

Druma does not train AI models on your data, and sends data to a model provider only to answer the request it was sent for. What each provider may keep, and its training terms, are under Models and providers.

The CMR reader learns one thing from your documents. On your paper CMRs it measures where the printed captions and boxes sit on each form type, so that it finds them on the next form of that type, for every customer. That geometry is the only part shared between customers: positions relative to the printed frame, identifiers of printed captions from a closed list, and counts, medians and spreads over several documents. It holds no filled-in text, no names, plates or other values, and no company identifier. No model is trained on it. Your documents, the text read from them and your corrections stay in your account for at most 12 months, or 30 days if nobody labelled or corrected the result. The terms are in section 4.1 of the Data Processing Agreement.

The driver scorecard uses no AI. It is a fixed, published weighted average of five components: on-time performance (35%), fuel efficiency against a benchmark (20%), idling (15%), incidents (15%) and compliance (15%). Nothing is sent to an AI provider to produce it. A component with no data is left out and the other weights are rescaled. Small samples are flagged as low confidence. The score is still profiling in the GDPR sense and is described as such in our Privacy Notice.

Human oversight

A person reviews and confirms AI output before:

  • an imported record or a setup value is saved;
  • a draft order becomes an active order;
  • an extracted invoice is matched and posted.

The CMR reading goes to a person when it is uncertain or does not match the order, and when Mistral cannot read it within four hours. Translations reach the reader without review, with the original shown next to them. Ask Druma never changes a record; asked to take an action, it gives you the steps. Druma AI changes a record only after the user confirms the change.

AI assistants are the exception. A step the company has switched to automatic runs without a person approving each action. Oversight then lies in the company's choice of mode for each step, the list of every action, and the undo, pause and take-over controls.

Opt-out

  • Each user can switch message translation off in their profile settings.
  • The Ask Druma data assistant and Druma AI are off until a company administrator switches them on.
  • AI reading of inbound order and invoice emails is off until the company switches it on per feature in Settings, Integrations.
  • CMR reading runs only after Druma switches it on for a company.
  • AI assistants are off until Druma sets one up at the company's request. Each skill can be switched off, and the company can pause its assistants at any time.
  • Smart Import, the setup-guide assistant, the Ask Druma help chat and AI Reports run only when a user asks them to.
  • The other AI features cannot be switched off per company: voice-note transcription, the comments on dashboard figures and the client-portal assistant.

EU AI Act

The EU AI Act (Regulation (EU) 2024/1689) entered into force on 1 August 2024 and applies in stages:

  • From 2 February 2025: the prohibited practices (Chapter II) and the AI-literacy duty (Art. 4).
  • From 2 August 2025: the duties of providers of general-purpose AI models, and the governance, notified-body and penalty rules.
  • From 2 August 2026: most of the Regulation, including the Article 50 transparency duties. For systems already on the market, the duty to mark generated content in a machine-readable way applies from 2 December 2026 (Article 111(4), added by Regulation (EU) 2026/1744).
  • From 2 August 2027: the end of the transition for general-purpose AI models placed on the market before 2 August 2025.

Regulation (EU) 2026/1744, the Digital Omnibus on AI, in force since 27 July 2026, moved the rules for the high-risk AI systems listed in Annex III to 2 December 2027, and those for high-risk AI in products covered by the Union harmonisation legislation in Annex I to 2 August 2028.

Our classification. The AI Act sets roles per AI system. Druma builds AI systems on general-purpose models from Google, Anthropic and Mistral AI: the document readers, Ask Druma, the setup-guide assistant, the client-portal assistant, Druma AI and the AI assistants. Druma is the provider of these systems, and a company that uses them in its business is a deployer. In Druma's assessment none of them is high-risk under Annex III or a prohibited practice. Druma does not use AI to score creditworthiness or decide access to essential services. Druma's AI does not evaluate workers or decide on its own who gets which work. When a user asks, Druma AI can propose putting an order on a truck; a planner confirms it, and the proposal names a truck, not a driver, and does not rank drivers by behaviour or personal traits. No AI assistant proposes or makes a driver assignment. The planning proposals, including those some screens call AI matching, come from fixed scoring and optimisation rules that use no machine-learning model, so they are not an AI system under Article 3(1). The driver scorecard uses no AI (see Training and data use).

Article 50(1). A person who interacts directly with an AI system must be told so, unless it is obvious. This applies to Ask Druma, the setup-guide assistant, the client-portal assistant, Druma AI and the AI assistants. In these panels you are talking to an AI system. The live-support hand-off inside Ask Druma is labelled Live Support and connects you to a person at Druma. The panels carry an assistant icon, and each Druma AI answer carries a badge such as "Read-only", "Using Druma's key" or "Using your provider key". A driver or carrier who receives a message from an AI assistant is told in that message that it comes from an AI assistant.

Article 50(2). Druma generates no synthetic images, audio or video. Its AI does generate text: answers in Ask Druma, the setup-guide assistant, the client-portal assistant and Druma AI, AI Reports and the comments on dashboard figures, and an assistant's translation of a driver's answer for the planner. This text is labelled as AI on screen and marked in machine-readable form. The data and the on-screen elements carry the standard IPTC source-type marker trainedAlgorithmicMedia, an email written by AI carries a header saying so, an exported spreadsheet carries it in its document properties, and text copied as plain text keeps a visible label.

Article 50(2) does not apply where a system performs an assistive function for standard editing or does not substantially alter the input data provided by the deployer or the semantics thereof. Druma treats reading a document into fields (including a CMR reading accepted without review), transcribing a voice note, and translating a message shown next to its original, as falling under that exception.