Germany's Bundesverband Digitale Wirtschaft (BVDW) on November 5, 2026 published its first "KI-Knigge," an etiquette guide for artificial intelligence at work that tells employees when to skip the chatbot altogether, how to behave when a meeting is being summarised by software, and who should own an AI agent before it is allowed to act.

The 25-page document, titled in full "Der BVDW-KI-Knigge: Verhaltenskodex für den nachhaltigen, sozialen und verantwortungsvollen Umgang mit KI in der Arbeitswelt" - a code of conduct for the sustainable, social and responsible use of AI in the working world - borrows its structure from the environmental, social and governance (ESG) framework companies already use for sustainability reporting. It borrows its name from Adolph Freiherr Knigge, whose 1788 book "Über den Umgang mit Menschen" made his surname the German shorthand for a manual of good manners. Quotations from the guide and from BVDW's announcement are translated from German by PPC Land.

In Short

Germany's digital economy association BVDW published a 25-page AI etiquette guide on November 5, 2026 that sorts everyday AI habits at work into environmental, social and governance chapters. It reaches anyone in a company who drafts social posts, summarises meetings, generates images or hands tasks to AI agents, and several of its points mirror labelling duties that the EU AI Act has applied since August 2, 2026. Nothing in it is binding, so what changes depends on whether employers turn its pointers into house rules on approved tools, human sign-off and documentation.

A code of conduct that refuses to be a rulebook

The guide was written inside BVDW's Lab Responsible AI for Business, a unit of the association's Working Group Digital Responsibility. Its lead author is Maike Scholz of Deutsche Telekom, who heads the lab. BVDW, which represents more than 600 member companies, presents the document as "practical recommendations" for companies that want to integrate AI deliberately along the three sustainability dimensions.

The authors are explicit about what the document is not. They write that they are convinced "that enabling users should come before bans," and that the guide is meant "not as a set of rules, but as guidance from practice for practice." The tone follows from that choice. Readers are addressed throughout with the informal German "Du," and each chapter ends with a list of what the authors call "Orientierungspunkte" - orientation points - rather than obligations.

Seventeen numbered sections carry the substance: four under Environmental, five under Social and eight under Governance. By PPC Land's count, those sections contain 94 orientation points - 23 environmental, 26 social and 45 governance. The governance chapter therefore holds almost half of the guide's practical content, and its final section on agentic AI systems, with ten points, is the longest single list in the document.

Scholz framed the guide around a single idea in BVDW's announcement. "AI can take work off our hands, but not the responsibility for its use," she said. "The AI Knigge shows how responsible AI use in everyday working life can succeed. To achieve that, we have to think about resources, human interaction and clear responsibilities together."

BVDW describes the guide as a snapshot. The announcement says it will be adapted "as needed" to new findings, requirements and practical experience, and the guide's closing outlook repeats that technological, regulatory, commercial and ethical conditions around AI keep shifting. One more detail sits in the imprint on the last page: "This paper was produced using artificial intelligence." The document does not say which tools were used, or for which parts.

Environmental: the cost of asking

The environmental chapter opens with a plain premise. "AI systems require considerable computing power and therefore energy," the authors write, and every prompt, generation and analysis consumes resources that "should be proportionate to the value created." They then reframe the usual question. "The question is not 'Do I need AI for this?' but 'Can AI optimise this process AND does the output justify the use of resources?'"

The chapter draws a line between employer and employee before it starts. "Selecting AI tools and checking their environmental impact is primarily the responsibility of companies," the guide says. What follows concentrates on what individuals can add in daily use.

Necessity before technique

Section E1 argues that the decisive choice comes before any prompt is typed. "Sustainability does not come from the best prompting technique, but from the right fundamental decision before the first prompt is written," the guide states. It warns against "over-automation," noting that a classic rule, a macro or a simple script is often enough compared with AI automation. It asks for a process check across the whole workflow, not just a single prompt, and for existing templates and documents to be consulted first.

Data gets the same treatment. Under what the authors call the principle of minimal necessary data, only the data a use case really needs should be generated, fed in and retained. "The most sustainable amount of data is the amount that is never generated in the first place," the guide says. Content is handled in similar terms: when AI speeds up production, the goal is better content, not more of it. "More output is not automatically more value."

One number, no source

Section E2 contains the only quantified claim in the document. "A search query consumes only about a tenth of the resources of a generative AI query," the guide states, as part of a recommendation to check briefly whether a classic search engine would do before turning to generative AI. No source is given, and the guide does not define whether "resources" means electricity, water, hardware or money.

The same section concedes that the measurement problem is unsolved. Uniform standards for measuring and reporting the energy consumption of AI workloads, and for converting it into a carbon footprint or costs, "have yet to be created," the authors write. That leaves the one-tenth ratio standing on its own. A similar multiple has circulated before: Alphabet chairman John Hennessy told Reuters in February 2023 that an exchange with a large language model likely cost around 10 times a standard keyword search - an estimate of cost, not of energy. Providers' own disclosures use different measures again. Google said in August 2025 that the median text prompt in its Gemini app used 0.24 watt-hours once idle machines, host processors and cooling were counted, and 0.10 watt-hours for the active chips alone, figures that sit underneath the inference cost of running AI at scale. At the aggregate level, an investigation by Paligo projected AI data centers consuming 945 terawatt-hours a year by 2030.

The remaining E2 pointers are operational, and several of them double as cost controls. The guide recommends storing recurring instructions - tone, target audience, format requirements, background knowledge - in the assistant's settings so they are not resent with every request. "That saves tokens, computing power and time." It asks for a new chat for a new topic, because long conversation histories increase the computing effort of each new request as the entire previous context is processed again.

Model choice gets its own pointers. The authors favour energy-efficient and, where possible, open models, and suggest checking whether providers use technical optimisations such as model compression or quantisation, which they say cut energy per request "considerably" without noticeably affecting quality. Work should start with the smallest or fastest model available, with "mini" or "flash" variants named as examples, moving to a more powerful model only if quality is insufficient - "not the other way round."

Two pointers target the way prompts are written. "Paper prompting" instead of "prompt ping-pong" means sketching goal, context, audience, format and length before opening the chat, like a mini briefing. For recurring tasks such as minutes, offer texts or social media posts, the guide proposes a fixed scheme - role, context, task, format, exclusion criteria - rather than reinventing the prompt each time. A final pointer asks for available energy data to be used when selecting tools, citing an "Energy Score" from platform providers without naming the platform, and the growing disclosure obligations on providers under the EU AI Act.

The rebound effect

Section E3 names a problem economists will recognise. Because AI makes tasks faster and easier, the guide argues, there is a temptation to create more content, iterate more often or take on additional tasks. This rebound effect can push resource consumption up without matching value.

The countermeasures are specific. Time saved by AI should go into a more thorough review of existing work rather than unrequested extras. Iterations should be capped with a fixed upper limit of prompt runs per task, "for example a maximum of three attempts," because endless optimisation raises consumption without proportional benefit. The guide also flags what it calls unnecessary AI-to-AI communication: when texts are repeatedly created, reformulated or summarised by AI on both sides of an exchange, resource use often rises without corresponding value. The description fits a pattern familiar to any team that drafts emails with one assistant while recipients summarise them with another.

Section E4 closes the chapter with knowledge sharing. A shared prompt library in a wiki or common folder, plus short documentation of successful use cases - task, model used, prompt structure, result - is meant to stop each person running their own trial-and-error cycles. The chapter's final line reads: "Just because something is possible does not mean that using it in the company makes sense."

Social: what the algorithm cannot see

The social chapter starts from the claim that collaboration in the AI era succeeds "not despite, but precisely through our human strengths." Section S1 is blunt: "Respect, trust and genuine collaboration arise between people, not between algorithms." AI, the authors write, cannot judge what a colleague is going through, what experience lies behind a proposal, or what dynamic a team currently has.

Three pointers in S1 deal with situations that have become routine in offices. Before an AI-supported summary or transcription of a meeting, participants should be asked whether they agree, to protect transparency, mutual trust and "the legitimate interests of all those involved." AI should not be used as an instance for assessing colleagues' work, which the guide says devalues their expertise. And because AI writing style "can often be recognised," the guide asks for AI-generated text to be reworked in the sender's own style before it goes out.

Section S2 treats transparency as the basis of trust with colleagues and customers alike. Every AI result should be checked before it is passed on, because it will be attributed to the person who sends it as their own work. AI-generated images, audio and video should be labelled when in doubt, even for purely internal use, "even if this goes beyond the legal obligations to label so-called deepfakes." A footnote points to Article 50 and following of the EU AI Act. The same footnote supports a second pointer: AI-supported touchpoints such as chatbots or product advisers should be designed so users can clearly recognise them as AI.

Section S3 sets out the epistemic position in one line. "AI delivers probabilities, not established facts." The guide warns that AI can produce invented information - it uses the German verb for hallucination - and asks for outputs to be checked against professional knowledge, with sources questioned. It flags bias with respect to gender, age, origin and other diversity characteristics, noting that outputs reflect historical inequalities in training data. Users are asked to keep their own skills sharp, because "only those who understand a topic can evaluate AI results soundly," and not to delegate decisions to the system.

Section S4 moves from the individual to the team, asking for shared ground rules on when and how AI is used, for AI-generated text to reflect the company's tone and values, and for generic, impersonal communication to be avoided.

Gender in the prompt

Section S5 is the most specific to its language. AI systems trained on historical data, the guide says, still tend to depict specialists and managers as male by default and to show women in stereotyped roles. Its first countermeasure is grammatical: when gender is not relevant, prompts should consistently use the feminine or a gender-neutral form, because, in the guide's words, AI follows the language of the prompt. The German example asks for "Expertinnen" - the feminine plural of experts - instead of "Experten," the masculine form traditionally used as a generic.

For images, the guide recommends specifying diversity of age, gender and background explicitly, such as "a group of executives of different genders and backgrounds," because without that instruction AI reproduces the default of its training data. Salary or performance analyses that AI builds from historical data should be questioned, since they can perpetuate structural disadvantages. The section ends with an observation about the workplace itself: "Access to AI know-how is becoming a career factor."

Governance: the longest chapter

The governance chapter runs to eight sections and 45 pointers. Its premise is that AI can support work, but responsibility "for content, decisions and their consequences always remains with people."

Degrees of human control

Section G1 introduces three levels of oversight. Depending on risk and use case, a person can approve decisions (Human in the Loop), supervise the process (Human on the Loop) or only intervene afterwards (Human out of the Loop). The degree of control, the guide says, should be chosen deliberately and be proportionate to the risk. It adds that establishing human-in-the-loop processes is in some cases legally required, or simply easier, as a way to minimise risk.

The same section places the labelling burden on users. "Users bear primary responsibility for the correct disclosure of AI content, even if platforms offer supporting tools," it states, citing Article 50(4) of the AI Act. Section G2 then lists quality checks: facts, figures and statements verified before reuse; sources checked to see whether they exist, are trustworthy and actually support the claims made; results tested for plausibility.

Data, approved tools and the shadow problem

Section G3 deals with data protection and confidentiality. Confidential or internal information may only be entered into approved AI solutions. A footnote explains the reasoning: "Enterprise solutions approved by the company are usually contractually designed so that company data is protected and, for example, not used to train the AI models." Personal data should not be entered unless this is expressly permitted and secured under data protection law; trade secrets stay out of unapproved systems; internal data classifications decide which information may be processed where. Third-party copyrighted content should only be uploaded if licence terms or usage rights allow it.

Two G3 pointers address different roles. Providers of generative AI systems are asked to ensure the technical traceability of generated content by building machine-readable labels and metadata into the system, "for example via C2PA or SynthID as a 'digital passport.'" Users are asked to label all AI-created or manipulated content that could be perceived as a deepfake, "in order to fulfil the legal transparency obligation."

Section G4 turns to shadow AI, defined as the use of unapproved AI tools where clear rules or suitable company offerings are missing. The pointers are the strictest in the guide: only company-approved AI tools for business tasks; no internal, confidential or sensitive information in publicly available SaaS AI services; no private AI accounts for business content or documents. The guide warns that using unapproved AI solutions "can have legal, employment-law or personal consequences," and asks employees to point colleagues towards existing policies and approved tools.

That section lands on a measurable problem. Browser telemetry from LayerX found that 47.11% of enterprise AI conversations ran through personal identities rather than corporate accounts, a figure Akamai reused in an enterprise AI usage risk report promoted in October 2026.

Section G5 links everyday documentation to regulation. Workflows, use cases and procedures should be documented now, the guide argues, because documentation is "an important basis for regulatory requirements, such as an AI system inventory and the risk assessment under the EU AI Act." Important prompts and outputs should be saved so they can be traced, reused and improved, and the use of AI in customer-facing work should be communicable transparently.

Section G6 covers copyright. Beyond checking usage rights before reusing AI-generated or third-party content, the guide asks users of commercial AI solutions to find out whether the provider offers contractual protection or liability arrangements for possible copyright infringements. The exposure is not abstract: a survey by CCC and Outsell found employees using externally published content with AI about 11 times a week, with each AI output reaching 96 people on average.

Section G7 opens with three words: "Every AI is IT." Existing security and IT policies, from the code of conduct to data security rules, apply to AI as well. The guide flags phishing and manipulated content, with a German reference point: "With AI, voices, images and messages can be generated to look deceptively real - this gives something like the grandparent scam a completely new dimension." The "Enkeltrick," in which fraudsters impersonate a grandchild in distress, is a well-known scam in Germany. The section also names prompt injection and manipulated inputs in external documents as AI-specific attack methods.

Agents with an owner and a kill switch

Section G8, on agentic AI systems, carries the most detailed instructions in the document. Because agents can carry out tasks independently and interact with other systems, the guide argues, risk and responsibility rise with autonomy. Its first pointer reads: "No AI algorithm without a clear owner." Every AI agent should have a named person or organisational unit that steers its use and takes responsibility.

The remaining nine pointers read like an operating checklist. Agents should receive precise tasks, rules and termination criteria. Human approval is required at defined points: "Critical decisions or external communication must not happen autonomously." Access rights should be limited strictly to the data and systems that are absolutely necessary. Behaviour and results should be monitored continuously, with intervention when deviations appear. Self-learning or self-expanding processes should not run without supervision. The tasks an agent takes on, and how its decisions come about, should be documented. Testing should happen in safe environments such as a sandbox before agents are integrated into real processes. Agents must not process or pass on sensitive data unless expressly permitted. And processes must be stoppable or correctable at any time, through automatic termination conditions or a kill switch.

The chapter closes with a sentence that sums up the guide's approach to regulation: "Security, data protection and compliance are not restrictions but the basis for trustworthy use."

Where the guide meets the AI Act

The Knigge cites the AI Act in three footnotes - twice to Article 50 and following, once to Article 50(4) - and mentions it twice more in the text, on providers' energy disclosures and on the AI system inventory. The Article 50 transparency duties became legally binding on August 2, 2026. Article 50(2) requires providers to mark AI outputs in a machine-readable, detectable form, and Article 50(4) requires deployers to disclose deep fakes and AI-generated text published on matters of public interest, unless there has been human review or editorial control. Fines reach up to 15 million euros or 3% of worldwide annual turnover, whichever is higher. The Commission had published free AI labelling icons before the deadline.

BVDW is not the first trade body in the region to translate those rules for its members, but its approach differs. IAB Austria's memo, written with a law firm, concluded that the producing agency is normally the responsible deployer and that a client becomes one only if it influences the AI use. VIA Nederland mapped four scenarios that trigger AI disclosure for ad agencies, from chatbots to in-house generative tools. Both documents allocate legal duties between companies. The BVDW guide speaks to the individual employee, and in places it asks for more than the law does, such as labelling AI-generated media used only internally.

The association has engaged with AI regulation directly before. In October 2025, BVDW warned about fragmented authority structures in Germany's draft law implementing the AI Act. The Knigge stays out of that debate; its one regulatory forecast is that AI use will become more standardised in coming years, which is why it wants documentation to start now.

Why marketing teams are the obvious readers

The guide is written for any office, but its examples come from marketing workflows: social media posts, offer texts and minutes as candidates for fixed prompt structures; chatbots and product advisers as touchpoints that must identify themselves; image generation as a source of stereotypes; AI-written communication whose style recipients can spot.

Survey data suggests the gap the guide addresses is real in the industry. IAB Europe's second Impact of AI on Digital Advertising report found that 86% of respondents use AI for marketing but only 48% have AI guidelines specifically for advertising and marketing, while 78% have identified a formal owner for AI governance. A smaller survey of 112 marketing professionals by ProGEO.ai found 75.9% using generative AI daily, with 76.8% working under a formal usage policy but only 43.8% seeing it enforced through technical controls.

The agent section connects to BVDW's earlier work. In January 2026, the association published a 25-page whitepaper on deploying agentic AI responsibly, which cited a Civey poll in which 71% of respondents could not envision AI agents handling tasks without human intervention. BVDW has also published practitioner documents on programmatic curation and an email marketing guide that anticipates AI agents reading the inbox. The Knigge differs from those in that it addresses behaviour rather than a channel or a technology.

Who wrote it, and what it leaves open

Besides Scholz, seven co-authors are listed: Anna-Lena Mikoteit-Zerb, formerly of Kleinanzeigen; Beatriz Bilfinger and Katharina Schneeberg of BVDW; Jakob Wößner of Weleda; Kornelia Haslbeck of BCN Group; Marius Menge of eagle lsp; and Tobias Kellner of Google Deutschland. Bilfinger, a senior programme manager for sustainability and digital responsibility, is BVDW's contact for the publication.

The guide names no vendor directly, though two of its examples correspond to product names. SynthID, cited next to the C2PA standard as a "digital passport" for AI content, is a watermarking system built by Google DeepMind. "Flash" is the suffix Google uses for its lighter Gemini models; it appears alongside "mini," a suffix used by OpenAI among others. The "Energy Score" the guide recommends consulting is attributed only to "platform providers."

Several questions remain with the reader's employer. The guide does not define which decisions count as "critical" for agents, how approval of AI tools should work, or who measures whether the rebound effect is being contained. Its one figure on resource use carries no source. And the document places responsibility on two levels without fully reconciling them: the environmental chapter says tool selection is "primarily the responsibility of companies," while the closing outlook states that "responsibility for a conscious, respectful and compliant use of AI, however, lies with each individual."

That tension is built into the format. A Knigge describes how people should behave; it does not enforce anything. Whether the 94 pointers change practice will depend on how many companies adopt them as policy - and on whether the next edition, which BVDW has already promised, arrives with the measurement standards the current one says do not yet exist.

Timeline

Summary

Who: The Bundesverband Digitale Wirtschaft (BVDW), Germany's digital economy association, through its Lab Responsible AI for Business. Lead author Maike Scholz of Deutsche Telekom, with co-authors from BVDW, Google Deutschland, Weleda, BCN Group, eagle lsp and formerly Kleinanzeigen.

What: "Der BVDW-KI-Knigge," a 25-page voluntary code of conduct for AI use at work, organised along environmental, social and governance lines in 17 sections. It covers resource use (including a suggested cap on prompt attempts and starting with the smallest model), workplace conduct (consent before AI meeting summaries, labelling, gender bias in prompts) and governance (approved tools only, documentation, copyright, security and named owners plus kill switches for AI agents).

When: Published November 5, 2026, with BVDW describing it as a snapshot to be updated as needed.

Where: Berlin, Germany, for companies and employees in the German digital economy; several pointers refer to Article 50 of the EU AI Act, whose transparency duties have applied across the EU since August 2, 2026.

Why: BVDW argues that whether AI improves work depends on how it is used, not only on efficiency, and that enabling users should come before bans. For marketing teams, which use AI heavily for content, images, chatbots and increasingly agents, the guide sets out behaviour that sits alongside, and in places goes beyond, the legal labelling duties already in force.