Eight rules, fifteen examples each. Every card opens to show what the right move would have been. Five at a time, so you can stop when you have seen enough.
An AI tool that receives work data is legally a processor. That requires a written agreement binding the vendor to your rules. No agreement: unlawful processing, full stop.
The sales team feeds a CRM export (names, emails, order values) through a free AI summariser with no processor agreement. The vendor becomes a de facto processor with no contract.
Only tools with a signed data processing agreement (Art 28(3)) are on the approved list. The export runs in a tool from that list, or not at all.
The CFO uploads supplier invoices with registration numbers to an AI bookkeeping app whose terms let it "improve the service" using customer data.
The processing agreement prohibits training and any other own use of your data. No such clause, no such tool.
The recruiter pastes candidates' CVs into a web-based AI parser without even checking whether the vendor is contracted as a processor.
Candidate data lives only in the recruitment system, which has a processing agreement. No external parser without a contract and a row in the register.
Marketing runs the entire customer list through an AI service to generate personalised emails. The tool stores the list on its own servers without a contract.
Personalisation happens inside the contracted marketing system. The customer list never leaves a system without a processing agreement.
The support team switches on an AI chatbot plugin that reads the whole ticket history (names, contact details) without IT approval or a contract.
Plugins with data access need IT approval and a processing agreement before they are switched on. Policy blocks the rest technically.
Reception uses a free AI note-taker during customer calls. The vendor's terms say nothing about processing on the company's behalf.
A note-taker that hears customer calls is approved only once a processing agreement is in place. Until then: pen and paper.
The company approves a cheap "team plan" of an AI chat. The plan runs on consumer terms with no processing agreement at all.
Buy the business or enterprise tier that ships with a processing agreement. If there is no such tier, the tool is not an option.
A subcontractor is allowed to run the company's customer data through its own AI tools. The contract has no sub-processor clause.
The processing agreement requires written approval for every sub-processor, AI vendors included. The subcontractor lists its tools.
Legal pastes a draft contract with the counterparty's names and personal data into an AI reviewer with no processing agreement.
Contract review happens in a tool listed in the register with its agreement date. Otherwise the text is anonymised first.
A browser extension for "AI email writing" reads the whole inbox. Nobody knows who the vendor is.
Extensions with data access are allow-listed by IT after a contract check. Everything else is blocked by browser policy.
The AI vendor changes its terms to allow training on customer data. Nobody at the company notices.
The processing agreement prohibits training on your data. Terms changes are reviewed annually and logged in the register.
HR uses an AI survey tool. The vendor hosts with a sub-processor that is never named.
The processing agreement lists every sub-processor by name and country, and you are notified before the list changes.
The CFO's "AI assistant" SaaS keeps personal data after the trial ends. There is no deletion clause.
The processing agreement requires deletion or return at contract end, with written confirmation.
A processing agreement exists but says nothing about the AI feature the vendor added later.
Every new AI feature triggers a contract review before it is switched on. The feature stays off until then.
Nobody can produce the processing agreement when a customer asks. It may be in someone's email.
Every processing agreement sits in the register, one row per vendor, retrievable in a minute.
Pocket Operator has three companies in the chain, all under a data processing agreement: Telegram carries the message, Hetzner stores it in Germany, Anthropic generates the reply. The whole list is published on thepocketoperator.com/data.
Data leaving the EU/EEA needs a safeguard: an adequacy decision, standard contractual clauses (SCC) or equivalent. "The tool was free and handy" is not a safeguard.
An employee transcribes a board meeting with a free AI tool whose audio files are stored on servers in China. No SCC exists.
Transcription uses a tool that processes in the EU, or one with SCC and a documented transfer impact assessment.
HR uses a US AI recruiting tool with no Data Privacy Framework certification or SCC to assess candidates.
The vendor is DPF-certified or has SCC in place, recorded in the register with the date.
A consultancy sends the client's business plan to a US-hosted AI model for proofreading. No transfer mechanism is documented.
Client material is proofread in an EU-hosted tool, or with a documented transfer safeguard the client has approved.
A developer uses an AI coding assistant that sends code, including test data with real customer details, to a server outside the EU without knowing it.
Test data is synthetic. The coding assistant is approved with a known processing region and a contract.
The project manager uploads a budget with staff data to an AI project tool running on a cloud server in India, with no adequacy decision.
Staff data stays in the HR system. The project tool gets a de-identified budget.
Marketing uses an AI image generator hosted in a country with no adequacy decision and uploads photos of employees.
Check the hosting country first. Without SCC or DPF, no personal data is uploaded, photos included.
The customer service AI routes tickets through a vendor whose fallback server sits outside the EU "during peak load".
The processing agreement fixes the processing region. Failover outside the EU needs the same safeguards, or no failover.
The AI vendor is EU-based, but its language model runs at a US sub-processor with no transfer mechanism.
The whole chain is mapped. Every transfer link has its own SCC or DPF, documented in the register.
An employee uses a personal VPN and a US-only tool for company work because "it is better".
Approved tools list. Personal workarounds are a policy breach met with training, not silence.
The analytics team exports EU customer data to a US colleague's AI tool for a "quick analysis".
Internal transfers are transfers too. The analysis runs in the EU-hosted tool the colleague is given access to.
A transfer impact assessment (TIA) has never been done for the non-EU AI vendor.
A TIA is on file for every non-EU AI vendor and updated when the legal situation changes.
The AI tool's DPF certification has lapsed. Nobody checked.
Certification status is verified at the annual review and logged.
The voice assistant vendor processes call audio in a data centre in Singapore.
EU processing region is written into the contract. If the vendor cannot promise it, no audio is sent.
The company relies on "consent" from employees to transfer HR data to a US AI tool.
Employee consent is rarely freely given. Use SCC and a TIA, not consent.
Backups of the AI tool's data replicate to a region outside the EU.
Backup locations are part of the transfer map and covered by the same safeguards.
Pocket Operator stores everything in Falkenstein, Germany. The only transfer is to Anthropic in the US, under an agreement with EU transfer safeguards. It is stated on the page before anything is created.
Health, union membership, ethnicity, sexual orientation: extra-protected categories. Feeding them into a general AI tool is almost always a serious breach, whatever the contract says.
HR feeds sick-leave certificates into an AI chat to "summarise absence patterns".
Certificates stay in the HR system with restricted access. Absence statistics are produced in aggregate, without diagnoses.
A manager asks an AI to analyse an employee survey that contains union membership data.
The survey is designed without union questions, or analysed only in aggregate by an authorised person.
A healthcare administrator pastes patient records into an AI transcription tool with no contract covering health data.
Health data is processed only in tools procured for care, with a contract, an Art 9(2) legal basis and a DPIA.
The recruiter lets an AI read cover letters mentioning disability or pregnancy to "assess suitability".
AI is not used for suitability assessment. Cover letters are read by a person who knows what must not be weighed.
Finance runs diversity statistics tied to ethnicity through an AI analytics service with no extra safeguards.
Diversity data is collected only voluntarily, anonymously and in aggregate, with a documented legal basis.
Occupational health uses an AI chat to draft return-to-work plans from medical certificates.
Special-category data stays in the health system. AI gets anonymised text at most, with an Art 9(2) basis and a DPIA.
An AI transcribes a union negotiation and records who said what.
Union meetings are not recorded without explicit agreement and a legal basis. Minutes are written by a person.
A recruiting AI infers "cultural fit" from names and photos, in practice an ethnicity proxy.
Inference of protected attributes is banned by policy. The tool is validated, assessed in a DPIA and reviewed by a human.
Wellness app data (steps, sleep) is fed to an AI to score "team energy".
Health data needs explicit consent and a clear purpose. Scoring teams is not one.
An AI chat is used to summarise a harassment complaint containing details of sexual orientation.
Sensitive investigations are handled in a controlled system with restricted access. No general-purpose AI.
Facial recognition on office cameras "for security" via an AI vendor. Biometrics with no basis.
Biometric identification needs an Art 9 exception and is limited by the AI Act. Use badge access.
AI sentiment analysis on employee chats to detect "mental health risk".
Health inference from communications is prohibited without explicit consent. Offer support channels instead.
Customer profiles with religious dietary preferences are uploaded to an AI marketing tool.
Religion is a special category. Minimise, get explicit consent and keep it out of general tools.
An AI-generated "diversity dashboard" is built from an HR export including ethnicity, with no legal basis.
Aggregate only, with a documented Art 9(2) basis and a DPIA before anything is built.
Sick-leave patterns are analysed by AI to predict who will be absent.
Predictive health profiling of employees is out. Absence is handled through policy and conversation.
Pocket Operator is a work assistant, not a health or HR record. Nothing in the product asks for special-category data, and the privacy notice tells you not to send anything sensitive.
Data must be deleted once the purpose is fulfilled. "The tool keeps everything forever, just in case" is itself a breach.
An employee has a free AI account tied to the work email where every conversation is kept forever with no deletion policy.
Work accounts in approved tools with retention set at onboarding. Conversations are deleted automatically after the period.
The company approves an AI CRM add-on that archives every customer call permanently "for future training".
Calls are kept only as long as the case requires. No training on customer data, written into the contract.
No routine exists for deleting AI-generated summaries of closed projects, even though the source data should have been purged long ago.
The purge routine covers derived documents too. When the original is deleted, the summaries go with it.
An AI meeting bot keeps every meeting note on a third-party server with no defined retention period.
Retention per data type is defined before the bot is switched on. Notes are deleted automatically when the period ends.
An employee exports customer data to an AI analytics SaaS with no automatic deletion when the service is cancelled.
The contract requires deletion on cancellation with written confirmation. Tools without that clause are not used.
The AI vendor keeps prompt history for three years by default. Nobody changes the setting.
Retention is set to the shortest workable period at onboarding and documented in the register.
Former employees' AI chat histories remain accessible after they have left.
The offboarding checklist deletes AI accounts and their data the same day.
The customer service AI keeps full chat logs with personal data forever "for quality".
Logs are anonymised once the case is resolved. Only what quality review needs is kept, time-boxed.
An export of customer data to an AI tool is still sitting on a laptop years later.
Local exports are prohibited or carry an expiry date. Storage locations are listed in the register.
Nobody can say how long the AI tool keeps uploaded files.
Retention per tool is documented in the register, sourced from the contract.
The AI tool's "memory" feature retains customer details across unrelated conversations.
Memory is scoped per purpose and cleared when the purpose ends.
Audit logs of AI usage contain full prompts with personal data and are kept for ten years.
Logs keep metadata, not content. Where content is needed it is pseudonymised.
Deleted customer records live on inside AI-generated summaries.
Deletion flows through to derived documents. Summaries are traceable to their sources.
The vendor's backups are kept for a year, longer than your policy, without you knowing.
Backup retention is part of the processing agreement and matches your policy.
An AI tool piloted two years ago still holds the company's data.
Pilots end with a deletion request and written confirmation, logged in the register.
Pocket Operator deletes operational logs automatically after 90 days. Your own notes and memory stay until you delete them. Two clocks, on purpose.
Decisions with legal or similarly significant effect on a person may not be made solely by an AI. Human review and a right to an explanation are required.
An AI tool automatically rejects job applications with no human reviewing the rejections.
AI may sort, but every rejection is reviewed by a person and the candidate can ask for an explanation.
Credit assessment of suppliers is done entirely by an AI model with no possibility of human review.
The model provides input. The decision is made by a person, and the supplier can request a review.
Employee performance evaluations are generated by AI and used directly for pay decisions without manager review.
The manager reviews and can override. The employee sees which inputs were used.
AI automatically decides which customers get a discount based on behavioural data, with no way for the customer to ask why.
Pricing with significant effect is reviewed by a human. The customer gets an explanation on request.
An employee is denied an internal promotion after an AI-generated match score, without even being told AI was behind it.
Promotions are decided by people. If AI input is used, it is disclosed and the employee can respond.
AI triages customer complaints and auto-rejects refunds below a certain score.
Refund denials are reviewed by a person. The customer can contest.
AI sets the shift schedule and gives "low performers" fewer hours, fully automatically.
Schedule suggestions from AI, decision by the manager, explanation available to the employee.
AI sets individual prices from behaviour, with no human check.
Human review and a right to an explanation. No solely automated decisions with significant effect.
AI flags employees as "attrition risk" and HR acts on it without talking to them.
Risk flags are conversation starters, not decisions. The employee is told and can respond.
An AI fraud detector suspends customer accounts automatically, with no appeal.
An appeal path to a human within a defined time. Suspensions are explained.
AI screens supplier tenders and excludes bidders automatically.
Every exclusion is assessed by a human who records why.
Credit for customers in the webshop is granted or denied by AI at checkout.
Art 22 exceptions only with safeguards: human review, explanation, right to contest.
AI performance ratings feed straight into bonus calculations.
The manager reviews and can override. The employee sees the inputs.
The AI chatbot decides which customers get to talk to a human.
Anyone can reach a human. The bot's routing is not a decision with legal effect.
Nobody knows which decisions in the company are automated.
An inventory of automated decisions. Each one is assessed against Art 22.
Pocket Operator drafts, suggests and reminds. It never makes a decision with legal effect on anyone. A human always acts.
People have the right to know when they are talking to, or being assessed by, an AI. Silence counts as a breach, not a grey area.
The website's customer service chatbot never says it is an AI and lets the visitor believe it is a person.
The first message says "I am an AI assistant". It is in the bot's instructions, not in a footnote document.
An internal HR tool generates staff evaluations without marking the content as AI-generated.
The evaluation is labelled "AI draft, reviewed by <name>". The employee knows AI contributed.
AI-generated summaries of customer calls are forwarded around the organisation without being marked as AI-created.
Summaries are labelled AI-generated and reviewed. The source is stated.
A voice-based AI assistant calls customers without introducing itself as an AI at the start of the call.
The voice bot introduces itself as an AI in the first sentence. The customer can ask for a human.
Marketing publishes AI-generated product images and copy without labels, despite the context requiring them.
Synthetic content is labelled where it could mislead, per Art 50(4). One labelling decision per campaign.
AI-written customer emails are signed with a person's name, with no mention of AI.
Disclose when it matters. Never give the impression a person wrote what an AI wrote.
An AI-generated video of the CEO is used in internal comms without a label.
AI-generated media is labelled per Art 50(4), internally too.
The AI voice in the phone queue mimics a named employee.
The voice bot says it is an AI at the start. It mimics no real person.
AI-generated "customer reviews" are published as real.
Never. It is consumer fraud too, not just a transparency breach.
The website chatbot is called "Anna" and has a stock photo of a human.
The name is fine, but "Anna is an AI assistant" is shown up front, and the avatar is not a human.
AI-generated rejection letters to candidates are sent as if HR wrote them.
The letter states that AI assisted and that a named person is accountable.
Emotion recognition runs on customer calls without callers being told.
Emotion recognition in the workplace is prohibited. Customers are told about any analysis that happens.
AI summaries of meetings are sent to clients as verbatim minutes.
Labelled "AI summary, reviewed by <name>". Never presented as minutes.
Social posts are generated by AI in the founder's voice, unmarked.
Labelling when synthetic content could mislead. The founder reviews what goes out in their name.
The internal policy says nothing about disclosure. Each team decides for itself.
One disclosure rule for the whole company, written into the AI policy.
Pocket Operator says it is an AI before it is created and again in its first message. It is written into every operator's instructions, so it holds even months later.
Appropriate technical and organisational measures are required. Shared passwords and unencrypted storage do not count.
Several employees share the same login to an AI tool. No traceability of who entered what.
Personal accounts via SSO. Every input can be traced to a person.
Sensitive data is sent to an AI tool via an insecure browser extension with no encryption.
Only approved, reviewed integrations with encryption in transit. Extensions are blocked by policy.
The AI tool's output, summaries with customer details, is saved in a shared, unencrypted folder open to the whole company.
Output is stored with the same access control as the source data. Need-to-know access, encryption at rest.
There is no access control over who in the organisation may use the AI tool to process customer data.
Roles and permissions are defined before rollout. Logged and reviewed.
API keys to the AI service are hard-coded in shared scripts, with no rotation or access restriction.
Keys in a secrets manager, rotated, with least privilege per use.
Prompts containing customer data are logged in plaintext to a shared log server.
Logs are pseudonymised, access-controlled and encrypted at rest.
An AI agent is given an admin API key "to be safe".
Least privilege: scoped keys, rotated, one per agent.
No multi-factor authentication on the AI tool's account, which holds an export of the customer database.
MFA and SSO on every tool that handles personal data.
An AI plugin can read every shared drive. Installed by one person.
Plugin permissions are reviewed by IT and scoped to what is needed.
Prompt injection: the AI reads an email telling it to forward files, and does it.
Untrusted content is data, never instructions. Tools with side effects require confirmation.
The same AI instance serves several customers, and memory leaks between them.
Tenant isolation per customer. Tested, not assumed.
The AI vendor had a breach. You find out on LinkedIn.
The processing agreement requires breach notice within hours. You have a 72-hour reporting playbook.
AI-generated code with a hard-coded password is pushed to production.
Secret scanning in CI. AI code is reviewed like any other code.
Employees use personal devices and accounts for AI work with company data.
Managed devices and accounts. Policy plus technical controls.
No data protection impact assessment (DPIA) for the AI tool processing customer data at scale.
DPIA done before go-live, reviewed yearly.
Every Pocket Operator reply runs in its own locked sandbox that sees one operator's folder and nothing else. Credentials are encrypted on disk and invisible to it. An automated check against the live server verifies it.
No certification is legally required. But missing basic governance is usually what lets every breach above happen in the first place.
There is no list of which AI tools are used in the organisation. "Shadow AI" is discovered only at an incident.
A register of every AI tool, its owner, its data and its contract. Updated with every new tool.
Expense approval of AI subscriptions happens routinely without anyone checking the vendor's data handling.
The expense flow has one step: "is the tool in the register?" No means stop until it is.
No risk assessment is done before a new AI tool is allowed to handle customer or staff data.
A light risk assessment per tool before any data flows. One page is enough, but it exists.
Employees get no training on what is okay to paste into an AI chat. Trade secrets and personal data mix freely.
One hour of training at onboarding and yearly: what may be pasted, what may not. With examples.
The incident response plan does not mention AI leaks at all. If data leaks through an AI tool there is no defined process.
The incident plan has an AI scenario: who is alerted, how the tool is cut off, what is reported within 72 hours.
An AI policy exists, but it was written by the vendor's sales team.
Own policy, owned by a named person, reviewed yearly.
Nobody owns AI risk. "IT handles it", says legal. "Legal handles it", says IT.
A named AI owner with a mandate and a budget.
AI vendors are evaluated on price and demo only.
Vendor checklist: contract, region, retention, security, sub-processors. Before the demo impresses.
No record of which AI decisions went wrong.
An incident and near-miss log for AI, reviewed quarterly.
Employees cannot report "I pasted something I should not have" without being punished.
A blameless reporting channel. Fix the process, not the person.
AI usage is measured by "adoption", never by risk.
KPIs include risk: tools in the register, contracts signed, incidents.
Model or vendor changes go live without any change control.
A change log. Significant changes are re-assessed before going live.
AI outputs are never tested for bias or errors before use on customers.
Evaluation before deployment, monitoring after.
Nobody can answer a customer's "which AI touches my data?"
A public, plain-language answer, like the map on thepocketoperator.com/data.
The compliance page says "we are compliant" with no evidence whatsoever.
Show the working, row by row, with the thing itself behind every row.
Pocket Operator is not certified and says so plainly. We follow the practices: one row per rule, evidence per row, and one page that changes first if anything changes.
For machines. The same 120 examples are one fetch away at https://thepocketoperator.com/pitfalls.json and listed in https://thepocketoperator.com/llms.txt. Not legal advice. Last generated 2026-09-16.
Read the article: The AI receipt you approved is the breach · Machine-readable copy (pitfalls.json)
Join the launch list and you get your personal start link the day it opens.