A UK small business can start with five contained AI automations: enquiry sorting, quote drafting, appointment follow-ups, invoice-chase drafts and weekly reporting. For each, map the personal data entering and leaving the tool, identify the lawful purpose, check the supplier's processor terms and keep a person in charge of consequential decisions before switching it on.
The practical UK GDPR test is a data-flow drawing, not a supplier's marketing badge. Write down where a record starts, which apps receive it, who can see the output, how long each copy remains and how a correction or deletion request would reach every relevant system. Ask a qualified adviser about an uncertain legal basis or a high-risk use.
The gate to use before any of the five set-ups
Choose one narrow job and make a sample record that resembles real work without containing a real person's details. Trace it from source to AI input, output, human reviewer, final destination and deletion point. If you cannot identify the data owner at each hop, pause the connection. The business may be the controller deciding why and how customer or staff data is used; a supplier processing it on the business's instructions may be a processor. Roles depend on the actual arrangement, so read the terms.
Article 28 of the UK GDPR, which the ICO's guidance on controller and processor contracts explains, requires written terms whenever a processor acts for a controller: processing only on documented instructions, confidentiality, security measures, rules for engaging sub-processors, help with rights requests, deletion or return of the data at the end of the contract, and the information needed for audits. Check the specific service and plan you intend to use. A privacy policy alone is not the same as the processing agreement. If a supplier uses the data for its own purposes, the roles and lawful-basis analysis may be different.
Keep an implementation record with seven fields: purpose; categories of people; exact fields sent; supplier and sub-processors; lawful basis and privacy notice location; retention and deletion route; human approval and escalation. Add security access, transfer and data protection impact assessment questions where risk warrants them. The ICO's AI and data protection risk toolkit is a useful prompt for those questions. The ICO marks parts of its AI guidance as under review following the Data (Use and Access) Act 2025, so read the version that is current on the day you assess. This is a working checklist, not a substitute for legal advice.
Consider a four-person electrical contractor with 20 enquiries a week. The figures below are illustrative. It wants faster admin, but its inbox includes names, addresses, phone numbers and occasional descriptions of a vulnerable customer's circumstances. The sensible first test uses synthetic enquiries and an internal reviewer. It does not connect the full inbox to a consumer assistant simply because the assistant can summarise an email.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy now1. Enquiry sorting: classify the work, not the person
Start by routing inbound enquiries into categories such as new quote, existing job update, invoice query and urgent safety concern. A rule can identify a form field or subject line; AI helps when the customer's free text is messy. The output should be a category, a short reason and a confidence warning, not an automatic rejection. An urgent electrical issue goes to the duty person even if the classifier is uncertain. For the broader routing design, see when rules are better than AI.
Data-flow check: the web form or inbox is the source. The automation may copy name, contact details and message into an AI service, then write a tag into a CRM or shared inbox. Ask whether the full message is necessary. For initial categorisation, a redacted text excerpt may suffice. Avoid attaching photographs, identity documents or unrelated email history. Document whether the AI provider stores prompts and outputs, who has access, where processing occurs and how records are deleted.
Processor check: inspect the contracts for the form, connector, AI service and CRM, not just the model provider. A no-code connector can receive the same personal data as the assistant. Confirm who is a processor for each step, the applicable written terms and how sub-processors are listed. The UK GDPR check also includes a documented lawful basis, privacy information that describes the purpose accurately, a rights-response route and appropriate access controls.
Illustrative prompt: “Classify this redacted enquiry into new quote, active job, billing or urgent human review. Do not infer health, finances or customer worth. If the text mentions danger or an unclear category, output urgent human review with the quoted phrase.” A sample output might say “new quote, because the customer asks for a socket installation.” If the message also says there are sparks at the consumer unit, the correct output is urgent human review. Test that failure before launch.
2. Quote drafting: keep price and scope with the estimator
A second useful automation turns a job survey and an approved rate sheet into a draft quote. It can organise the scope, exclusions and questions for the estimator. It should not invent a parts price, commit to a start date or accept the job. The estimator checks measurements, materials, labour, tax treatment and terms. A small service firm can first use AI on a redacted description with no customer identity, then paste the reviewed text into its quote system.
Data-flow check: a survey may include a household address, photographs and notes about occupants. Write down which of those fields truly help draft the scope. Often the AI needs “replace three fittings in occupied property” rather than the exact address or the customer's contact details. Keep photographs out until the supplier's handling is checked and the use is necessary. Record the destination of the generated quote and whether the source remains in a chat history.
Processor check: the quote platform and AI service each need appropriate terms if they process personal data on your behalf. Check whether a staff member can paste the entire customer email into an unapproved personal account. A written staff rule should name the approved account, permitted fields and the human who signs off every quote. If the quote could materially affect the customer, the reviewer must exercise real judgement rather than rubber-stamp text.
In the electrical contractor example, a surveyed job has two rooms, four outlets and one existing circuit to check. An illustrative AI draft includes a new consumer unit because “that is commonly required”. The estimator removes it: there is no evidence the unit needs replacement. The prompt is then tightened to say “use only listed line items; put missing information under questions”. The correction log records the invented line so the next three sample jobs can test for it.
3. Appointment follow-ups: send from verified booking data
Appointment reminders are often a better early automation than open-ended chat. The booking system knows the visit date, time window and contact preference. A deterministic trigger can send a standard reminder; AI may adjust a short explanation when the job type requires preparation. The booking record, not an AI conversation, remains the authority. Send a sample to an internal address before any customer message goes out.
Data-flow check: the booking app passes a name, contact address, date, location and perhaps job description to the messaging service. Minimise the payload. A basic reminder needs the date, time window and reply route, not a detailed description of a customer's home. Check whether a provider processes message content, delivery metadata and replies. Include each tool in the data map and set access permissions so a junior user cannot download the full booking list.
Processor check: compare the booking and messaging providers' terms, sub-processor information and retention controls. Document what happens to messages when a customer asks for correction or erasure, bearing in mind any records you must retain for another lawful purpose. The ICO's approach to lawfulness calls for a purpose and appropriate basis, with accurate privacy information. A service reminder and a marketing offer should not be treated as the same message merely because both can be sent by the same tool.
For illustration, the system holds a visit on Tuesday between 09:00 and 11:00. AI drafts, “Your engineer will arrive at 09:00.” That is too certain. The approved version says, “Your visit is booked for Tuesday, 09:00–11:00. Reply if access details have changed.” If the diary changes, suppress the original reminder and use a reviewed change notice. This small test catches a common trust problem: the assistant turning a window into a promise.
4. Invoice chasing: draft the tone, verify the ledger
Invoice chasing can save repetitive writing, but the ledger must decide whether an invoice is outstanding. Feed AI an invoice reference, due date and approved status only after the accounts person has checked credits, disputed work and recent payments. Ask it for a polite draft. Do not let it decide the debt exists, add a late fee or send a threat. For the mechanics of the reminder sequence, see invoice reminders for trades.
Data-flow check: the accounting app holds payer details, amounts, history and sometimes bank information. The AI draft usually needs a smaller set: first name placeholder, invoice reference, approved balance, due date and a contact route for disputes. Keep bank details in the approved invoice or payment portal rather than retyping them in a generated email. Record where the output is saved and who can access the audit trail.
Processor check: confirm the accounting system, automation connector, AI service and email platform's roles and terms. Review data transfers and retention. Do not assume that an account's default settings cover a separately connected API or add-on. The UK GDPR check also includes a clear purpose, secure access and a process for correcting an erroneous invoice record. Ask a qualified adviser if the lawful basis or retention position is uncertain.
In the illustrative contractor, invoice 1048 is for 620 and appears 12 days overdue, but the customer paid 400 yesterday. An AI prompt based on a stale export writes “Your 620 balance is overdue.” The accounts person catches it at approval and refreshes the ledger, leaving 220 for review. The fix is a data freshness gate: the amount and payment timestamp must be checked immediately before a draft is generated. A polished email with a wrong balance is still a failed automation.
5. Weekly reporting: aggregate before AI sees the detail
A weekly summary of enquiries, booked work, completed jobs and overdue invoices can be useful with little personal data. Calculate counts and totals in the source systems first. Give AI an aggregate table and ask for plain-English observations, anomalies and questions. Do not feed it a customer-level export merely to produce a paragraph for Monday's meeting. This workflow is often the least complex privacy choice when the numbers are genuinely aggregated.
Data-flow check: note which systems supply the figures, where the summary is compiled and whether any small group could identify an individual. A report saying “one complaint from the only customer on Friday” may reveal a person even without a name. Remove identifying examples or use a broader period. Confirm who can open the report and how long it remains in the assistant or document system. Compare the output with the original totals before circulation.
Processor check: if the table is truly anonymous, the particular AI input may fall outside personal-data rules, but the surrounding workflow may still contain personal data in source apps, logs or drafts. Do not declare a dataset anonymous just because names are absent. If personal data enters a processor, the contract and security checks remain. Record the assessment and revisit it when the report gains drill-down links or staff performance fields.
Illustrative output: “Enquiries rose from 18 to 26, a 44% increase.” The manager checks the calculation: eight divided by 18 is about 44%. Then she spots that three of the 26 were duplicate form submissions. The corrected comparison is 18 to 23, about 28%. AI can explain a verified table, but an attractive percentage is not evidence that the source data was deduplicated. The messy data cleaning tutorial helps with that upstream task.
Use a one-page UK GDPR decision record
For each of the five, record the answer to the same questions in a single page. Who decides the purpose? What personal data is involved? What lawful basis has been assessed? Which people are told about the processing, and where? Which supplier receives which fields? Are there processor terms covering the real service and any sub-processors? How are access, deletion, corrections and incidents handled? Does the use create enough risk to require a formal impact assessment? Who reviews the output before it affects someone?
A filled-in miniature record for enquiry sorting might read: “Purpose: route service requests. Source: web form. Fields sent: redacted request text and enquiry ID; contact details stay in the inbox. Output: category and human-review flag. Reviewer: office manager. Supplier terms: contract checked for AI service and connector. Retention: confirm in both products. Privacy notice: update before live use. Stop rule: any urgent request misrouted.” The unfinished retention field is deliberately visible; it blocks launch until answered.
The ICO's guidance makes a useful distinction between decision support and solely automated decisions with significant effects. A human reviewer has to be able to understand, challenge and override an AI suggestion. Merely pressing approve after seeing a score is not meaningful review. For routine admin, design the system so the reviewer sees source evidence and an exception reason. For recruitment, credit, employment and other consequential uses, seek specialist advice before treating a similar pattern as safe.
Check transfers and vendor changes during procurement. Ask the supplier where the relevant service processes and stores data, what transfer mechanism it relies on where applicable, how sub-processors are notified, and how your business retrieves or deletes data at contract end. Keep a dated copy of the terms you reviewed. A future plan change or new connected app can alter the data flow, so the review is tied to the exact configuration, not the brand name.
Choose the first launch by risk and reversibility
Score each candidate on four practical dimensions from one to three: personal-data sensitivity, cost of a wrong output, ease of human review and ease of rollback. A weekly aggregate summary might score low on the first two; an automated invoice chase may score higher because a wrong demand affects trust. The numbers are your internal prioritisation tool, not a legal test. If a use could significantly affect a person, stop the simple scoring and ask for a proper assessment.
Run the chosen workflow in shadow mode for two weeks: AI drafts or tags, but no external message is sent automatically. Compare at least ten varied cases, including a complaint, a duplicate, a changed booking and an unusual request. Record false routes, invented details, privacy leaks and time spent correcting them. The shadow-mode pilot tutorial gives the testing pattern. Launch only the part that passes and keep the manual route available.
Finally, schedule a review after a month. Check whether the source fields, vendor terms, access list, retention settings and privacy wording still match the live process. Review corrections and complaints, not just time saved. A modest workflow that a person can explain and reverse is a stronger first step than an impressive chain whose data copies and decisions nobody can trace.
Put the review date and the named owner in the decision record. When that person leaves or the tool is replaced, hand over the sample tests, the current data map and the last approved supplier terms. That small administrative step prevents an apparently identical automation from continuing under an unreviewed account or contract.
Frequently asked questions
Does removing a customer's name make an AI prompt anonymous?
Not necessarily. An address, order number, unusual complaint or combination of fields may still identify someone. Test what the recipient of the prompt could infer, minimise the fields, and treat pseudonymised records as personal data where identification remains possible.
Do I need consent for every AI automation?
No single lawful basis fits every workflow. Document the purpose and assess the appropriate basis with someone responsible for data protection. Consent is not a shortcut for poor necessity or unclear vendor terms. The ICO expects the chosen basis and purposes to be documented and explained in privacy information.
Can a supplier's GDPR badge replace a processor contract?
No. Check whether the supplier processes personal data on your instructions, obtain the required written terms where it is a processor, and review security, sub-processors, retention and transfer arrangements. The controller remains responsible for choosing a provider that offers sufficient guarantees.
Further reads
- GDPR and AI Tools: What a Small Business Must Do — Review the general privacy duties behind these five workflows.
- Are ChatGPT, Claude, Gemini and Copilot GDPR-Compliant? — Compare assistant accounts and business data controls.
- How to Add Human Approval Steps to AI Automations — Build a recorded approval point into the workflow.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: UK GDPR Article 28 (legislation.gov.uk); ICO guidance How do we ensure individual rights in our AI systems? and AI and data protection risk toolkit; ICO AI audit toolkit pages on lawfulness, human review, and contracts and third parties. All business figures are illustrative.