Outlook
Part of The marketing automation outlook for England, with four labels that keep forecasts honest
Marketing automation risk scenarios: five controls to test
Use five marketing automation risk scenarios for England to test AI outputs, customer contact, data access, supplier change and operational recovery.
Marketing automation risk scenarios usually begin with a claim nobody checked, a contact reused on the wrong channel, or a release nobody re-tested. These five scenarios pair each control question with a publicly documented case, and an England template shows how to record owner, evidence and stopping rule.
Base date: 5 September 2026. These are editorial exercises for England, not predictions, incident reports or legal findings.
The reviewer chose each case because current UK regulator or security guidance raises the same control question. No platform, message, customer record, response plan or outcome was tested.
The five scenarios are the ones to test when reviewing a marketing-automation platform, whether the stack is Microsoft Customer Insights, HubSpot, Mailchimp, Klaviyo or Salesforce Marketing Cloud. Each pairs a platform review question with the control to test, the case that shows why it matters, and the evidence a buyer or operator should ask to see.
What to take away
- Product settings alone do not establish permission under PECR for email, text and similar messages.
- Keep a versioned test pack, a named change owner and a rollback path for releases.
- Air Canada and DPD anchor the invented-claim case. HelloFresh anchors the permission case.
- Samsung, DeepSeek, CrowdStrike, TSB and Facebook show how data, releases and recoveries fail.
- The England worked example gives the template fields and a filled-in example.
Scenario 1: Air Canada's chatbot invents a bereavement fare
A customer asked Air Canada's website chatbot about bereavement fares in late 2022. The bot told him he could book a full-price ticket and claim a refund within 90 days. Air Canada's real policy did not allow that, and the airline argued in tribunal that the chatbot was a separate legal entity.
The British Columbia Civil Resolution Tribunal rejected that argument in February 2024. It held the airline responsible for its chatbot's words and ordered it to pay C$812 in damages and fees.
DPD supplies a second case. In January 2024 its AI chatbot, added by a supplier, swore at a customer and wrote a poem criticising the company. DPD disabled the AI element within days and linked the behaviour to a system update.
The CMA's guidance on AI agents, published on 9 March 2026, says businesses remain responsible for unlawful agent conduct and should monitor outputs. Qualified consumer-law and advertising review is required for the real case.
Test whether the system blocks release until a person verifies claims against an approved record. Stop the workflow if it cannot preserve the draft, source, approver and final change.
Control owner: marketing operations lead; evidence artifact: approved claims record.
Scenario 2: HelloFresh's £140,000 spam fine
The ICO fined HelloFresh £140,000 in November 2024. Its investigation found the company's subsidiaries sent about 79 million emails and one million texts over seven months, including to customers who had opted out. PECR penalties can reach £500,000.
A journey that reuses a contact for a different message or channel needs a fresh rule check. The ICO's electronic-mail guidance, updated 2024, explains PECR rules for email, text and similar messages. Product settings do not themselves establish permission.
England worked example using the template. This is illustrative, not a real company case.
England worked example
- Owner
- Lifecycle marketing manager named in the journey log
- Evidence
- Signed consent record, journey version and suppression log
- Stopping rule
- Pause the send if suppression fails or the consent purpose does not match the message
- Rehearsal date
- 15 October 2026
Rehearse suppression, preference changes and complaint escalation with synthetic records. Halt the send if the team cannot show the applicable recipient, purpose, channel and exception analysis.
Control owner: lifecycle marketing manager; evidence artifact: consent record and suppression log.
Scenario 3: Samsung staff paste code into ChatGPT
In April 2023 Samsung engineers shared internal source code and meeting notes with ChatGPT. The company banned external generative AI tools and began building an internal alternative. That exposure was not a marketing incident, but the control question is the same.
A January 2025 DeepSeek exposure left chat history and API keys reachable. For a marketing team, the equivalent control question is whether customer data, prompts or suppression logic can leave through an AI tool.
The UK Business Data Survey reports the share of businesses uncomfortable with their data training external AI models. The survey page gives the exact figure; this is an attitude measure, not an incident rate.
The ICO's AI guidance says AI adoption may require governance and risk appetite to be reconsidered. Test access controls, redaction, logging and deletion evidence. Escalate the data-protection analysis to the privacy team.
Control owner: data protection lead; evidence artifact: access log and redaction record.
Scenario 4: a faulty release breaks a live journey
Supplier announcement evidence, 2026: Microsoft's Customer Insights release plan distinguishes previews, planned availability and released features, while warning that plans can change. That does not prove any defect.
On 19 July 2024 a faulty CrowdStrike Falcon sensor update crashed about 8.5 million Windows devices. Flights were grounded and GP appointments were cancelled.
Insurer Parametrix estimated Fortune 500 losses at $5.4 billion, a modelled figure rather than a measured total.
McDonald's ended a voice-ordering pilot with IBM in June 2024 after order errors were filmed and shared. The chain said it would keep testing voice AI with other partners.
For marketing automation, the control question is whether a release changes a journey rule, consent logic or suppression list without a re-test.
Editorial scenario: a release changes a journey rule or permission. Keep a versioned test pack, named change owner and rollback path. Pause deployment when release notes and observed behaviour disagree.
Control owner: release manager; evidence artifact: versioned test pack and rollback log.
Scenario 5: TSB's recovery ran for weeks
TSB's April 2018 platform migration locked around 1.9 million customers out of their accounts, and faults persisted for weeks. The bank put the cost at about £330 million, and its chief executive, Paul Pester, resigned in September 2018.
Facebook's six-hour outage on 4 October 2021 shows the same pattern. A faulty backbone configuration cut off Facebook, Instagram and WhatsApp, and staff lost access to internal tools and building systems, which slowed the fix.
The NCSC secure-development guidance, published 27 November 2023, recommends documenting data, models and prompts and planning across the supply chain. Apply those ideas to dependencies, logs and recovery evidence without treating the guidance as certification.
Rehearse manual suppression, queue cancellation, data export and customer-support handover. Stop expansion if recovery exceeds the buyer's stated tolerance or leaves uncertain message status.
Control owner: incident response lead; evidence artifact: recovery timeline and manual suppression log.
Review block: review these scenarios quarterly through December 2027 and after a regulatory update, platform release, near miss, complaint or material workflow change. Record what changed and whether the control still works. Do not assign probabilities until the organisation has a defensible incident denominator and observation period.
Turning the scenarios into a platform review
In a demo, a trial or a renewal review, ask to see the five controls in order: a claim held until a person verifies it; a fresh permission check when a contact is reused on another channel; AI features confined to approved data with access, redaction and deletion logs; versioned releases with a named change owner and a rollback path; and manual export, suppression and support handover during an outage. A platform that cannot show those five has not passed the review, whatever the feature list says.
Before you act
- 1. Verify claims against an approved record before release.
- 2. Rehearse suppression and preference changes with synthetic records.
- 3. Test access controls, redaction, logging and deletion evidence.
- 4. Keep a versioned test pack and rollback path.
- 5. Rehearse manual suppression and queue cancellation, and record owner, evidence and stopping rule on the England scenario template.
Common questions
What does the CMA guidance say about responsibility for AI agents?
The CMA guidance says businesses remain responsible for unlawful agent conduct and should monitor outputs, and the Air Canada ruling points the same way. A workflow should block release until a person verifies claims against an approved record.
Why do product settings not settle permission for a send?
The ICO electronic-mail guidance explains PECR rules for email, text and similar messages, and product settings do not themselves establish permission. A journey that reuses a contact for a different message or channel needs a fresh rule check. The ICO's HelloFresh penalty reached £140,000.
What should happen when release notes and observed behaviour disagree?
Pause deployment. Keep a versioned test pack, a named change owner and a rollback path, and stop expansion if recovery exceeds the buyer's stated tolerance or leaves uncertain message status. Microsoft warns that release plans can change.
