Who Owns Customer Data Risk When Marketing Uses AI?

The business deploying the marketing workflow owns the customer-facing response to data risk. Vendors can carry separate legal duties, and the exact allocation varies by market and contract. The company whose name is on the customer journey still needs to trace the information, explain its use and name who has authority to act.

A customer enters her budget, timing and business problem into a chat window on your website. The assistant recommends a package, logs the exchange and passes a summary to a salesperson. Weeks later, she asks where her information went, what happened to it and who can correct the outcome.

Singapore's July 2026 PDPC guidance makes that management problem concrete. It identifies model providers, system providers and system deployers as distinct stakeholders, and says the deployer bears primary responsibility for ensuring its chosen system can meet the deployer's PDPA obligations. The guidance is advisory, the PDPA and subsidiary legislation prevail, and providers may carry duties of their own.

Full Court Press uses two operating terms to turn that responsibility into an inspectable customer journey.

FCP Customer Data Accountability Test

Customer Data Path

The Customer Data Path is the recorded route of one customer's information through every person, system, provider and output involved in an AI-assisted interaction.

It answers: Where did the information go, what happened to it, and where does it remain?

The word one is deliberate. Follow one complaint, sales call, recommendation, image or customer request from first input through processing, sharing, storage, output, retention and deletion. A real interaction gives leaders something they can inspect.

Customer Data Response Owner

The Customer Data Response Owner is the named person who can reconstruct the Customer Data Path, coordinate the teams and providers involved, correct the customer-facing outcome and change the workflow.

It answers: Who can bring the complete answer together and act on it?

This is operating accountability. Legal responsibility depends on the circumstances, contracts and statutory roles.

The FCP Customer Data Accountability Test asks two questions before an AI-assisted customer workflow is approved:

1. Can we reconstruct the Customer Data Path for one real interaction?
2. Is there a named Customer Data Response Owner with authority to act?

One-page approval matrix

A clean answer to both questions gives leadership a documented basis for approval. A missing route or an owner without authority sends the workflow back for correction.

1. The path: follow one customer interaction

A supplier register lists approved tools and contracts. The Customer Data Path follows what actually happened to one customer's information across the people, accounts, settings, integrations and providers involved.

The route might begin in a public chatbot, continue through a CRM summary and lead score, produce a generated service reply, and end as a follow-up task for a salesperson. It might begin when an employee pastes a complaint into a consumer AI account. A white-labelled product can pass the same information through a wrapper, an integration partner and an underlying model provider.

For each step, record:

• the customer information received or inferred;
• the person, account, provider or integration that handled it;
• the purpose and available legal basis or exception;
• the output created and any customer-facing decision;
• where the input, output and logs are stored or shared;
• the retention and deletion position; and
• the person who can correct the outcome or change the route.

The resulting path should let a leader start with the original interaction and reach every copy, summary, output and action that remains relevant.

2. The breaks in the path

Most gaps appear between the approved product name and the way people use the product in daily work.

Product plan and account settings

The exact product and account matter. OpenAI states that inputs and outputs from its business offerings, including ChatGPT Business, ChatGPT Enterprise and its API platform, are excluded from model training by default. Consumer ChatGPT has separate data controls, including a setting that determines whether conversations help improve models.

Anthropic states that chats and coding sessions from Claude for Work and its API are excluded from training by default, subject to limited routes such as explicit feedback or customer participation in a development programme. Its Free, Pro and Max consumer plans have separate controls and may use chats or coding sessions for model improvement when the user allows it, through explicit feedback or in safety-review cases.

Training is one entry in the path. Leadership also needs the account owner, retention setting, connected services, access permissions, public-sharing controls and the categories of information employees may enter.

Integrations, outputs and retention

A chatbot response can become a CRM note, an email draft, a lead score, an analytics event or a task assigned to another team. Each new output can have a different access list and retention period. The path needs to continue until the business can say where each material copy remains and how deletion or correction reaches it.

Provider documentation should also identify subprocessors, hosting locations, security access, model or product improvement uses, and the changes that require a fresh approval. A contract name alone cannot reconstruct those handoffs.

Public sharing

Anthropic's sharing guidance says consumer chats are private by default and that creating a share link makes a snapshot available to anyone with the link. Team and Enterprise sharing is limited to people in the same organisation.

Search Engine Land reported on 28 July 2026 that some public Claude share URLs appeared in search results. Its headline described "private chats", while the reported mechanism involved public share pages. TechCrunch separately reported public Claude share links and Artifacts appearing in search. No reviewed source established exposure of ordinary unshared chats.

The leadership question is specific: which public snapshots or Artifacts exist, what customer information appears in them, who created them and who can revoke them?

3. The owner: give one person authority to close the response

The Customer Data Response Owner can sit in commercial leadership, operations, product, privacy or another function. The decisive requirement is the authority attached to the named person.

That person must be able to:

• obtain the complete Customer Data Path from every team and provider involved;
• coordinate legal, privacy, security, procurement, marketing and customer-service action;
• explain the outcome to the customer in usable language;
• correct the customer-facing response or record where appropriate; and
• pause or change the workflow when the evidence or control is weak.

The role also protects the employee who faces the customer. A comparative field study in the Journal of Management Studies' 2026 volume, first published online in November 2025, examined experts across banking, biotechnology and recruitment who had to present AI-generated decisions they could not easily overrule. A Customer Data Response Owner gives that employee access to an explanation, an escalation route and someone with authority to challenge the workflow.

Naming a committee or shared inbox leaves the central question unanswered. Leadership should be able to identify the person who can bring the complete answer together and decide what changes next.

4. The evidence leadership should expect

The evidence should connect the real customer interaction, the provider controls, the customer explanation and the named decision-maker.

Purpose, notice and customer choice

The final PDPC Advisory Guidelines on the Use of Personal Data in Generative AI were issued on 20 July 2026. They build on the PDPC's March 2024 Advisory Guidelines on the Use of Personal Data in AI Recommendation and Decision Systems, which cover recommendation, prediction and decision uses such as personalisation and recommendation engines.

The July guidance addresses user data used for large-scale Generative AI model training or fine-tuning. Unless deemed consent or a relevant statutory exception applies, consent is required and the individual must be told the purpose. Paragraph 4.6 encourages organisations, to the extent practicable, to explain the model functions that require personal data, the types of personal data involved, how the data will be used for training or fine-tuning, and how an individual can decline or withdraw consent.

Paragraph 4.8 allows consent to this additional use to be a condition of service only to the extent reasonable for providing that product or service. Paragraph 5.1 confirms that properly anonymised data falls outside the PDPA. The 2024 guidance says the Business Improvement Exception can be relevant to personalisation or recommendation engines when its conditions are met. These positions require a documented, use-case-specific assessment.

The separate Generative AI Chatbot Transparency Guidelines are voluntary. They encourage a Chatbot Information Card explaining in plain language what the chatbot does, its limitations, how data may be handled and how users can report issues. DBS, Google, Meta, OCBC and Singapore Airlines have indicated that they intend to use the guidelines as a reference. CNA reported on 20 July 2026 that organisations intended to refer to the guidance over the next six to twelve months; the source does not establish completed information cards.

An accurate card depends on an accurate Customer Data Path. A usable reporting route depends on a Customer Data Response Owner who can act on what the customer reports.

Provider and workflow evidence

Before approval, leadership should expect five connected records:

These records should be reviewed while the campaign design, provider and customer promise can still change.

Market-specific overlay

The operating test stays consistent across markets. The legal analysis must follow the organisation, jurisdiction, sector, data and intended use.

The EU GDPR places accountability around controllers and processors. The EU AI Act adds duties based on use and risk, including transparency provisions applying from 2 August 2026. The UK ICO says all data-protection provisions in the Data (Use and Access) Act 2025 were in force by 19 June 2026. A Congressional Research Service review describes the US federal framework as sectoral and records at least 19 state comprehensive consumer privacy laws. The Australian privacy regulator says obligations can apply to personal information entered into an AI product and to information generated or inferred in its outputs.

This table is a commercial orientation. Jurisdiction-specific advice should determine the final legal position for each deployment.

5. Test one live customer journey in the next meeting

Choose one recent complaint, sales call, recommendation, image or customer request. Ask the team to reconstruct it in the room:

• What information did the customer provide, and what did the workflow infer?
• Which exact product plan, account and settings were used?
• Which people, providers, integrations and subprocessors handled the information?
• Which outputs, summaries, logs or public shares were created?
• Where does each material copy remain, and how would correction or deletion reach it?
• What was the customer told about purpose, AI use, sharing and choice?
• Which market, sector and data-type requirements apply?
• Who is the Customer Data Response Owner, and can that person correct the outcome, revoke a share, pause the workflow and require a new approval?
• Which provider, model, purpose or data change sends the workflow back for review?

A generic policy answer should trigger a request for the actual interaction, account settings, provider evidence and named decision-maker.