What are the AI trust and data privacy deal-breakers for a marketing platform?
Six answers end a marketing AI vendor evaluation on the spot: training on your data, hidden sub-processors, no data residency, no logs, undisclosed automated decisions, no deletion proof.
This is for marketing leads at banks, insurers and fintechs who sign off AI tools or inherit them. You can score any vendor, including those already live.
Twenty questions sort a vendor into pass, conditional or fail. Six are AI trust and data privacy deal-breakers by Leapbuzz rule: the buyer cannot discharge its own duty without the vendor's answer. A strong demo does not fix a fail on any of the six, though a fail may mean limiting the tool to non-personal data, not a ban.
Regulator texts here are guidance we read, not legal review.
- It trains models on your prompts, outputs or uploads by default.
- It will not name its sub-processors.
- It cannot say where your data is stored and processed.
- Model activity is not logged, or the audit log cannot be exported.
- A feature shapes decisions about people and the vendor cannot say how.
- It cannot delete your data and prove it.
Usage policy and review gates sit in our guide to enterprise AI content governance; this post is the procurement scorecard, part of our AI strategy work. For governance of AI systems beyond marketing, see Origin Pi, our parent company.
Why does the buyer, not the AI vendor, carry the privacy risk?
Regulators in Singapore and Australia say the organisation using an AI tool keeps its obligations; Canada's commissioners say so for AI-driven decisions.
The Monetary Authority of Singapore set it out in its release of 7 October 2026: "FIs remain accountable for AI used in the services they deliver, including AI developed, operated or provided by third parties." Paragraph 5.11, in Section 5 of the Guidelines, adds that where residual risk cannot be kept within appetite, the institution "should consider limiting or suspending the third-party AI service, or replacing the relevant third-party provider".
Sections 3 and 4 apply from 7 October 2027 and Sections 5 and 6 by 7 October 2028. Paragraph 2.1 scales the Guidelines to size and risk, so a copywriting tool at a bank or financial institution may sit at the basic end.
The PDPC's Advisory Guidelines on Use of Personal Data in Generative AI, issued 20 July 2026, state at 9.1 that "System Deployers bear primary responsibility for ensuring that the Generative AI Systems they have chosen to use can meet their obligations under the PDPA."
"Realising these benefits sustainably requires financial institutions to understand and manage the risks that come with increasingly capable AI systems."
Ho Hern Shin, Deputy Managing Director, Monetary Authority of Singapore, media release, 7 October 2026 (MAS)
Australia's OAIC fact sheet of September 2026 says the transparency obligation will "typically remain with the APP entity". Canada's federal, provincial and territorial privacy regulators wrote on 7 December 2023, in non-binding principles, that "accountability for decisions rests with the organization, and not with any kind of automated system". Malaysia and the US: no equivalent AI text read, so the six live in your contract.
| Deal-breaker | Singapore | Australia | Canada |
|---|---|---|---|
| Trains on your data | PDPC GenAI 7.1: providers using prompts for model development must meet the PDPA | OAIC: check terms that collect inputs "for further training" | OPC: consent where it is the legal authority |
| Hidden sub-processors | MAS 5.11(b) supply chain and (c) concentration, for FIs | OAIC: whether data "could be disclosed outside of Australia" | Contract term |
| No data residency | PDPC GenAI 7.5: providers should document residency and retention (example: info sheet) | OAIC: ask "where the servers are located" | Contract term |
| No logs | PDPC GenAI 9.3: agent activity data (our inference) | Needed to evidence an APP 1.8 statement (our inference) | OPC Principle 7: retention schedules (our inference) |
| Undisclosed automated decisions | PDPC 2024 Guidelines 9.5: what to tell consumers | APP 1.7 to 1.9, from 10 Dec 2026 | Quebec: notice of decisions based exclusively on automated processing (CAI) |
| No deletion proof | PDPA s.25 retention, per PDPC GenAI 7.2 | OAIC: data "will need to be destroyed or de-identified" | OPC Principle 7: prompts "should not be retained" unless required |
Which 20 questions should you ask a marketing AI vendor about data privacy?
Ask 20 questions in five groups: data, model, decisioning, security and contract. Six are deal-breakers; the other 14 separate pass from conditional.
Send them in writing. A verbal "we don't train on your data" from a sales engineer is worth nothing at audit; paragraph 5.11(a) of the MAS Guidelines (Section 5) says an FI may consider independent assessments, "i.e. not self-attestations".
Pass standards are Leapbuzz procurement criteria informed by regulator guidance, not regulator requirements; "our inference" marks a mapping we drew.
| Question | Pass | Conditional | Fail | Reference |
|---|---|---|---|---|
| 1. Does the vendor train any model on our prompts, outputs or uploads? deal-breaker | Off in the contract | Off via a setting | On or silent | OAIC AI products guidance; PDPC GenAI 7.1 |
| 2. Where are our prompts, outputs and files stored and processed? deal-breaker | Named regions, pinned | Region list, no pinning | "Global" | OAIC AI products guidance; PDPC GenAI 7.5; pinning is Leapbuzz's |
| 3. Can you delete our data, inference logs included, and confirm it in writing? deal-breaker | Window plus certificate | Deletion on request | No path | PDPC GenAI 7.3, 7.5; OAIC; certificate is Leapbuzz's |
| 4. What is the default retention for prompts and outputs? | Stated, configurable | Stated, fixed | Unstated | PDPC GenAI 7.2 (PDPA s.25) |
| 5. Can we block categories of personal data at input? | Field rules or DLP | Policy only | No control | PDPC GenAI 9.3; OAIC best practice (our inference) |
| 6. Does the tool infer or generate personal information? | Documented, can be switched off | Documented | Unknown | OAIC: inferred data is APP 3 collection |
| 7. Which foundation models sit underneath? | Named, versioned | Named | Withheld | MAS 5.11(b) |
| 8. Will you notify us before adding AI features or changing the model? | Contractual notice | Release notes | No notice | MAS 5.11(d), (f) |
| 9. Is there an exportable audit log of model activity, with a retention limit? deal-breaker | Exportable, retention capped | Logged, no export | Not logged | OPC Principle 7; PDPC GenAI 9.3 (our inference) |
| 10. Is there an independent assessment covering AI risks? | Third-party audit or certification | Under way | Self-attestation only | MAS 5.11(a) ("may consider") |
| 11. Does any feature make or substantially inform a decision that significantly affects a person? deal-breaker | No, or yes and documented | Yes, documentation in progress | Yes, cannot describe it | OAIC APP 1.7 fact sheet; CAI (Quebec) |
| 12. Will you give plain-language wording on how personal data feeds those decisions? | Ready-made text | On request | Refused | OAIC fact sheet; PDPC 2024 9.5 |
| 13. Can a human review and override an output, with the override recorded? | Yes, logged | Yes, unlogged | No | OPC Principle 5 (our inference); CAI (Quebec) |
| 14. What is your breach notification timeline to us? | Stated in hours | "Prompt" | None | PDPC GenAI 8.3(c) |
| 15. Do you support SSO, role-based access and encryption? | All three | Two | One or none | PDPC GenAI 7.5 example; SSO is Leapbuzz's |
| 16. Have you tested for prompt injection and data leakage? | Metrics shared | Tested, no metrics | Not tested | PDPC GenAI 8.2, 8.3(b) |
| 17. Is our data kept separate from other customers' data and pooled training sets? | Contractual segregation | Logical separation | Mixed | Leapbuzz rule |
| 18. Will you list every sub-processor with its location, and tell us before changes? deal-breaker | Full list plus notice | List, no notice | No list | MAS 5.11(b), (c); OAIC overseas disclosure |
| 19. Do we get audit rights and data protection clauses? | Both | Data protection only | Neither | MAS 5.11(f) |
| 20. Who owns outputs, and do you indemnify IP claims? | We own; indemnity | We own; no indemnity | Vendor owns or unclear | Leapbuzz rule |
Half the data questions are deal-breakers. Security has none, since weak answers are often fixable by contract, but two security misses cannot pass.
How should the procurement gate use the checklist?
Run the six deal-breakers first as a hard gate, then score the other 14; a vendor needs 12 of those 14, with at most one security miss, to pass outright.
Start with tools already live. MAS footnote 21 calls out SaaS "which may not be explicitly sold as AI, but contains AI features". Our marketing stack audit lists them fast.
Made-up round numbers. A Malaysian insurer lists 10 AI tools. Two fail the gate: a copywriting add-on that trains on inputs by default, and a chatbot plugin with no logs. Five are conditional, mostly on breach timelines. Three pass. The fails get 30 days to fix or go.
If you cannot say which tools train on customer prompts or where the data sits, that is the gap. We run the 20 questions across your stack and return a scored vendor map with fixes per tool; ask us for one.
Vendor scorer: pass, conditional or fail in 20 ticks
Tick a question only when the written answer meets the Pass column; Conditional counts as unticked, and the six deal-breakers take Pass, Conditional or Fail.
FAIL
Failed deal-breaker: 9 (exportable audit log). Do not onboard until fixed. Other questions passed: 12 of 14.
Formula: any deal-breaker at Fail, or 7 or fewer of the other 14, = FAIL. PASS needs all six at Pass, 12 to 14 of the other 14 and 3 of 4 security questions (14 to 17). Anything else is CONDITIONAL. Illustrative heuristic, not a regulator's method.
The same scoring as a static table
| Result | Rule |
|---|---|
| FAIL | Any deal-breaker at Fail, or 7 or fewer of the 14 |
| CONDITIONAL | No Fail and 8 or more of the 14, but a Conditional deal-breaker, two security misses or under 12 of the 14 |
| PASS | All six at Pass, 12 to 14 of the 14, at most one security miss |
Which marketing AI decisions need disclosure from 10 December 2026?
From 10 December 2026, programmatic advertising for significant goods or services and insurance eligibility programs are the marketing cases the OAIC lists as generally in scope.
The OAIC release of 30 September 2026 sets the date. Its fact sheet lists "programmatic advertising relating to the decision to sell significant goods or services" among decisions generally considered in scope, beside programs "involved in assessing an individual's eligibility for insurance".
The test has three parts: the program makes, or is substantially and directly related to making, a decision; the decision could reasonably be expected to significantly affect a person; and personal information is used. A tool that uses none is out. For insurance marketers, a quote flow that filters who sees an offer is, in our inference, the obvious candidate. Where unsure, the OAIC says to "take a cautious approach" and disclose.
Quebec applies a different test. The Commission d'accès à l'information says a business must inform a person subject to a decision based exclusively on automated processing, and let them present observations to a reviewer.
A vendor's trust page is not your disclosure. The obligation stays with you, so question 11 fails if the vendor cannot describe in writing what its feature decides.
Do this today
Send the six AI trust and data privacy deal-breakers, as written questions, to the vendors of your three busiest AI tools this week.
See company data in AI tools and our EU AI Act marketing checklist. Leapbuzz runs AI strategy and vendor reviews across Singapore, Malaysia, Australia, the US and Canada.
- Book every answer into the AI inventory.
- Re-run all 20 on each new AI feature.
Decision rule: a vendor that fails any of the six does not touch personal data, and one that has not answered all six in writing after 30 days is limited to non-personal data until it does. Marketing guidance, not legal advice.