In short
Before uploading a confidential client ad to an AI review tool, establish who may process it, what leaves the workspace, who can see the output and how deletion works across active systems and backups. “Not used for training,” “EU hosted” and “encrypted” answer different questions. This checklist turns those claims into evidence requests, with SaliencyLab's current disclosures as a worked example.
Reviewed 16 October 2026. SaliencyLab publishes this guide and provides creative-review services and software. This is an operational procurement checklist, not a legal opinion or an independent security assessment. Your client's requirements and signed agreement determine whether a particular workflow is acceptable.
Begin with the material, not the vendor badge
A public ad, an unreleased product film and a cut containing identifiable customers have different risks. The file may contain more than the finished creative: unreleased claims, an internal codename, an unlicensed track, names in subtitles, a voice recording or a frame from a private dashboard.
Make an inventory before choosing a tool. Record the asset owner, launch embargo, permitted recipients, whether people are identifiable and whether any confidential information is unnecessary for the review. A screenshot with customer names removed may answer the creative question just as well as the original dashboard.
Obtain the client's permission for the intended processing. Permission to edit an ad does not automatically settle permission to send it to an AI provider, a recruited panel or a public sharing link. Confidential business information can matter even when a file contains no personal data.
Where personal data are processed on a controller's behalf, GDPR Article 28 sets requirements for the processor relationship, including documented instructions and contractual terms. Ask the responsible privacy or legal reviewer to establish the applicable roles and requirements. A checklist tick does not itself establish compliance.
The procurement evidence to collect
| Question | Evidence to request | Decision it supports |
|---|---|---|
| Who can authorize this upload? | Client instruction, engagement scope and asset rights | Whether the team may use the proposed workflow |
| What is transmitted? | Data-flow description by processing feature | Whether the transfer is necessary and permitted |
| Who receives it? | Current sub-processor list and human-access scope | Whether all recipients fit the agreement |
| Where is it processed? | Storage and processing locations separately | Whether regional and transfer requirements are satisfied |
| Is it used for training? | Policy for the exact product, account and API | Whether the stated restriction covers this use |
| How long do copies persist? | Active-data, log, cache and backup schedules | Whether retention matches the client's deadline |
| What does deletion prove? | Deletion procedure and restoration safeguards | Whether removed material could return through recovery |
| What can be shared? | Link permissions, expiry, revocation and export behavior | Whether access remains within the approved group |
| What assurance exists? | Report owner, scope, date and exceptions | Whether independent assurance covers the actual supplier |
| What happens after an incident? | Named contact and contractual notification process | Whether the client can respond appropriately |
Download the confidential creative-review checklist (CSV). Record the source and review date beside every answer. Use “unknown” where evidence is missing; do not substitute an optimistic assumption.
Separate four claims that often get blurred
No training does not mean no retention
A provider can prohibit training on customer content while retaining information for security monitoring, caching or service delivery. Google's data-retention documentation separates its training restriction from retention conditions and actions needed for zero retention. This is why an agency should ask about the exact service and configuration, rather than extending a slogan to every data flow.
For any vendor, request separate answers for model training, temporary processing, abuse monitoring, application storage and backups. Check whether a consumer chat product and a commercial API have different terms. Do not assume a private chat window changes the underlying contract.
Storage location does not describe every processor
The primary database may be in one country while transcription, analytics, email delivery or support operate elsewhere. Ask for a feature-level explanation: file storage, image or frame analysis, audio transcription, report storage and operational telemetry.
If a client requires all processing to remain in a particular region, a supplier's primary storage address is insufficient evidence. Resolve that requirement before upload, and include the permitted configuration in the agreement where appropriate.
Encryption does not define access
Encryption in transit and at rest helps protect data in specific circumstances. It does not tell you which workspace users, administrators, processors or link recipients can read it through authorized paths.
Check how collaborators are invited, how permissions change when a contractor leaves, and whether an exported report contains images, transcripts or revealing filenames. A downloaded file can outlive the workspace permission that originally protected it.
A provider's certification does not certify the application
Ask whose assurance report you are reading. A cloud provider's assessed services and controls are distinct from the creative-review vendor's application, staff practices and operational procedures. Request the relevant scope, audit period, exceptions and customer responsibilities. A security-header scan or a successful backup test answers a narrower question than a full independent assessment.
Apply the checklist to SaliencyLab
This is a dated summary of SaliencyLab's own published controls and exceptions. It is not an assurance opinion. Read the linked policies and agree requirements before sharing confidential work.
The processing path
SaliencyLab's sub-processor inventory identifies Supabase for the primary database and object storage in Paris, France. Vertex AI uses Belgium by default; Video Intelligence and Speech-to-Text use global endpoints, and other operational providers can process internationally. An EU-only processing claim would therefore be too broad.
The security page describes private creative storage, signed media access and workspace isolation. Those controls do not mean every report-sharing choice is private to the same audience. Review recipients before distributing an export or creating a share link. If the project needs a stricter sharing policy, settle that requirement before the review begins.
SaliencyLab states that it does not use customer uploads to train its own models and describes Google Cloud processing. This should be read alongside the provider's terms, retention conditions and the particular features used; it is not a blanket claim of zero data retention.
The retention exception matters
SaliencyLab's published retention disclosure distinguishes its commitments from the configuration observed on 8 October 2026. The stated 90-day cold-archive limit was not fully enforced across backup tiers: the check found older snapshots and a 400-day cold-export lifecycle, with seven days of soft-delete recovery. Remediation remains open in the published disclosure reviewed for this article.
That disclosure does not extend a contractual deadline or prove that residual copies have been erased. If a client requires verified deletion within a period the current system cannot demonstrate, stop the upload until the requirement is resolved. Choosing SaliencyLab for a useful diagnostic does not remove this procurement decision.
Deletion from the active workspace and deletion from backups are separate events. Ask how erasure requests are reapplied before restored data returns to service, and what evidence establishes the scope of a recovery test. Recovering one table does not demonstrate recovery of every file and identity or prevention of every erased record's return.
Assurance and contracting
SaliencyLab does not hold its own SOC 2 Type II report or ISO/IEC 27001 certification. Links to infrastructure-provider assurance documents are not a substitute. A procurement process requiring the supplier's own report should record that requirement as unmet unless verified evidence changes it.
The published DPA provides processing and signature terms. Some public pages describe availability by plan differently. Do not infer contractual eligibility from a pricing snippet: confirm the executed agreement, permitted processing and exceptions for your specific engagement before uploading.
The creative specifications page provides a native fictional sample and another checklist for confidential work. Inspect it with non-confidential material first. A sample demonstrates the format and makes it easier to review outputs; it does not prove security, accuracy or commercial results.
Use a staged approval process
Stage 1: test with permitted material
Start with an asset already cleared for public use or a fictional example. Confirm the upload flow, what appears in the report, export contents and recipient controls. Record the product and configuration used. This establishes whether the workflow fits your team before confidential assets enter it.
Do not use an unreleased client asset as a convenient security test. The consequences of a mistaken assumption would already have occurred by the time the test exposed it.
Stage 2: review the actual engagement
List the precise files and processing features needed. Match them to the client's permission and vendor agreement. Have the designated owner resolve missing evidence. A useful approval record says who accepted which requirement on which date, and which material remains prohibited.
For example, a client might permit a public product image but prohibit an unreleased testimonial video until its participant permissions and processing terms are confirmed. “Approved for AI” is too broad to preserve that distinction.
Stage 3: run the creative review with limited access
Invite only the people who need the material. Give the reviewer the minimum context needed to interpret the ad. Keep source observations separate from generated suggestions, and verify every factual claim before inserting it into client-facing copy.
An AI suggestion can introduce an unsupported promise that was absent from the original. SaliencyLab's fictional native QA example explicitly limits what the creative establishes; adopt that habit in the editing brief. Confidentiality and claim accuracy need separate checks.
Stage 4: close the engagement deliberately
Record deliverable recipients, revoke unnecessary access and follow the agreed deletion procedure. Ask what remains in backups and for how long. Preserve the approval and deletion evidence in the client's authorized system rather than spreading another copy of the creative across informal notes.
When to pause the upload
Pause when the client has not authorized the processors, an essential location or retention answer is unknown, a required assurance report does not exist, or an open exception conflicts with the contract. That is a specific unresolved condition, not a verdict that every use of AI review is unsuitable.
SaliencyLab can be evaluated for creative diagnosis and expert interpretation within an agreed scope. For confidential work, fit also depends on the policies above and the client's requirements. Use the evidence checklist to evaluate performance claims separately, and the three-ad budget guide to compare what each research approach actually buys.