AI Can Make the Ad. Your Evidence File Makes It Publishable.
Generative AI can produce a sophisticated campaign faster than a traditional creative process. It cannot make the campaign publishable on its own.
That requires an evidence file: who approved the concept, which likeness and voice rights were licensed, what source material entered each model, which product claims were substantiated, who reviewed the final asset, where it may run, and how it can be withdrawn.
SKF’s new industrial campaign featuring an AI-generated Greta Garbo is a useful current example. It is also a reminder that the most valuable marketing control is not a generic “AI approved” checkbox. It is a complete chain of evidence tied to the exact creative asset that reaches the market.
What happened
Documented by SKF: In a September 30 company announcement, SKF said it recreated Greta Garbo using AI for a campaign promoting magnetic-bearing solutions and a payment model tied to energy savings. The company connected the concept to Garbo’s appearance in an early film promoting Swedish industry.
Reported by Reuters: Reuters reported on October 2 that the production used multiple image and motion models from ByteDance, Google, and Kuaishou. According to SKF, the visual likeness was generated from text prompts rather than archival images or footage. The voice used material from an early Garbo film with permission, and a human performer helped establish pacing and cadence.
Rights and approvals reported: SKF told Reuters that the production was developed with Garbo’s estate and family. That is a documented company statement; the specific contracts, territories, duration, permitted media, and approval terms are not public in the sources reviewed.
Company claims, not independent findings: SKF describes potential energy, wear, maintenance, availability, and operating benefits from its magnetic-bearing technology. This article does not independently validate those performance claims or the economics of SKF’s payment model.
Public reaction is opinion: Commentary on the campaign has ranged from praise to strong criticism. Those reactions show reputational uncertainty, not proof that the campaign succeeded or failed commercially.
The operating lesson: the creative file is only half the deliverable
AI marketing workflows often focus on prompts, outputs, and speed. A production-ready workflow must also preserve the authority to use the output and the evidence behind every objective claim.
This matters especially in industrial and B2B marketing, where creative content can include equipment performance, energy savings, maintenance intervals, safety language, ROI projections, customer references, technical drawings, certifications, employee likenesses, and partner trademarks. A polished video or product page can move through the organization faster than its evidence does.
The solution is a campaign evidence file managed through the marketing process and connected to the company’s business systems.
What belongs in the evidence file
1. A source and rights register
List every input used to create the asset: photographs, video, audio, scripts, product documentation, customer data, brand assets, stock media, performer contributions, and reference material. For each item, record the owner, license, permitted uses, territories, duration, media, attribution requirements, and restrictions.
A text prompt does not automatically eliminate rights questions. The U.S. Copyright Office defines a digital replica as a digitally created or manipulated image, video, or audio recording that realistically but falsely depicts an individual. Its AI initiative also notes that digital replicas can be authorized or unauthorized and that legal protections vary.
2. Identity and performance consent
For any recognizable person, voice, customer, employee, expert, or synthetic stand-in, document who authorized the use and what the authorization covers. Include whether the approval extends to edits, translations, derivative versions, future campaigns, paid media, partner channels, and model training.
For a deceased figure or historical likeness, do not assume that copyright clearance, estate cooperation, publicity rights, and voice rights are the same thing. Route the specific use through qualified counsel where the stakes justify it.
3. A model and vendor ledger
Record every model, version, vendor, account, and material setting used for images, motion, voice, copy, translation, and editing. Include the relevant commercial-use terms, retention settings, training opt-outs, geographic processing commitments, and known restrictions. Multi-model creative chains need the same vendor and model risk review as other production systems.
4. Claim substantiation before generation
Create an approved claim library before the model writes copy. Each objective statement should link to the test, specification, calculation, certification, or other evidence that supports it, along with the owner and expiration or review date.
The FTC’s advertising-substantiation policy says advertisers should possess a reasonable basis for objective claims before dissemination. AI does not change that sequence. A model can restate or exaggerate a claim in seconds; it cannot create the evidence the business should already hold.
5. Deterministic content checks
Before human review, automatically compare the asset against approved product names, model numbers, specifications, prices, dates, trademarks, mandatory disclosures, prohibited phrases, and geographic restrictions. Flag unsupported numerical claims, altered disclaimers, missing units, and statements that do not match the claim library.
6. A real approval gate
The final approver should see more than a thumbnail. Present the exact asset, source list, rights status, claim references, intended channels, audience, geography, disclosure decision, expiration date, and rollback owner. Bridge Road’s human approval controls should identify what the approver is accepting—not merely record that someone clicked “approve.”
7. A publication and version record
Assign the approved asset a version, checksum, publication window, channel list, and campaign owner. Do not allow a later model edit, translation, crop, voice change, or platform adaptation to inherit approval automatically. Material changes return to validation and approval.
8. Monitoring and rollback
Monitor complaints, takedown requests, platform rejections, claim challenges, partner concerns, and unexpected audience interpretation. The rollback plan should identify who can pause ads, replace web assets, notify distributors or partners, preserve evidence, and move to a non-synthetic backup creative.
A workable release flow
- Brief: Define the audience, business objective, channels, claims, identities, and risk level.
- Clear: Verify source rights, likeness and voice authority, trademarks, product evidence, and vendor terms.
- Generate: Use approved tools and preserve prompts, inputs, models, versions, and operator decisions.
- Validate: Run deterministic checks against the product record and claim library.
- Review: Require brand, technical, and risk approval appropriate to the asset.
- Release: Publish only the approved version to approved channels and territories.
- Monitor: Track response, exceptions, complaints, and claim performance.
- Withdraw: Use a named owner and tested process to pause, replace, correct, or archive the campaign.
How this helps an operating business
An industrial manufacturer can generate product video, translated distributor content, service illustrations, and application-specific ads much faster. An eCommerce team can create channel variations at scale. A distributor can turn manufacturer documentation into vertical-specific campaigns. The evidence file keeps speed from separating creative output from the product truth.
For example, an AI-generated ad that says a component “cuts maintenance by 30%” should not reach a customer because the wording sounded plausible. The workflow should retrieve an approved claim, preserve the supporting source, confirm the applicable product and operating conditions, and show that evidence to the approver.
That structure also protects creative momentum. Teams spend less time reconstructing how an asset was made, which release covered it, or where an unsupported number came from. Exceptions go to a named owner instead of stopping the entire operation.
What to do when evidence is incomplete
Do not ask the model to “make it safer.” Remove the unsupported claim, replace the restricted input, use a generic or human-created alternative, narrow the channels, or hold publication until the evidence is complete. The manual fallback is a simpler approved asset—not a rushed synthetic version with an undocumented origin.
If a rights or claim issue appears after publication, activate the incident-response process: pause distribution, preserve the evidence package, identify affected channels, notify the appropriate owners, correct or withdraw the asset, and approve any replacement through the normal gate.
The management question
Before releasing AI-generated creative, ask: Could we hand one evidence package to a customer, platform, partner, executive, or counsel and explain every meaningful input, claim, approval, and publication decision?
If the answer is yes, AI has accelerated a governed marketing process. If the answer is no, the asset is not finished.
