AI Content Tools: The Security and Data Questions to Ask

A head of brand at a mid-size company told me their procurement review of an AI video tool took eleven weeks and produced a single actionable finding. Eleven weeks. The finding was that the vendor's uploaded assets were retained indefinitely by default. Useful, but it should have taken one email.

The reason these reviews drag is that most security questionnaires were written for SaaS that stores records, not for tools that take your product photos, your executive's face and your voice, run them through third-party model providers, and hand back a file you'll put on your brand's channels. The risk surface is different, so the questions should be different too.

Here's the list I'd actually send, why each question matters, and what a reasonable answer sounds like. Nothing here requires a legal team to run — a marketing lead can send it and read the replies.

Question 1: Is my content used to train models?

This is the one everyone asks and the one most often answered imprecisely. Split it into three:

• Does the platform train on my uploads or outputs?
• Do the underlying model providers the platform routes to train on inputs passed through the API?
• Does the answer change by plan tier?

That third part catches people out. It's common for consumer or free tiers to have different data terms than paid business tiers, which means a team trialing a tool on free credits may be operating under terms they wouldn't accept in production. Ask for the answer in writing, per tier.

A reasonable answer: "No training on customer content on paid plans; provider-side terms are X, Y, Z for the models we route to." An unreasonable answer: "We take privacy seriously."

Question 2: What's retained, for how long, and can I delete it?

Generation platforms hold more than you think: uploaded reference images, generated outputs, prompts, and often intermediate frames. Each is a separate retention question.

Ask specifically:

• Default retention for uploads vs. outputs vs. prompts
• Whether deletion is real deletion or a soft flag
• Backup retention window after deletion
• Whether deleting a workspace deletes member-created assets

The practical risk is rarely a breach. It's an unreviewed product shot or an unapproved executive likeness sitting in a library for two years and surfacing in a share link. Retention hygiene is mostly about reducing that surface.

Question 3: Who on my team can see and do what?

Access control questions for content tools are unusually concrete because the actions are unusually consequential.

Capability	Why it needs a control	Reasonable default
Publish to connected social accounts	A mistake is public and immediate	Restricted to a named few
Upload voice or face references	Consent and likeness exposure	Restricted, logged
Create share links	Silent external distribution	Expiring links
Spend credits	Budget	Visible balance, spend history
Disconnect social accounts	Breaks scheduled posting	Admin only

The one people forget is the fifth row. A departing contractor who disconnects an account can quietly break a week of scheduled publishing.

Question 4: What are the commercial rights on the output?

This is where content tools diverge sharply from other software, and where the answer has real money attached.

Three sub-questions:

1. Is commercial use permitted, and on which plans? Many platforms permit it only on paid tiers. Versely, for example, allows commercial use on paid plans and doesn't watermark output — but you should confirm the equivalent for any vendor rather than assume it.
2. Does it vary by model? It can. A platform routing to a dozen model providers inherits a dozen sets of terms. Ask whether the platform normalizes them or passes them through.
3. Who owns the output? And can the vendor use your generations in their marketing without asking?

Get all three in writing before a campaign, not after. Re-generating a finished ad because of a licensing surprise is the most avoidable expense in this category.

Question 5: Which subprocessors touch my data?

A modern AI content platform is a router. Your prompt and your reference image travel to whichever model provider serves that model. That's not a flaw — it's how you get access to 60+ video models and 100+ image models in one place — but it means the subprocessor list is the real data map.

Ask for the list, and ask two follow-ups: does the platform tell you which provider serves a given model, and can you restrict generation to a subset if your policy requires it. If a vendor can't tell you which provider ran your job, that's a meaningful gap.

Also worth asking: where does generated media get stored and served from, and are asset URLs public, signed, or expiring? Publicly guessable asset URLs are a quiet leak vector for unreleased product imagery.

Question 6: What does it actually cost, and can I see the spend?

Security reviews rarely cover pricing, and they should — unpredictable spend is an operational risk, and opaque billing makes it impossible to attribute cost to a campaign.

The questions that matter for a pricing guide conversation with any AI content vendor:

• Is billing per seat, per generation, or credit-based? Credit-based models scale with usage rather than headcount, which suits small teams making a lot of content and penalizes teams making a little.
• Can I see a per-generation spend history, and is it attributable to a person or a campaign?
• What happens on a failed generation — am I charged?
• Are there free daily credits for evaluation, and do the data terms differ on them?
• Does price vary by model? It should, because model cost varies enormously. A fast tier and a premium tier costing the same is a sign the pricing is averaging over you.

Versely bills in credits with free daily credits and a visible spend history; the current plan structure is on the pricing page, and we broke down how to budget against credits in credits explained.

Question 7: What happens to my consent obligations?

Two capabilities carry obligations no vendor can absorb for you: voice cloning and likeness/avatars.

If you clone a narrator's voice or build a digital twin of a founder, you need documented, specific, revocable consent — covering the uses, the duration, and what happens when they leave. The platform can store the voice; it can't store your permission to use it.

Related, and increasingly non-optional: disclosure. Rules on labeling synthetic media in advertising vary by market and platform, and they've been tightening. We covered the practical version in AI ad disclosure compliance. Treat disclosure as a creative constraint you design for, not a legal footnote.

A one-page evaluation you can actually finish

If you do nothing else, send this five-line email:

1. Do you train on customer content on paid plans? Do your model providers?
2. What's the default retention for uploads, outputs and prompts, and is deletion permanent?
3. Which subprocessors receive our data, and can we see which serves each model?
4. Are outputs cleared for commercial use on our plan, watermark-free, and does that vary by model?
5. Is spend visible per generation and attributable per user?

Five answers gets you 90% of the risk picture. If the replies are vague on any of them, that vagueness is the finding.

FAQ

Should I avoid AI content tools that route to multiple model providers?

No — routing is what gives you model choice, and single-provider tools just hide the same dependency behind one name. What matters is disclosure: the vendor should tell you which provider serves each model and list subprocessors. Multi-provider with transparency beats single-provider with silence.

Is free-tier usage safe for evaluating with real brand assets?

Often not, and this is the most common quiet mistake. Data terms and commercial-use rights frequently differ between free and paid tiers. Evaluate with non-sensitive stand-in assets, or confirm the free-tier terms in writing before uploading anything real.

Do I need a written consent form for voice cloning a colleague?

Yes. Get specific, documented consent covering permitted uses, duration, and revocation — including what happens if they leave the company. Platform-side controls can restrict who uploads a voice, but the permission itself is your responsibility.

How do I control spend across a team using a credit-based tool?

Set a per-campaign generation cap before work starts, keep exploration on fast model tiers and finish on premium ones, and review the spend history weekly. Credit-based pricing is easier to control than seat pricing precisely because you can see where it went.

What's the single most overlooked question in these reviews?

Asset URL access. Teams scrutinize training and retention, then share a generated unreleased product video via a link that never expires. Ask whether media URLs are signed and expiring, and set share links to expire by default.

Run the five-line email against whatever you're evaluating, then pilot on one format rather than the whole calendar — the operational guidance in integrating AI content into your existing martech pairs well with the answers you'll get back.