AI Tools and Workflows · Procurement
Ten questions to ask before anything of yours goes through somebody else’s tool.
Every question below exists because a published position makes it necessary, and every one is answerable. A vendor who will not answer in writing has answered.
Last checked September 2026. Written for producers, line producers, heads of department and facility managers.
The ten
Ask all ten. Get the answers in an email, not in a meeting.
| Ask | Why it is the question |
|---|---|
| What is your model trained on, and will you put that in writing? | Of the products we have checked, only Adobe makes a positive licensed-training claim, and only for its own Firefly models. Almost everyone else says nothing. Silence is not compliance, and a compliance team will read it as a flag. |
| Do you train on our inputs or outputs, and is that the default or must we switch it off? | Netflix’s published guidance requires that tools “cannot store, reuse, or train on production data inputs or outputs”. Of the major video generators we checked, none makes that commitment on its public pages. Among the general assistants, only Anthropic’s published pricing states training is off by default on paid plans. |
| Where is our material stored, in which jurisdiction, and for how long? | Every upload of a script, a crew list, a medical form or a child licence is a processing decision under UK data protection law, not a productivity decision. |
| Will you sign a data processing agreement, and will you certify deletion? | Deletion and a certificate of deletion are contract terms. They do not happen by default and they cannot be added afterwards. |
| Who owns the output, and will you indemnify us against third-party IP claims? | Ownership and indemnity are separate questions and vendors answer them separately. Ask for both. |
| What security accreditation do you hold? | TPN and SOC 2 are the ones this industry recognises. Of the tools on our register, Filmustage is the only one that names both, which is exactly the credential a studio’s content-security team asks for before a script is uploaded. |
| Which third-party models does your product route to, and under whose licence? | Ask this of Adobe, Luma, Higgsfield and Autodesk in particular. All four now front other companies’ models, and a commercial-safety claim made for a vendor’s own model does not travel to the ones it resells. |
| Does the output carry Content Credentials or a watermark, and can they be stripped? | Google states it couples SynthID with interoperable C2PA Content Credentials on its current image model, and Adobe attaches Content Credentials to Firefly output. Most vendors say nothing. From August 2026 the EU’s transparency regime expects generated output to be detectable as generated. |
| What licence tier do we need, given our revenue? | Topaz, Stability AI and Midjourney all gate commercial use at a million dollars of annual revenue. A production company above that threshold on a standard plan is out of compliance, and nobody finds out until somebody asks. |
| What is the product’s status, and what notice would we get if you discontinued it? | OpenAI discontinued Sora. Its largest announced commercial partner is reported to have found out less than an hour before the public. |
In the contract
What to get written down, not agreed verbally.
A no-training warranty
From any vendor touching production material. This is Netflix’s second principle turned into a clause.
Clause
A scanning clause, supplied before the scanning day
Naming purpose, storage location, retention period, deletion trigger and the named data controller. Not handed over on the day, in a queue, with a crew waiting.
Clause
Consent scoped to the production
Not to the company and its affiliates in perpetuity. Separate consents for the scan, for the creation of a replica, for each use of it, for any training on the resulting data, and for any transfer to a third party.
Clause
Editing use separated from training use
Pact says scans are “most commonly for editing purposes”. If that is true, the contract can say so and forbid the rest.
Clause
Indemnity, deletion and certification
Indemnity for third-party IP claims arising from generated output, a deletion obligation, and a certificate confirming it happened.
Clause
Disclosure flowing both ways
The production tells the performer; the producer tells the commissioner. The BBC requires external producers to raise any AI use potentially needing disclosure with their commissioner.
Clause
For writers
No requirement to use generative AI; generated output is not literary material and not source material for credit; disclosure if supplied material was generated. That is Article 72, and it is the clearest language in the industry.
Clause
Three things that catch people
None of these is exotic. All three are in tools people already have.
The model you bought is not the model that made the pixels
Adobe’s generative panel inside Premiere now routes to Firefly, Veo, Kling, Runway and Luma. An editor generating an effect from the timeline may be reaching a model whose licence nobody on the production has read, under a commercial-safety claim that does not cover it.
The default setting is the exposure
Midjourney’s default is public: an art department generating unreleased designs on an entry plan is publishing them, and privacy is a paid tier. Consumer assistant tiers may train on what you paste. Neither is a technology problem and both cost a tier upgrade to fix.
Revenue thresholds bite quietly
Several licences that read as free or cheap change entirely above a million dollars of annual revenue, “regardless of the source of that revenue”. The person who signed up was not the person who knew the turnover.
Next
