Where AI fits in agency work
| Work | AI suitability | Why |
|---|---|---|
| Concept visualisation and pitch boards | Excellent | Speed matters more than fidelity |
| Social cutdowns and variants | Excellent | Volume and iteration are the point |
| B-roll and texture | Good | Low risk, low scrutiny |
| Product hero footage | Conditional | Accuracy requirements dominate |
| Anything with a real person | Careful | Rights and disclosure obligations |
| Regulated sector advertising | Careful | Claims and substantiation risk |
Scope per shot, not per project: a project rarely divides cleanly into AI and non-AI. Decide shot by shot, because the constraint is usually one specific shot rather than the whole job.
Scoping honestly
- Say which shots will be AI-generated in the brief, before quoting.
- State the expected number of regeneration rounds, so revision expectations are set early.
- Flag anything the client may assume is shot footage.
- Be explicit about what AI will not do well on this brief — clients remember surprises far more than limitations stated upfront.
The failure mode is not a client rejecting AI. It is a client discovering AI after approving something they believed was shot. Disclosure at brief stage costs almost nothing; discovery costs the relationship.
Rights and disclosure in writing
| Item | Why it matters |
|---|---|
| Likeness and voice clearance | Synthetic use of a recognisable person needs specific written consent |
| Music and audio rights | Generated or licensed audio has its own terms |
| Disclosure obligations | Several markets require labelling; platforms may require it regardless |
| Claims substantiation | Generated copy can invent product claims you cannot support |
| Third-party asset licences | Reference images and stock inputs carry their own terms |
Get it in the contract: warranties that you hold necessary rights will be asked for. Being able to answer immediately is a commercial advantage.
Approval gates
-
Approve the concept before generating
Cheapest possible gate. Never generate to find out what the client wants.
-
Approve style frames before the full set
Agree look and treatment on one or two frames, then scale.
-
Approve at native resolution before upscaling
Never spend refinement budget on unapproved footage.
-
Approve the edit before final grade and audio
Structural changes are cheapest before finishing.
-
Approve disclosure wording before delivery
Avoids relabelling after release.
Revision economics
- Generation is cheap; re-briefing is not. Price revisions as scope, not as compute.
- Define what counts as a revision round in the contract — regeneration count is a poor proxy for effort.
- Keep seeds and prompts so an approved shot can be reproduced exactly rather than approximately.
- Track approval rate per client and per content type; it is your best predictor of margin.
How to price it
| Model | When it works | Risk |
|---|---|---|
| Fixed fee per deliverable | Repeatable content types | You absorb regeneration risk |
| Fee plus revision rounds | Most agency work | Needs clear round definition |
| Day rate | Exploratory work | Clients dislike open-ended AI iteration |
| Value-based | High-impact campaigns | Hardest to justify without benchmarks |
Do not price off generation cost: clients do not buy seconds of compute. Price the outcome, and treat cheap generation as margin rather than as a reason to discount.
What to deliver
- Final masters in the required specs, plus the project files.
- Prompts, seeds and reference images, so the work is reproducible.
- Provenance and consent records alongside the assets.
- Disclosure wording used, and where it was applied.
- An archive of approved intermediate versions, in case a later revision needs them.