Why handoffs fail even when the contract is fine
Ownership and delivery are different problems. A licensing agreement settles who may use an image, for how long, in which territories. It rarely specifies the technical state of the files that arrive — and that is the part that determines whether the brand can still use the asset in two years.
Four failure modes account for most of it:
| Failure | What the brand receives | What surfaces later |
|---|---|---|
| Derivative-only delivery | Flattened, resized JPEGs | No re-crop, no colour correction, no new sizes |
| Naming drift | IMG_4471_final_v3_APPROVED.jpg | Cannot map files to SKUs at ingest |
| Metadata stripped | Clean pixels, empty fields | No creator, no rights, no AI provenance |
| Single-channel transfer | A WeTransfer link | Expires in 7 days; becomes the only copy |
None of these are visible at the moment of delivery. The images look right, the invoice gets paid, and the defect is discovered at the exact moment the asset is needed most urgently. The cost is not the original shoot fee — it is the reshoot, plus the delay to whatever launch depended on it.
Masters, derivatives, and the difference that matters
Every handoff should distinguish three tiers of file, and the agreement should name all three explicitly.
Masters are the highest-fidelity, editable state: raw captures, layered PSD or TIFF files with retouching intact, and any cutout paths or alpha channels. These are what make future work cheap. Approved renditions are the flattened, colour-managed, full-resolution finals — the source of truth for anything published. Derivatives are the channel-specific exports: the 2048px square for the marketplace, the 1200px for the product page, the cropped 4:5 for paid social.
Accepting derivatives as the deliverable. If the brand only receives channel exports, every future channel requires going back to the agency — or reshooting. Derivatives are disposable and should be regenerable at any time from the approved rendition.
The practical rule: derivatives should be treated as cache, not as assets. A brand that can rebuild every derivative from its masters in an afternoon has a working catalog. A brand whose derivatives are its archive has a liability, and the liability compounds with every new channel it adds.
Raw files deserve their own line in the agreement. Many agencies retain raws by default as a matter of craft — the position is defensible, but it must be a decision, not a discovery. If raws are not transferring, the brand should know that before the shoot, not when it wants a different crop.
The metadata that has to survive the transfer
Metadata is the part of the handoff that everyone assumes is someone else's job. It is also the cheapest thing to get right and the most expensive to reconstruct.
Four fields carry nearly all the value:
- Identifier — the SKU or style code, embedded in the file, not just in the filename. Filenames get changed by every system they pass through; embedded IPTC fields survive far more of the journey.
- Creator and copyright — who shot it and under what terms. This is what makes a takedown or a licensing question answerable years later.
- Colour profile — an embedded ICC profile, not an assumption. Untagged files get interpreted differently by every downstream pipeline, which is where "the product looks wrong online" usually begins.
- Provenance — whether AI tools were involved in creating or editing the image, and which.
That last field stopped being optional in the EU. The AI Act's Article 50 transparency obligations, applying from 2 August 2026, require AI-generated or AI-manipulated content to carry machine-readable markers identifying it as such. The C2PA Content Credentials specification is the standard most tooling has converged on — a cryptographically signed manifest embedded in the file rather than a visible watermark.
Storefront and marketplace pipelines re-encode uploads, and embedded manifests frequently do not survive. Embedding provenance in the file is necessary but not sufficient — the handoff should also deliver a sidecar record (a CSV or manifest) stating which assets involved AI, so the brand can answer the question without depending on bytes that a CDN may have discarded.
Write the deliverable manifest before the shoot
The fix for almost everything above is a single document agreed before any camera comes out: a deliverable manifest that states exactly what will be handed over. It takes about an hour to write and it is the highest-leverage hour in the whole engagement.
A workable manifest covers eight things:
- Shot list mapped to SKUs — every intended image tied to an identifier, with the required angles per SKU.
- File tiers — which of masters, raws, approved renditions and derivatives transfer, stated individually.
- Formats and specs — file type, bit depth, colour space, minimum long edge for each tier.
- Naming convention — the exact pattern, with an example, and who is responsible for applying it.
- Embedded metadata — the required IPTC fields and provenance markers.
- Transfer mechanism — where files land, and who holds durable access.
- Rights summary — a plain-language line per asset class, cross-referenced to the contract.
- Acceptance criteria — what makes the delivery complete, and the window for raising defects.
The acceptance criteria matter more than they look. Without them, "delivered" is whenever the agency says so, and the brand's leverage evaporates the moment the final invoice clears. With them, there is a defined window in which a missing master is the agency's problem rather than the brand's.
Run the ingest check while you still have leverage
A handoff is not complete when files arrive. It is complete when the brand has verified them — and that verification has to happen before the engagement closes.
The check is mechanical and mostly automatable:
Run six assertions against the delivered set: every SKU on the shot list has its required angles; every file matches the naming pattern; every file carries an embedded colour profile; required IPTC fields are populated; master files are genuinely layered rather than flattened files with a .psd extension; and the AI-provenance record covers every asset that needs one.
Most of this is a short script over the delivered folder, and it is worth writing once and reusing on every engagement. The failure it prevents — discovering in month nine that a third of the masters were never delivered — is not recoverable by any amount of goodwill.
This ingest gate is the same technical check that belongs in an internal product image review and approval workflow; the only difference is that here it also functions as the acceptance test for a contract.
Offboarding: the handoff nobody schedules
The riskiest transfer is the one that happens when a relationship ends — because by then nobody is being paid to make it go well.
Three things reliably go missing at offboarding. The first is anything held in the agency's own tools: retouching presets, colour recipes, camera profiles and lighting diagrams that make a catalog look consistent. These are rarely named in contracts and are the reason a new agency's work looks subtly different from the old one's.
The second is access. Shared drives, DAM seats and asset libraries are usually provisioned under the agency's account. When that account closes, the brand can lose access to files it owns outright.
The third is the mapping. An agency's internal job numbers often connect files to SKUs, seasons and shoots. Without that mapping exported, the brand inherits a folder of correctly named files and no history.
Unplanned offboarding
- Final invoice, then silence
- Assets scattered across expired links
- Presets and recipes lost
- New agency starts from zero
- Consistency breaks mid-catalog
Scheduled offboarding
- Transfer session on the calendar
- Full archive to brand-owned storage
- Presets and specs documented
- SKU mapping exported
- Visual continuity preserved
The remedy is to write offboarding into the engagement at the start: a named final transfer, a verification pass against the manifest, and a documented specification of the look so it can be reproduced by someone else.
What changes when part of the catalog is AI-generated
AI-assisted imagery has made handoffs both simpler and more demanding. Simpler, because much of what used to require a reshoot — a background swap, a wider crop, a new colourway — can now be produced from an existing approved rendition, which reduces the brand's dependence on any single agency. More demanding, because the provenance record now has to be maintained and transferred alongside the pixels.
Practically, this shifts what the brand should insist on receiving. A clean, high-resolution, correctly profiled master with an accurate cutout path is worth considerably more than it used to be, because it is now the input to a great deal of downstream generation. The decision about whether a flawed image needs a fresh capture at all has moved too, which is worth thinking through separately when you choose between reshooting and retouching.
First: a per-asset flag for whether AI tools were used in creation or editing, delivered as a sidecar file as well as embedded metadata. Second: a statement of which generation or editing tool produced each AI-involved asset. Both are trivial to record at production time and effectively impossible to reconstruct afterwards.
Tools like Retouchable sit on the derivative side of this line: given a good master, they regenerate backgrounds, crops and channel exports on demand, which is precisely why the master tier is the part of the handoff worth fighting for. The rights position does not change because AI was involved — but it does get harder to answer without records, which is covered in more depth in our guide to product photography usage rights.