Getting started
How a creative agency can pilot licensing its own operating records
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A creative agency can pilot licensing its own operating records by choosing one record family it clearly controls, such as brief versions, revision rounds and approval decisions, limiting it to clients whose contracts allow it, and excluding finished deliverables, client materials and third-party assets. A one-page brief should fix the scope, the rights basis for each account, masking rules, the utility test and stop conditions before any export.
License terms
Creative agency data pilot
Key takeaways
- Pilot the agency's process records, not the creative work, which contracts often assign to the client.
- Embedded stock, fonts, music and talent carry license terms that often bar AI use, so keep files out.
- Leave out any client account whose contract is silent rather than reading it generously.
- Mask identities but keep the order of versions, reviewer roles, decisions and the comments that caused them.
- Write stop conditions into the brief, including embargoed campaigns and legal claims reviews.
How can a creative agency pilot licensing its own operating records?
Pick one record family the agency clearly controls, such as brief versions, revision rounds and approval decisions from its project and proofing tools, limit it to a date range and a handful of clients whose contracts allow it, and leave the creative deliverables and any third-party assets out. Write that scope into a one-page pilot brief that also names the rights basis for each record family and the task the records could teach.
The pilot exists to answer two questions cheaply: can the agency show it has the right to license these records, and are they useful once client names and finished work are removed? A pilot that tries to include the portfolio answers neither, because ownership of the work itself is usually the hardest question an agency faces.
This guide sets out which agency records are usually the agency's own, why the deliverable normally stays out, a scope, rights and utility pilot brief you can copy, the approvals and conditions that change what may be delivered, an illustrative worked example, and the exception that should stop a pilot before it starts.
Which agency records are the agency's own, and which belong to clients or others?
Agency records fall into three layers: the agency's own process records, client materials the agency holds, and third-party assets licensed for a campaign. Only the first layer is a sensible pilot candidate, and even there client contracts can restrict use.
Process records describe how work moved: brief versions, scoping notes, task and status changes, reviewer comments, approval decisions, revision counts and turnaround times. They are created by the agency in its own systems. Client materials such as brand guidelines, product data, unreleased launch plans and customer research are usually confidential to the client. Third-party assets such as stock imagery, fonts, music and talent footage come with their own licenses.
| Record | Usual control | Pilot default | Why |
|---|---|---|---|
| Brief versions and scope changes | Agency, subject to client confidentiality | Include, with client names and products masked | Shows how requirements shifted and why |
| Revision rounds and reviewer comments | Agency, subject to client confidentiality | Include, with comments screened for client secrets | The core signal: what feedback changed the work |
| Approval and rejection decisions | Agency workflow records | Include, with reviewer roles instead of names | Shows who decided, at which stage, on what grounds |
| Final deliverables and comps | Often assigned to the client by contract | Exclude | Ownership and licensing of the work itself is a separate question |
| Client brand guidelines and research | Client | Exclude | Confidential client material, not agency records |
| Stock, fonts, music and talent assets | Third-party licensors | Exclude | License terms usually limit reuse and often bar AI use |
| Timesheets and resourcing | Agency, includes employee data | Exclude from a first pilot | Personal data and commercial terms add review without adding much workflow signal |
Why does the finished creative work usually stay out of the pilot?
Because the agency often does not own it, or owns it subject to promises that a licensing deal would break. Ownership of creative work turns on employment status and contract wording, and the answer can differ for each deliverable.
The U.S. Copyright Office's Circular 30 explains the two routes to a work made for hire. Work made by an employee within the scope of employment qualifies. Work by a freelancer or outside studio qualifies only if it falls into one of nine statutory categories, such as a contribution to a collective work or part of an audiovisual work, and the parties sign a written agreement that expressly calls it a work made for hire. Agency contracts commonly add assignment clauses on top, transferring rights in deliverables to the client on payment.
Third-party assets add another layer. Getty Images' content license agreement, for example, says that unless an invoice, sales order or license agreement explicitly authorizes it, licensed content, including captions, keywords and other metadata, may not be used for any machine learning or artificial intelligence purposes. A comp or final file that embeds licensed stock carries that restriction with it, which is one more reason to keep files out and process records in.
The broader law on training models with copyrighted works is still being worked out. The Copyright Office released a pre-publication version of its report on generative AI training in May 2025, covering the legal implications of training on copyrighted works and licensing. A pilot built on process records avoids staking the agency's first data deal on those open questions.
The scope, rights and utility pilot brief
Keep the brief to one page and fill every row before any record is exported. Each row names the evidence that supports it and the person who signs it off, so that a client services lead, the agency's counsel and a buyer can all read the same document. The rows below are the template.
| Brief field | What to write | Evidence to attach | Signed off by |
|---|---|---|---|
| Pilot question | The one thing the pilot must show, such as whether revision histories predict approval | Short rationale | Agency lead |
| Record family | One family, such as proof revisions and decisions | System name and export method | Operations or production lead |
| Date range and clients | Months covered and the client accounts included | List of accounts with contract dates | Client services lead |
| Rights basis | Why the agency may license these records for each client | Relevant contract clauses on confidentiality, data and work product | Counsel |
| Excluded material | Deliverables, client materials, third-party assets and personal data left out | Exclusion rules applied to the export | Production lead and counsel |
| Masking rules | How client names, products, people and campaign details are replaced | Masking dictionary and a before-and-after sample | Data or IT lead |
| Reviewer roles | The role labels used in place of names, such as account director or client approver | Role mapping table | Operations lead |
| Utility test | The task a buyer could build and how a sample will be judged | Three illustrative record chains | Agency lead |
| Stop conditions | Findings that pause the pilot, such as an embargoed campaign or a contract that bars reuse | Written list | Counsel |
| Exit and deletion | What happens to pilot copies if the pilot ends without a license | Retention note | Agency lead |
Which approvals, rights conditions or revisions change what may be delivered?
Any condition attached to a client, campaign or asset can remove a record from the pilot or change how it is masked, so check these before export rather than after a buyer has seen a sample.
- A client contract that defines confidential information broadly enough to cover the agency's notes about the client's work. That clause, not the masking, decides whether the records can be licensed.
- A campaign still under embargo. Records about unreleased products should stay out until the launch is public, and some clients will want them out permanently.
- Talent, music or stock usage terms that limit territory, media or duration. Process records that describe these assets in detail can reveal terms the agency agreed to keep private.
- A revision that went through legal or claims review. Comments from a regulated claims review can contain the client's legal reasoning and should be excluded unless the client agrees.
- An approval made outside the system, by email or in a meeting. The record shows a status change with no reason; keep it, but label the reason as unrecorded rather than inventing one.
- A change to the agency's own terms. FTC staff have warned that adopting more permissive data practices, such as using data for AI training, and disclosing the change only through a surreptitious, retroactive amendment to terms of service or a privacy policy may be unfair or deceptive. Amend client agreements openly and prospectively, with the client's agreement, if at all.
How do you keep revision and approval decisions useful after masking?
Keep the sequence and the reasons, and remove the identities. A revision history is useful because it shows a brief, a draft, a reviewer's comment, a change, and a decision, in order. Masking should replace names, brands and products with consistent placeholders so that the chain still reads correctly.
Check what your proofing tool actually stores before you plan the export. Adobe's documentation for Workfront's proof approval report describes fields such as the approver's decision, the decision date and the approver's stage, and notes that the report lists only the current version of each proof that needs approval. Workfront's proof review FAQ says a print summary lists all comments, markups and decisions made on a proof. A pilot that relied on the approval report alone would miss earlier versions; the version history and comment trail are what carry the revision signal.
Two more rules keep the records honest. Record each reviewer by role and stage, because a buyer learning approval behavior needs to know that a client approver, not an internal art director, rejected a draft. And keep rejections and the comments that caused them, because they show where the line was drawn far better than a run of approvals.
| Field to keep | Masking rule | Why it matters |
|---|---|---|
| Brief version and change note | Replace client and product names with stable placeholders | Shows how the requirement moved between rounds |
| Proof version number and date | Keep as is | Preserves the order of revisions and turnaround |
| Reviewer role and stage | Map names to role labels | Separates internal review from client approval |
| Decision and decision date | Keep as is | The label a model would learn from |
| Comment text | Screen for client secrets, people and claims language | Explains why the work changed |
Worked example (illustrative): a 60-person agency pilots its proof histories
This example is illustrative and not drawn from a real agency. It uses counts and durations only.
A brand and digital agency with about 60 staff runs its production work through a project and proofing tool. Its pilot question is whether revision histories show which kinds of feedback lead to client approval. The brief limits the pilot to 18 months of proof versions, comments and decisions for clients whose contracts allow anonymized use of the agency's own process records.
- Of 22 active client accounts, counsel clears 9. Seven contracts are silent, so those accounts are left out rather than read generously; 6 define confidential information to include all project communications.
- The export covers about 1,400 proofs with an average of 3 rounds each. Files and thumbnails are excluded; only version metadata, comments, roles and decisions are kept.
- Masking replaces client, brand and product names using a dictionary of 240 terms. A reviewer reads 50 random comment chains after masking and finds 4 that still name a product line, so the dictionary is extended and the check repeated.
- Exception: one account's comments include notes from the client's legal claims reviewer. That account is removed in full, even though the contract allowed anonymized use, because the comments carry the client's legal reasoning.
- Utility test: in three sample chains, the comment that preceded each client approval is clearly identifiable, which supports the pilot question. The agency records this as a qualitative finding, not a measured result.
What should stop the pilot before any record leaves the agency?
Stop when a record family's rights basis rests on silence in a contract, when the records cannot be separated from the work product, or when masking leaves the client recognizable. Each of those means the pilot would test the agency's legal risk rather than the value of its records.
Recognizability deserves a hard look in creative work. A comment that describes a distinctive campaign idea can identify the client to anyone in the industry, even after names are removed. NIST's report on de-identification notes that risks can remain in de-identified data and that data use agreements are an additional control, so plan for contractual limits on the recipient as well as masking.
Bring in specialists early. A creative operations lead can confirm which systems hold the full revision history, and a rights reviewer or counsel should read the client agreements and any third-party licenses before the brief is signed. Datasets that document how they were assembled and what was excluded are easier to review; Datasheets for Datasets is a widely used model for that documentation.
Turning a pilot into a licensed dataset
If the pilot shows the rights basis holds and the masked records still answer the pilot question, the brief becomes the first draft of a dataset description: scope, exclusions, masking, roles and known gaps. Widening the scope later is then a matter of adding clients and date ranges under the same rules.
An agency that wants to go further can license the selected process records through SourceX, which runs rights review, contracting, preparation, delivery and payment for that kind of deal. Its opening assessment concerns the agency as a company, so no proofs or comments are uploaded, and no record is delivered before the agency has approved the scope and terms and signed.
Frequently asked questions
What should a scope, rights and utility pilot brief include?
It should state the pilot question, the record family, the date range and client accounts, the rights basis for each account, what is excluded, the masking rules, reviewer role labels, the utility test, stop conditions and what happens to copies if the pilot ends. Each row should name its evidence and who signs it off.
Can an agency license the creative work it made for clients?
Often not without the client, because contracts commonly assign deliverables to the client or define them as works made for hire, and embedded stock, fonts, music or talent carry their own license terms. A first pilot is safer when it uses the agency's process records and leaves the work itself out.
Do client confidentiality clauses cover the agency's own project records?
They can. Some agreements define confidential information to include all project communications, which would capture comments and approval notes. Counsel should read each agreement; an account whose contract is silent is better left out of a pilot than read generously.
Is masking client names enough to make agency records shareable?
Not always. Campaign ideas, product details and industry context can identify a client after names are removed. Combine masking with a recognizability review on a random sample and contractual limits on the recipient.
Can an agency update its client terms to allow AI training on project records?
Only openly and with the client's agreement. FTC staff have warned that expanding data use, such as for AI training, through a quiet retroactive change to terms or privacy policies may be unfair or deceptive.
Sources
- The U.S. Copyright Office explains that an employee's work within the scope of employment is a work made for hire, while a commissioned work qualifies only in nine categories with a signed written agreement. Source
- Getty Images' content license agreement says licensed content, including captions, keywords and metadata, may not be used for machine learning or AI purposes unless explicitly authorized. Source
- Adobe's Workfront documentation describes the proof approval report's approver decision, decision date and stage fields, and says the report lists only the current version of each proof needing approval. Source
- Workfront Proof's review FAQ says a print summary lists all comments, markups and decisions made on a proof. Source
- FTC staff warned that expanding data use, such as for AI training, through a surreptitious retroactive change to terms of service or a privacy policy may be unfair or deceptive. Source
- The Copyright Office released a pre-publication version of its report on generative AI training in May 2025, covering legal implications of training on copyrighted works and licensing. Source
- NIST IR 8053 notes that risks can remain in de-identified data and describes data use agreements as an additional control. Source
- Datasheets for Datasets proposes documenting a dataset's motivation, composition, collection process and recommended uses. Source
See if your company qualifies
A short company assessment. No data uploads are needed.