How to Keep a Record of AI Work for Clients

A finished AI image looks the same whether it took one prompt or forty, whether it came from a free plan or a paid one, and whether you have the right to hand it over for a commercial campaign or not. The file itself does not carry that information. You have to keep it. This guide covers what to record for client work, why it matters more now than it did two years ago, and a simple method that takes a few minutes per project rather than a new piece of software. It is practical guidance from working experience, not legal advice; for contracts and compliance, talk to a lawyer.

What to record for every deliverable

If this looks like a lot, notice that the platform already holds most of it. The work is copying it somewhere you control before the history scrolls away or the account changes.

Why clients are asking now

Three things have changed the conversation.

• Disclosure requests. Agencies, publishers and larger brands increasingly ask suppliers to declare whether and how AI was used, sometimes in the contract and sometimes at delivery. A record lets you answer in a sentence instead of a scramble.
• Regulation. The EU AI Act includes transparency duties for AI-generated and manipulated content, and these are being phased in over the coming years. The details of who must label what, and how, are still settling. Whatever the final form, you cannot label what you cannot trace.
• Licences that differ by plan. Several platforms grant broader commercial rights on paid plans than on free ones, and some have changed their terms more than once. The question 'were we allowed to use this?' is only answerable if you know which plan you were on when you made it.

There is a quieter reason too. Clients come back. 'Can we have three more in the same style?' is an easy brief if you kept the prompt, model, seed and style references, and a slow reverse-engineering job if you did not. The practical side of this is covered in our Flux models explained guide, where licence terms vary within a single model family.

A folder and sheet method

You do not need a database. One folder structure and one spreadsheet per client do the job.

1. One folder per project, named with the client, project and date, for example acme-spring-campaign-2026-03.
2. Four subfolders inside it: 01-references for anything you uploaded or used as a style guide, 02-generations for raw outputs worth keeping, 03-edits for upscaled or retouched versions, and 04-delivered for exactly what the client received.
3. One sheet in the project folder with a row per kept generation: filename, date, platform, plan, model and version, prompt, key settings, references used, status (candidate, approved, delivered), and notes.
4. A terms snapshot. Save the platform's terms and pricing page as a PDF on the day you start the project and drop it in 01-references. Do it again if the project runs for months.
5. Fill the sheet as you go, not at the end. Thirty seconds after each keeper is painless; two hours of archaeology after delivery is not.

Name files so they sort and self-describe: 2026-03-14_kling-2_hero-v3_approved.mp4 tells you most of the story before you open the sheet. Where a platform lets you download generation metadata or a prompt history, export it into 02-generations as well.

Handling delivery and disclosure

At delivery, the client should receive three things: the final files, a short note on how they were made, and the rights position as you understand it. Keep the note plain. 'Images generated with [platform] on a [plan name] plan, [model and version], [date]; upscaled in [tool]; no real people or third-party brand assets used as references.' If the client wants an AI disclosure statement for their own compliance, this note is most of it.

Two habits protect you further. Keep the delivered files separate from the working files, so there is never confusion about which version went out. And keep the record for as long as the work is in use, not just until the invoice is paid; questions about licensing or disclosure tend to arrive long after the project closes.

Try it

Create a header row in a new spreadsheet, then paste this into the first data row the next time you generate something you might deliver.

filename | date | platform | plan | model + version | prompt | settings (ratio, seed, mode, duration) | references | status | notes
2026-03-14_hero-v3.png | 2026-03-14 | [platform] | [plan] | [model] | [full prompt] | 16:9, seed 4821, raw mode | ref-01.jpg (own photo) | approved | upscaled 2x, delivered 18 Mar

Adapt the columns to your tools: video work wants duration and camera-move fields, design work wants style and palette fields. The columns matter less than the habit of filling them in while the generation is still on screen.