Where an AI creative tool fits in a larger SOP

Most teams do not need a new process for AI. They need one extra step in the process they already have, plus a few extra things to write down. This page gives you a template for that. I am not repeating the basics of how to write an SOP: for that, read how to write an SOP and free SOP examples for a small business. What you get here is a content production SOP you can fill in, with the AI step already placed.

It suits a small business or agency of about 3 to 30 people that makes images and short videos for its own marketing or for clients.

This page may contain affiliate links from which Blur Studio earns commissions.

What this SOP covers

The SOP covers one flow: from a request for a piece of content to a published, filed asset. It names the owner of each step, says what the step takes in and puts out, and marks where AI is used. It does not cover which AI tool to buy. If you are still choosing, see all-in-one AI creative platform versus separate tools.

Keep the whole document to one or two pages. A long SOP gets skimmed once and ignored, which is the same lesson as in a short contract SOP your team can follow.

The SOP with owners (fill in the brackets)

#	Step	Owner	In	Out	Time limit
1	Request and brief	[Account lead]	Client or internal request	One-page brief: goal, audience, channel, size, must-show items, banned items	[1 day]
2	Rights and reference check	[Producer]	Brief, reference files	List of allowed references, written consent for any real person, brand files	[1 day]
3	AI generation step	[Creator]	Brief, approved references, saved prompt	[Number] drafts (for example 6), plus a reject log	[Half a day]
4	Select and edit	[Creator]	Drafts	1 to 2 finalists, edited and sized for the channel	[1 day]
5	Review gate (by risk tier)	[Creative lead, then client if needed]	Finalists	Approved or changes marked as blocker or suggestion	[2 days, 2 rounds at most]
6	Label and finalise	[Producer]	Approved file	File with the platform label set, metadata kept, captions added	[Half a day]
7	Publish or hand over	[Account lead]	Final file	Posted piece or delivered files	[As scheduled]
8	File and record	[Producer]	Final file, log	Asset stored in the shared library with the record fields below	[Same day]

One named person owns each row. If a row has two owners, it has none.

Where the AI step sits, and how to write it

Step 3 is the only AI step. Write it as a small box with an input and an output, like any other step. A good version reads:

• Input: the approved brief, only the references cleared in step 2, and the saved prompt from the team's prompt library if one exists.
• Action: generate up to [N] drafts in the approved tool, within the credit budget for this job. Credits are the points an AI tool charges for each image or clip.
• Output: the drafts, plus a reject log: for each draft thrown away, one line on why (wrong product shape, odd hands, unreadable text, off-brand colour).
• Stop rule: if there is no usable draft after [3] rounds, stop and use the fallback in the section below.

The reject log matters more than it looks. After a month it shows which prompts and settings waste credits and which kinds of jobs the tool is bad at. It is also the evidence you need if a client asks how a result was made.

Name the approved tools in the SOP, with the plan each one is on. For example, if the team uses Higgsfield (an AI image and video platform that offers several models in one account), write down the plan and who holds the login. A rule on which tools are allowed, and what ages, faces and content are off limits, belongs in a separate policy. See an AI creative tool policy for teams, and point step 3 at it.

Review gates by risk

Not every asset needs the same checking. If everything needs five sign-offs, people will skip them. Match the gate to the risk, using three tiers you can paste into the SOP.

Tier	Examples	Who must approve	What they check
1. Internal	Mood boards, concept sketches, internal pitches	[Creative lead] only	Is it clearly marked as a draft? No client data used in the tool?
2. Public, organic	Social posts, blog images, short clips	[Creative lead] plus [Brand owner]	Brand look, accuracy of anything shown, label set, nothing that looks like a real person without consent
3. Paid, client or sensitive	Paid ads, client deliverables, product shots buyers rely on, anything with a real person, regulated claims	[Creative lead], [Brand owner], [Rights check], then [Client]	Brand, rights and consent on file, claims and product accuracy, ad platform rules, label

A two-gate model is common in agency tools: an internal gate first, then a client gate. Vendor guides on approval workflows describe the same idea, with extra checks added only when the content is public or paid. The SOP above does the same in plain words.

Three rules keep review from dragging on:

1. Mark each comment as a blocker or a suggestion. A blocker must be fixed. A suggestion can be ignored by the creator, who explains why in one line.
2. Cap rounds at two. A third round needs the account lead to agree on a new scope. See how to get written approval when a client changes the scope.
3. One person has the final say. Name them. Without that, review turns into a vote.

If your clients approve work through a portal or by email, my guides to a creative client approval workflow and a content approval process go deeper on the client gate.

Fields to record for each AI asset

A normal asset record has a title, owner, date and file path. AI assets need a few more fields, because later you may need to prove where something came from, rerun it, or change it when a model closes. Copy this list into your asset library or spreadsheet:

1. Asset ID and job name.
2. Tool, plan and model with version. Models change names and some close. OpenAI's Sora video app, for instance, ended on 26 April 2026, with its API ending on 24 September 2026.
3. Prompt saved (the exact text, plus settings and seed or reference IDs if the tool shows them).
4. Reference files used, and where each came from.
5. Rights and consent check: who checked, on what date, and a link to the signed consent if a real person appears.
6. Risk tier and who approved it, with dates.
7. Label set: yes or no, per platform. Check the platform labelling rules and ad platform policy traps.
8. Provenance data kept: some tools add content credentials (C2PA, a standard way of marking where a file came from). Do not strip them. See why you should not remove C2PA marks.
9. Credits or cost used for the job, so you can work out cost per approved asset.
10. File location in the shared library. A tool's own history is not a storage plan. See building an AI asset library.

Keep client confidential material out of the tool unless you have checked its data terms. Some services may use what you upload to train their models unless you are on a business agreement that says otherwise. Higgsfield's terms, for instance, say it may train on content except under enterprise agreements, so an agency on a team plan cannot promise a client that nothing it uploads will ever train a model.

Fallbacks: when the tool is down or the credits are gone

Add one line to the SOP for each of these cases, with an owner. Teams usually forget this until a deadline is at risk.

• The tool is down or slow: [Creator] tells [Account lead] within [1 hour]. Switch to the second approved tool or to stock footage, and note it in the record.
• Credits are used up: [Account lead] decides between a small top-up, waiting for the next billing cycle, or the no-AI route. Many tools let subscription credits expire at the end of the month, so unused credits are not a saving. Set a monthly credit limit per job so one project cannot drain the account.
• The result keeps failing: after [3] rounds, switch to a phone shoot, stock or a freelancer. The cheapest AI draft is the one you stop making. For a comparison, see AI video versus stock footage versus filming yourself.
• A model is removed or changed: [Producer] flags affected saved prompts and re-tests the top ones before the next job.

Where to keep it: a doc, Notion or Trainual

An SOP only works if people can find it and you can tell they read it. You have three realistic places to keep this one.

A shared document or a Notion page. Google Docs is a free word processor you share by link. Notion is a workspace tool for notes and wikis. For a team under about 10 people with no need to prove training happened, either one is enough. Give the SOP an owner, a review date and a version note. My comparison Trainual versus Notion for SOPs and the list of free SOP software cover the options, how to choose SOP software explains what to look for, and how to manage SOPs for a small business covers keeping them current.

Trainual. Trainual is software for writing a team's procedures and tracking who has read and completed them. It makes sense when you onboard people regularly, work with freelancers, or need to show a client that your team was trained on the AI rules. You can turn this SOP into a short course with a quiz, assign it to new hires, and see who has finished. Trainual does not publish its prices; its pricing page says there is a one-time $1,000 implementation fee, so ask for a quote and compare that with the free options first. My sign-up link offers a 7-day trial; confirm the terms on the sign-up page before relying on that. For setup, see how to set up Trainual for a 15-person team, or free alternatives to Trainual if the fee is too much.

Your approval tool. Some tools hold their own approval steps. Canva (a design tool) lists approvals only on its Business plan, for example. That is useful for the review gate, but it is not a replacement for the written SOP.

Whatever you choose, write the review date in the SOP. I would re-read it every quarter, and after any big model or platform-rule change.

Test the SOP before you rely on it

Run one real job through every row, and mark each row pass or fail.

• Pass: every row had an owner who knew it was theirs; the reviewer used the right tier; the record is complete; someone not on the project could find the file and re-run the prompt.
• Fail: a step was skipped, an owner was unclear, a comment was not marked blocker or suggestion, or a field was left empty.

Fix the SOP, not the person, then run one more job.

Who should skip this

If you make a couple of AI images a month for your own social posts, a full SOP is too much. Use steps 1, 3, 6 and 8 and a note of the prompt. Add the rest when you have more than one person making content, work for clients, or run paid ads.

How I'd approach this

I would start with the table, fill in the brackets with real names this week, and run one live job through it. Keep the AI step small and the record fields few. The review tiers are the part that saves the most time, because they stop low-risk drafts waiting for a client who does not need to see them.

If your team is bigger or the volume is high, a document stops being enough. Blur Studio builds the repeatable workflow, approval portal or shared asset library around these tools, so the SOP is enforced by the system and not only by memory.

This page is general information, not legal advice. Check local law, your contracts and each platform's current rules.