Adobe spent last week in Miami Beach telling the creative world that its tools are now open to outside AI agents. At Adobe MAX 2026, which ran from 10 to 12 November, the company previewed local MCP integrations. In plain terms, third-party agents will soon be able to edit native Photoshop, Illustrator, Premiere and After Effects files. If your brand lives in those files, and most brands do, it is time to write AI brand guidelines before the agents arrive.
This is not a panic piece. Agents will do a great deal of dull production work well, and I am glad of it. However, the news exposes a gap that most brand guidelines were never built to cover. Below, I explain what changed, why the typical guidelines PDF fails the AI test, and what an AI-ready brand system needs to contain. There is also a practical section on how we approach the work at SOHO.
What actually changed at Adobe MAX 2026?
Adobe’s announcement on 10 November covers a lot of ground, from Creative Agent across Creative Cloud to Firefly rebuilt as an AI creative studio. The item that matters most for brand owners is quieter. Local MCP integrations, due in private beta in early 2027 according to Creative Bloq, will let the agents your team already uses edit real project files rather than flattened exports.
Two reactions from the week are worth reading together. Creative Bloq’s Ian Dean summed up the mood in one line: “How much control you keep is the big question”. Then, on 13 November, CMSWire’s Greg Kihlstrom argued that MAX hands marketing teams a governance job. His point is simple: an agency can now point its own agent at a client’s native files, and Adobe’s example workflow runs no brand check. He also cited Adobe-commissioned research in which senior leaders estimated that 9 per cent of public-facing content was off-brand. Practitioners put the figure at 27 per cent. In other words, the people closest to the work already saw three times more drift than their leaders did.
Why do most brand guidelines fail the AI test?
Most guidelines documents address people with taste, and that is the problem AI brand guidelines have to solve. They show a logo, a palette, three typefaces and a page of adjectives, and then they rely on a designer to interpret the rest. That interpretation is the job. It is also the one thing no agent can inherit from a PDF.
An agent reads instructions literally. Ask it to keep the layout “confident and calm” and it will do something, but not necessarily your something. Instead, tell it that clear space equals twice the x-height of the wordmark, and that the mark never sits on a photograph without a 40 per cent scrim. Now it can comply, and you can check that it did. That gap is the gap between brand guidelines for AI tools and brand guidelines for a studio wall.
There is a second failure, and it is the one Kihlstrom is pointing at. Guidelines describe the finished look and say nothing about the route to it. Who may generate, who may edit, who approves, and what happens when a file leaves your Creative Cloud for someone else’s agent? Those are governance questions, and they tend to be nobody’s job until something goes wrong in public.

What should AI brand guidelines contain?
Start with what you already have, then add five things. None of them needs new software, and all of them make the brand easier for humans to use as well.
1. Tokens, not just swatches
Colours, type scales, spacing and corner radii should exist as named values in a file a machine can read. With a design system, this already exists. Without one, our explainer on what a design system is is the right place to begin.
2. Hard rules written as rules
Clear space, minimum sizes, forbidden combinations, image treatment and motion limits, one instruction per line. Each rule should be testable with a yes or a no. If you cannot imagine checking it, it is a preference, not a rule.
3. Voice with examples
Three to five pairs of “we say this, not that”, drawn from real copy that shipped. Adjectives alone are useless to an agent and nearly useless to a new freelancer, so give both of them sentences to copy.
4. Permissions and routes
Decide which tools and agents may touch native files, which are limited to exports, and who owns each route. Kihlstrom’s advice to decide in writing which source of brand rules wins is sound, and we have adopted it.
5. Disclosure and provenance
Say where AI-made or AI-edited assets must be labelled, and how content credentials survive the production chain. We covered the public-facing side in our piece on AI disclosure in marketing.
How a small senior agency does this in practice
At SOHO we do not sell a separate AI brand guidelines product. The useful work is the same work a good brand system always needed, done with more precision. Here is how we run it.
We begin by auditing the brand as it is actually used, not as the PDF describes it. Pull the last fifty things that went out, from LinkedIn tiles to investor decks, and mark what drifted. The drift tells you which rules were never explicit in the first place.
Then we ask the client three questions. Which tools will your team and your partners realistically use next year? Next, which assets are sacred, meaning no automated change ships without a human looking at it? And who signs off when the agency and the client disagree about an agent’s output?
Finally, we write the rules into the design system rather than beside it. Tokens live in Figma and in code, and the hard rules sit next to the components they govern. Koto’s Giffgaff refresh, covered by Creative Boom on 11 November, built the new digital system on the brand’s existing design tokens, which is exactly the right instinct. Our identity work for deep-tech clients such as Simudyne and Helsing has to survive engineers, investors and partners using it without us in the room. An agent is simply the newest colleague in that room.
Three mistakes to avoid: writing a hundred rules when fifteen matter, banning AI outright (people will use it anyway, just quietly), and treating the document as finished. Review it every quarter, because the tools will change faster than the brand.
Who owns the brand when the agent does the work?
My answer is unchanged by last week. The client owns the decisions and the agency holds the system. What changes is that both parties now need that system to be legible to software, not just to people. So, if your agency keeps the master files, ask which agents they intend to run against those files once MCP integrations open up. Then ask whether a brand check sits anywhere in that workflow. A shrug is your first governance gap.
For the same reason, I would not let any partner’s agent loose on native files until your AI brand guidelines exist. Exports first, native files later, and only once you have seen the output with your own eyes.
Where to start this quarter
You do not need to wait for the beta. Three steps before the end of the year will give you a first draft of AI brand guidelines, and each is a few days of work rather than a project.
- Audit the last quarter’s output for drift and list the rules that would have prevented it.
- Turn palette, type and spacing into named tokens, even in a spreadsheet, if a design system does not exist yet.
- Write the permissions page: tools, routes, owners, and the one person who decides.
Talk to us about an AI-ready brand system
If your identity was built before agents could open your files, it is worth an hour’s conversation. Our visual identity work includes the documentation an agent can follow and a person can check. You can start that conversation here.
Frequently asked questions
Do we need AI brand guidelines if we do not use AI?
Almost certainly, because your partners and platforms do. Agents now sit inside the tools your agency, your printers and your social team use every day, so the brand will meet them whether or not you choose to.
Are AI brand guidelines the same as a design system?
They overlap but are not identical. A design system gives you tokens and components. AI brand guidelines add the hard rules, permissions and disclosure policy that tell any tool, human or otherwise, how far it may go.
Should we let an agency’s agent edit our native files?
Not yet. Start with exports, look at the output, and only open native files once the rules and the approval route are written down and someone owns them.
