When I read about Adobe’s new Firefly AI Assistant, my first thought was simple: this is super cool. My second thought was a question the tool cannot answer for me. Whose standard is it working to?
We build an AI marketer, so we have a stake in this.
Why this excites me
I am not a designer, but I think in pictures. For years I struggled to tell designers what I could see in my head. I had neither the language nor the skills to say it well.
The best I could do was sketch bad drawings or collect reference images. Those only ever caught a shadow of what I imagined. It was frustrating.
My job was also to protect the brand and the business. Sometimes the designs that came back were great and still needed fine tuning. Explaining that fine tuning was as hard as writing the first brief.
What Adobe is building
On 15 April 2026, Adobe described the Firefly AI Assistant. You state the outcome you want in your own words, and the assistant runs multi-step workflows across apps such as Photoshop, Illustrator and Premiere.
Adobe says creators provide the vision, judgement and direction while the assistant handles the execution.
Several pieces matter for brand work. Adobe lists pre-built Creative Skills, including generating content across social channels. It says the assistant can learn a creator’s preferences over time, and that it understands brand assets.
It can also read reviewer feedback in Frame.io and apply the changes, which Adobe says shortens the loop from review to final.
Five days later, at Adobe Summit, Adobe announced Brand Intelligence. Adobe describes it as an engine that learns from review feedback, annotations, rejections and approvals, instead of relying only on static brand guidelines.
For someone like me, this is an amplified way to talk with designers, human and AI. I can describe, react and refine in the same conversation.
The question underneath: whose standard?
A faster conversation only helps if everyone is talking about the same brand. If a designer, a marketer and an AI agent each hold a slightly different picture, a better channel just spreads the disagreement faster.
So whose opinion would actually fit the brand? Which standards are we following?
A tool that learns one creator’s preferences is a good fit for that creator. A brand is different, because a team makes it. Analyst coverage of Adobe MAX by diginomica, published on 11 November, notes that much brand knowledge is tacit and not written down.
That gap has a cost. An Optimizely and Savanta survey of over 2,000 marketing leaders found that one in four knowingly publish off-brand AI content. The respondents came from seven countries.
The article reporting it, from July 2026, names neither the countries nor the survey date.
Skills grow, and knowledge has to centralise
This is my view, not a finding. As we keep working with AI, we will build more skills and find better ways to use these systems. That means moving away from knowledge scattered across people, files and chat threads, and towards one shared place.
Learning from rejections and approvals is a sensible design. It also depends on reviewers who give consistent reasons.
If one reviewer rejects a draft as “off tone” and another approves almost the same draft, the signal is noisy. That is our opinion, and it applies to any tool that learns from review.
Turn your rejections into rules this week
You do not need any software for this. You need an afternoon and your recent review history.
- Collect the last 20 drafts that someone rejected or heavily rewrote.
- Beside each, write the reason in one plain sentence.
- Group the reasons that repeat. Three or four groups usually cover most of them.
- Turn each group into a rule, with one example that passes and one that fails.
- Mark each rule as “never” or “usually”. Judgement calls are fine, as long as they are labelled.
- Give the rules to someone new and ask them to review five drafts. Compare their verdicts with yours.
Where the newcomer disagrees with you, the rule is still too vague. Rewrite it until a stranger can apply it. Keep the result to one page and start with the ten rules that cause most of your rejections.
How Mavic approaches it
Mavic exists to find the balance in this. We know a brand is a mix of business, visual identity, emotion, experience and history. We know a team makes it, so it cannot live in one head.
We also know AI agents need complex, constantly evolving information to understand what is on brand. With Mavic, we bring these considerations into a single workspace.
In practice, Mavic reads your Brand DNA each time it plans a task. Brand DNA holds your writing styles, personality, customer profiles and ban list. The help centre puts basic setup at under two hours.
It has limits. The ban list is a strong guideline, not a strict filter, so a banned term can still appear, especially in longer documents.
With Tandem, one brief is split across specialist agents, and every agent works from the same saved Brand DNA. We do not claim Mavic can judge whether work is spiritually on brand. It follows the standard your team wrote down, and you decide what goes out.
Keep a named person at the gate
A checker is a useful first reader. It is not the one who answers for the brand when something goes wrong.
Keep one named person who can say no, and publish nothing without their approval. Then review their rejections once a month. They are the best source of your next rule.
Sources
- Adobe, “Adobe Ushers in a New Era of Creativity with New Creative Agent and Generative AI Innovations in Adobe Firefly”, press release (15 April 2026)
- Adobe, “Adobe introduces Brand Intelligence” (20 April 2026)
- diginomica, Phil Wainewright, analysis of Adobe MAX 2026 (11 November 2026)
- Okoone, reporting an Optimizely and Savanta survey (6 July 2026)
- Mavic help centre, Brand DNA overview
- Mavic help centre, Ban List
