A marketing team of four puts out around 20 assets a week. Checking each against the brand guidelines by eye was a human bottleneck, and it's now a system.

• Magi's Brand Audit Agent checks copy and images, from content made in Magi, a live URL, or an upload, against a brand defined as nine sections of rules in BrandOS. Every violation links back to the rule it broke.

• It ranks all the violations to display what's high, medium, and low.

• Severity is judged by the agent from what the brand itself said matters: an off-brand word outranks a sentence-case slip, for instance.

• It can catch a logo rendered green instead of #FF72FF inside an AI-generated image, and you can use Copilot to regenerate the image correctly.

  
  

Tirth Patel is a founding AI engineer at Magi, and the thing he's been building is aimed at what he calls one of the less glamorous problems marketers deal with: brand consistency. Nobody gets promoted for it; everyone gets blamed when it slips.

His arithmetic: a marketing team of four puts out something like 20 things a week. Posts, pages, emails, decks. Someone has to know whether each one follows the brand guidelines, and until recently that someone was a person reading everything, remembering rules nobody had written in one place, and getting tired by Thursday.

The first two posts in this series were about removing the transfer tax: the cost of moving finished information into new containers. Harshini stopped re-briefing the brand before every post. Deekshaa stopped copy-pasting collateral into templates. Both changes push the same number up: how much gets shipped. Which is exactly the number that breaks the person reading everything.

AI can now make 20 assets. Can your reviewer check 20 at the same time?

When drafting is slow, brand governance comes free with the bottleneck. Four people can only make so much, so one careful reviewer can see all of it.

Take the drafting cost out, and the reviewer is the constraint. Twenty assets a week across five voices and several channels is not a volume one careful person can hold in their head, and the failure is quiet: nobody notices the sixth email used the old tagline until a customer forwards it. The alternative most teams reach for is slowing production back down, which gives up the gain that created the problem.

Tirth's answer is that if the brand is the thing being checked, the brand has to exist somewhere a check can run against it, which means as rules rather than as a PDF.

You can't codify taste. You can codify consistency.

The objection from every brand person: you cannot reduce a brand to a rulebook. The reason a line lands, the judgment that this metaphor is ours and that one isn't, the sense of when to break your own pattern. None of that survives being written as a rule.

Correct. And it describes a small share of what actually goes wrong at volume.

Look at what a reviewer flags in a real week.

• A product name written the way the market writes it instead of the way the brand does.

• A hex colour that's near the palette but not in it. Sentence case where the brand uses title case.

• A list missing the Oxford comma the brand's language rules require. A logo mark recoloured to match the background.

• A claim phrased in a way legal signed off on once and never again. These are not taste failures.

  
  

They're rule failures, and they're the ones that multiply with output, because a person checking twenty assets will catch the first three and start skimming. So the useful question is whether the rule-shaped violations, which are the majority, can be caught by something that doesn't skim. Taste stays with the person, who now has the attention to spend on it.

A brand book describes your brand. A brand system can enforce it

In Magi, the brand is an ontology in BrandOS rather than a document, with nine sections: Brand Identity, Voice & Tone, Messaging, Language & Grammar, Visual Identity, Keywords & Topics, Content Guidelines, Legal & Compliance, and Social Media. Each holds rules, examples and, for visual identity, the approved assets themselves.

That structure is what makes the audit possible. When the Brand Audit Agent raises a violation, it points at a specific rule in a specific section. The Oxford comma flag on a Magi post cited the Language & Grammar rule that requires it, quoted the offending list, and stated the fix. The reviewer can respond in one of two ways: the rule is wrong, so change the rule, or the rule is right, so fix the draft. Neither response is available when a person has read the brand book and formed an impression. Traceability is what turns "this feels off-brand" into a decision someone can act on.

It also means the audit gets stricter as the brand gets more defined. A rule added to BrandOS on Tuesday is enforced on Wednesday's twenty assets without anyone being told.

Not all off-brand is equally off-brand: ranked violations

Every violation is graded high, medium or low. The grade isn't a fixed table Tirth wrote; the agent judges it from two inputs. What the brand put into BrandOS, including which rules it wrote down with force and which are stylistic preferences. And general principles of how content works: using the wrong term for your own brand is a higher-order problem than a sentence-case slip, because one changes what the reader believes and the other changes how the sentence looks.

So a competitor's name for your category appearing in your own copy is high. The paper-plane logo rendered in the wrong colour is high, because the logo lock-up is the brand's most protected asset. A missing Oxford comma is low unless the brand's language rules say otherwise. A reviewer can override or dismiss any flag, and the audit history keeps the record either way.

The point of grading is triage. Twenty assets a week might produce sixty flags. A team needs to know which five to fix before publishing and which fifty-five can wait for a batch pass.

Your AI content engine needs a brand firewall

The clearest example came from Magi's own content. A LinkedIn post about BrandOS, drafted in the Magi editor, came with a generated visual. The Audit tab ran against the Magi brand and returned one flag, on the image, not the copy: the Magi paper-plane mark had been rendered green, alongside the wordmark. The rule cited was Visual Identity, primary colour #FF72FF. The recommendation named the fix.

Tirth asked Copilot to apply it. The image was regenerated with the correct mark, and the surrounding colours moved closer to the palette as a side effect. Total elapsed time was shorter than it would take to open the file in a design tool.

Show Image After the Copilot fix: the agent's trace on the left, the regenerated BrandOS visual on the right.

Two things about this matter beyond the fix. First, the error was AI-made: a generated image had drifted off-brand in a way a person skimming a thumbnail would likely miss. The check caught a failure the same class of tooling introduced. Second, the check ran inside the editor where the post was written, as one of three tabs (Chat, Design, Audit), so it happened before publishing rather than after a customer forwarded the post.

Anything a Magi user makes can be audited against the brand, including live URLs

The agent lives in Workforce, the section of Magi for agents that take one job and run it end to end. To run an audit you pick the brand and a source. The source can be content generated in Magi, a live URL, or an uploaded file, with an optional content type so the agent applies the right template rules.

The live-URL path is what makes it a brand compliance tool rather than a drafting aid. A Magi user can point it at anything they've published, the homepage included, and get the same graded, rule-linked report. Magi's own audit history shows the magihq.com homepage, blog posts, a MCP campaign page and a set of LinkedIn drafts, each with format, violation count, status and timestamps. That history is also how a team sees whether it's drifting over time or tightening.

Automation works best where the rules are explicit and judgment still matters

Three limits worth stating plainly.

The agent can only enforce rules that exist. If a brand has never written down that its product name is one word and lowercase, the audit won't flag the version that isn't. The first weeks of using it tend to be about discovering which rules were assumed rather than recorded, and writing them.

Severity is judgment, so it can be wrong. That's why override exists, and why the history is kept. A reviewer who disagrees with a grade twice should look at whether the rule in BrandOS is stated with the force they intended.

And the audit doesn't tell you whether the post is any good. It tells you whether the post is yours. A perfectly compliant asset can still be dull, and the person freed from checking Oxford commas is the one who has to notice that.

Diagnose the brand problem before you automate the fix

Pull the last month of published assets and list every brand slip you can find. Then sort the list, not by frequency, but by how much each one would have cost if a prospect had seen it. That ordering is your severity model, and it's the input any brand compliance software needs and none can invent for you.

If most of the list is rule-shaped, you have a volume problem and a system fixes it. If most of it is taste, you have a definition problem, and the fix is upstream in the brand itself.

Frequently asked questions

What is brand compliance? Brand compliance is the degree to which published assets follow a brand's defined rules for voice, language, messaging, visual identity and legal claims. It is checked per asset, unlike a brand audit, which assesses how a brand is perceived overall.

What is a brand audit agent? A brand audit agent is software that checks a piece of content against a brand's defined rules and reports violations. Magi's Brand Audit Agent checks both copy and images, grades each violation high, medium or low, and links every flag to the specific rule in BrandOS it breaks.

How does Magi's Brand Audit Agent work? Magi's Brand Audit Agent takes a brand and a source, which can be content generated in Magi, a live URL, or an uploaded file, and checks it against the nine-section brand ontology in BrandOS. It returns graded violations with the rule cited and a recommended fix, which Copilot can apply, including regenerating an image.

Can the Brand Audit Agent check images? Yes. It audits visuals against the Visual Identity section of BrandOS, including logo usage and colour. In Magi's own content it flagged a paper-plane logo mark rendered green instead of the primary #FF72FF and regenerated the image correctly.

What determines whether a brand violation is high or low severity? The agent judges severity from what the brand defined in BrandOS and from general content principles. A wrong term for the brand or an incorrect logo is high because it changes what a reader believes; a sentence-case or punctuation slip is low unless the brand's rules elevate it. Reviewers can override any grade.

Can Magi audit a live website for brand compliance? Yes. A Magi user can point the Brand Audit Agent at any live URL they've published, including the homepage, and receive the same graded, rule-linked report as for content drafted in Magi. Results are kept in an audit history.

What is brand compliance software? Brand compliance software checks assets against a brand's rules before or after publishing and reports where they diverge. Effective tools require the brand to be defined as structured rules rather than a PDF, so violations can be traced to the rule they break.

What are the nine sections of the BrandOS brand ontology? The nine sections are Brand Identity, Voice & Tone, Messaging, Language & Grammar, Visual Identity, Keywords & Topics, Content Guidelines, Legal & Compliance, and Social Media. Every Brand Audit violation links to a rule in one of them.

About this post. Third in The 100x Marketer, a series on how Magi's own team uses the product. Sourced from Tirth Patel's recorded workflow episode, the Brand Audit Agent as configured for Magi's own brand, and Magi's audit history. Published by Magi HQ, an agentic marketing automation platform for lean B2B teams. Customers include 100ms, Payactiv, Lyric and Accuknox.

TL;DR. Tirth Patel, AI engineer at Magi, built the Brand Audit Agent to solve brand compliance at volume: a team of four shipping around 20 assets a week can no longer be checked by one careful reader. The agent lives in Magi's Workforce section and checks copy and images from content generated in Magi, a live URL, or an uploaded file, against the brand as defined in BrandOS, an ontology of nine sections: Brand Identity, Voice & Tone, Messaging, Language & Grammar, Visual Identity, Keywords & Topics, Content Guidelines, Legal & Compliance, and Social Media. Every violation links to the rule it breaks and is graded high, medium or low by the agent based on what the brand defined and on content principles, so a wrong brand term outranks a casing slip; reviewers can override. In Magi's own content it caught a logo rendered green instead of #FF72FF inside a generated image, and Copilot regenerated it. The Audit tab sits in the editor beside Chat and Design. Limits: unwritten rules can't be enforced, severity is judgment, and compliance is not quality.