What is a brand agent, and what does it need to be trusted?
An AI agent that works in your name needs what you would give a new hire: a clear job, limits, someone checking the work, and a paper trail.

A brand agent is an AI agent that does a recurring job in your company's name, such as drafting posts, answering buyers or preparing articles, working from your approved brand thinking instead of whatever the model already believes. You can trust one when it has a narrow job, written limits on what it may do alone, a word that stops it, a check by someone who did not make the work, and a record of every run.
On 11 November United Airlines said it would restrict AI bots’ access to its digital platforms, citing data security, both directly and through the travel agencies and aggregators that sell its tickets. CNBC reported United’s concern that customers sharing account details with third-party AI assistants could expose sensitive information and lead to unintended purchases.
That is a company deciding which agents may act in its customers’ names. Any brand that puts its own agent to work has the same decision to make about it: what may this thing do in our name, and who answers when it gets it wrong? We have had to write our answer down, because agents already run parts of our company.
What is a brand agent?
A brand agent is an AI agent given a recurring job in your company’s name and the approved brand thinking to do it with. The job might be drafting the week’s posts, answering a buyer’s first question on your site, or preparing an article like this one. What separates it from an AI tool is that it acts: it starts on a schedule or a trigger, works through several steps, and produces something that leaves the building or changes something inside it.
That last part is where the risk sits. A chat assistant that writes a weak paragraph costs you a minute. An agent that posts the weak paragraph at nine in the morning has spoken for you.
A tool waits for you. An agent acts for you, so it needs a job description.
What does a brand agent need before you trust it?
Six things, written down before the first run. They are what you would give a new hire handed your company’s name: somewhere to learn the brand, a job, limits, an off switch, a manager and a paper trail.
One approved source The agent reads your approved positioning, claims, voice and visual rules from one place your company owns, not from the model’s training or last month’s chat. If the brand lives in five places, every run picks its own version.
One narrow job Name the job, what it may touch and what it may not. “Publish two reviewed articles a week on our blog” is a job. “Help with marketing” is a wish.
A limit on what it does alone Decide, per kind of work, what the agent may do without asking and what waits for a person. Start with everything waiting. Widen the limit one kind of work at a time, and only after clean runs.
A word that stops it One plain command, held by a named person, that pauses the agent at once. If nobody knows the word, the agent has no off switch.
A reviewer who did not make it The agent that made a piece is the worst judge of it. A separate reviewer, a person or a second agent starting fresh, returns go, revise or hold before anything ships.
A receipt for every run A short record of what it read, what it changed, what a person decided and what is still open. Without it you cannot tell a good week from a lucky one.
We have written more on the last item in why every AI agent should leave a receipt, and on deciding who approves what in what AI brand governance is.
What does a trusted brand agent look like in practice?
You are reading the work of one. A Claude agent on our blog desk wrote this article from our Brand Brain, the files where we keep our approved positioning, claims and voice. It works on a clock, twice a week, under a standing approval Kevin Reilly, one of our two founders, recorded in writing: at most one article a day, no paid images, our own blog and nowhere else. A separate reviewer that did not write it had to return go before it could publish. The site’s publisher refuses any article without an approved cover. One command from a founder, STOP BLOG, pauses publishing.
Our other agents follow the same pattern. Grok runs our company account on X under its own written rulebook and stop word, and posts only through X’s official API, never by signing in through a browser. Codex makes our images and builds site changes on a separate branch, and never merges its own work into the live site.
The limits came from mistakes. On 17 September one of our articles went live showing “Cover pending” where its cover should have been, and the reviewer had passed it. We did not write a rule about that one cover. We changed what the reviewer asks, whether a good colleague would ship the piece as it stands, and made the publisher refuse any article without an approved cover.
This guide was meant to run on a Thursday. The run that writes it did not start on the Wednesday it was scheduled, so the article waited for the next slot, because our publisher releases nothing without an approved cover and a reviewer’s go. A late article costs us a few days. An unchecked one in our name costs more.
Should we let outside AI agents act on our site?
Decide it on purpose, because it changes how a buyer’s assistant can reach you. Companies are already choosing differently: CNBC notes that Amazon has blocked Meta’s AI agent Muse while Walmart has partnered with it, and that Expedia and Booking are working with OpenAI. United says it will restrict them.
Ask of an outside agent what you ask of your own: who is it acting for, what may it do, what happens when it gets something wrong, and who answers for that. United’s stated worry, purchases nobody intended, is the same failure a brand fears from its own agent.
Keep one thing separate. Stopping an agent from buying on your site does not stop assistants from describing you. They can still answer questions about your company from what they find, so what your pages say still matters. We covered that in how to check what AI says about your brand, and correct it.
Where should we start with a brand agent?
Pick one job your team does every week, in your name, that a reviewer can check in ten minutes. Write the six things for it on one page before the first run. Then run it with a person approving every piece, and widen the limit only when the receipts show clean weeks.
The source it works from takes longest to get right, and every later agent reads from it too. That is the job of a Brand Brain: the approved brand thinking your people and AI tools read before they make anything, kept in files your company owns. If your agents keep losing a correction between runs, read why AI forgets the correction you made last week.
Where this comes from.
-
News report
CNBC · United Airlines moves to block AI agents from its platform over security concerns
Used: United said on 11 November it would restrict AI bots' access to its digital platforms, citing data security, directly and through travel agencies and aggregators; its concern that sharing account credentials with third-party AI assistants could expose sensitive information and lead to unintended purchases; Amazon has blocked Meta's AI agent Muse while Walmart has partnered with it; Expedia and Booking are working with OpenAI. No effective date or list of permitted agents is given, so none is stated. United's own newsroom statement was not found; the report is the source.
-
First-party working record
Adaptius · How the blog desk, the X desk and the studio desk run (Brand Brain working record)
Used: the blog's standing approval (two reviewed articles a week, at most one a day, no paid generation, our blog only, STOP BLOG), the publisher's refusal without an approved cover and the 17 September 'Cover pending' incident that caused it, Grok's X rulebook on the official API, Codex never merging its own work, and the Wednesday 11 November writing run that did not start. Method in use for our own company, not a customer result. Checked against the Brand Brain Nov 12, 2026.


