How to Brief AI on Your Brand: Guidelines Your AI Tools Can Follow

By Yamm Labs · Updated October 2026 · 14 min read

A content lead at a life insurer uploads the full brand book PDF into a chatbot and asks for three LinkedIn posts on term cover. The drafts arrive in under a minute. One calls the company by its app nickname, one promises “complete peace of mind”, and none carries the insurer’s registered name. Compliance sends all three back the same afternoon. The brand book was not the problem. It was written for designers who read it once and absorb it. An AI tool needs short rules it applies every time, plus a list of regulated lines it may only copy. Someone also has to keep both current.

What do you need before writing brand guidelines for AI?

You need four owners, two written approvals and a folder of inputs before anyone writes a rule. Skip any of them and the brief stalls at the first review.

The owners come first. The marketing or brand head owns the brief and settles arguments about tone. A named compliance reviewer approves every regulated line the tools are allowed to use. Each product owner (cards, loans, term plans, a mutual fund scheme) answers for the rates, fees and limits of that product. A tools owner, usually in marketing operations or IT security, knows which AI tools the company has approved and where each one accepts instructions.

The approvals are a list of approved AI tools and a written rule on what data may go into them. Customer data stays out of public AI tools. India’s Digital Personal Data Protection Act, 2023 requires consent that is free, specific, informed, unconditional and unambiguous, and a prompt pasted with a customer’s policy details was never part of what that customer agreed to. If you are regulated by the Reserve Bank of India (RBI), check whether your board has adopted an AI policy. RBI’s committee on a Framework for Responsible and Ethical Enablement of AI (FREE-AI) recommended in its August 2025 report that regulated entities adopt a board-approved AI policy. Your brief should sit under that policy, as one of its working documents.

The inputs are documents you already have. Collect the current brand book and tone guide, a sample of recently approved copy and recently rejected copy, every disclaimer compliance has signed off, the live product pages with their rates and fees, and a list of every name the company goes by. The rejected copy teaches more than the approved copy, because it shows where the house has drawn its lines in practice.

One more decision belongs here: who is accountable when an AI tool gets it wrong. The FREE-AI report says accountability stays with the entity that deploys AI and cannot be handed to the model. Write the name of the person who approves AI-assisted content into the brief itself. That one line settles most later arguments.

Phase one: your brand book has to become instructions a model can follow

A brand book describes a feeling, and a model needs instructions it can obey and you can check. “Our voice is warm and confident” is a description. “Write in sentences under 20 words, address the reader as you, never use an exclamation mark” is an instruction. Phase one converts the first kind into the second.

1. Step 1: write the identity block. In one short paragraph, state the registered company name, the trade name, the app name and the product names, in the form each one must take. Say what the company is licensed to do and what it is not, in plain words: a mutual fund distributor arranges investments and does not give investment advice. List the names the company must never be called, such as an old brand name after a merger or a nickname customers use on social media. Add the markets you serve and the languages you publish in.
2. Step 2: turn your voice into rules with examples. Write ten to fifteen rules, each one testable by a reviewer who has never met your team. Pair each rule with one approved sentence and one rejected sentence from your own archive. Add a banned-phrase list for claims your compliance team has refused before, such as promises of returns, “risk-free” or “best in India”. Say which spelling convention to use, how to write rupee amounts (₹5 lakh, not ₹500,000), and whether the tool may translate copy or must leave translation to your approved translators.

The identity block carries the first regulator-driven decision on this page: entity naming. Under IRDAI’s 2024 regulations on protecting policyholders’ interests, every insurer’s advertisement must show its registered name with its trade name or logo prominently. So the identity block for an insurer tells the tool to write the full registered name on first mention in any advertising copy, then the short name after that. A mutual fund distributor makes a similar call: AMFI requires the “AMFI-registered Mutual Fund Distributor” tagline with the distributor’s name in its communication, so that tagline belongs in the identity block, word for word.

The table below shows the conversion in practice. The company name is hypothetical.

Keep the identity block and the voice rules short enough to read in five minutes. Long briefs get truncated by some tools and skimmed by every person, and the rules that matter most end up buried on page six.

SIGNALNewsletter by Yamm Labs

The next piece, straight to your inbox.

New writing on BFSI brands, B2B and influencer marketing, and AI search.

Email *

Subscribe

Subscription form is not available at the moment

One click to unsubscribe. Privacy policy

Phase two: regulated lines are copied from a library, never generated

In brand guidelines for AI, any line a regulator controls should reach the tool as fixed text to copy, never as an idea to write up. A model paraphrases by default, and a paraphrased disclaimer is a different disclaimer. Phase two builds the two files that stop that from happening.

1. Step 3: build the fixed-lines library. Each entry holds the exact approved text, where it goes, the format rule, the languages it is approved in, the approver’s name and the approval date. Tell the tool, in the brief, to copy entries character for character and to stop and flag the draft if no entry fits.
2. Step 4: build the facts register. List every rate, fee, limit and eligibility condition the tools might mention. Each row carries the value, the page on your own website where it is published, the product owner, the date it was last checked and the date it must be checked again. Tell the tool that any number missing from the register must be replaced with the word CHECK in capitals, so a human sees the gap before anyone else does.

Three examples show what goes in the library. For a mutual fund, SEBI’s advertisement code requires the standard warning, “Mutual fund investments are subject to market risks, read all scheme related documents carefully”, on communication that solicits investment in a scheme. In audio-visual advertisements SEBI requires that warning to be displayed and voiced for five seconds. So the library entry for video scripts carries the text and a timing rule, and the tool writes the script around the warning instead of squeezing it in at the end.

For a lender, the decision is what the tool must never do. RBI’s rules on the Key Facts Statement (KFS) require lenders to give retail and MSME term loan borrowers a KFS that includes the annual percentage rate (APR) before the loan contract is executed. That applied to loans sanctioned from 1 October 2024. An AI tool should never calculate or restate an APR in marketing copy. The library entry tells it to point the reader to the KFS and to use only the rate wording product has approved.

For an insurer working through partners, IRDAI’s 2024 regulations say an advertisement issued by a distribution channel needs the insurer’s written approval before it is issued. Any copy an AI tool drafts for a bank partner or a broker is therefore labelled as a draft for insurer approval, and the brief says so.

The facts register carries the second decision: how a regulated fact stays current. When a fixed deposit rate changes, the product owner updates one row, and that row feeds both the AI brief and the public product page, which shows the figure with an “as of” date. That matters beyond your own tools. AI answer engines quote what they find on your public pages, so a stale rate on a forgotten page can end up in an answer with your name on it. Our AI Search and GEO work starts from this register for that reason.

Google’s own guidance on generative AI content backs the habit. It warns that generative models do not retrieve facts and tells publishers to fact-check and review AI output before publishing, including titles, meta descriptions and structured data. A facts register is how a marketing team does that review at speed.

Phase three: a brief only counts once it is loaded and tested in every tool

Brand guidelines for AI that sit in a shared drive change nothing. It has to be placed where each tool reads instructions, then tested against the same prompts every time something changes.

1. Step 5: load the brief into each tool. Chat assistants usually accept custom instructions or project files. In-house tools built on a model take a system prompt that your developers control. Agency partners use their own tools, so the brief and the library go into the agency contract, with a duty to tell you which tools they use. Put the identity block and the fixed-lines rule at the top of every instruction set, attach the library and the register as separate files, and record which version each tool holds.
2. Step 6: test, log and review. Keep a fixed prompt set of about twenty requests that covers each product, each channel and a handful of traps: a request to promise returns, a request to compare with a named competitor, a request to translate the standard warning into Hindi. Score every output from 0 to 2 on naming, voice, fixed lines and facts. Re-run the whole set after any change to the brief, the tool or the model version, and keep the scores in a log.

Customer-facing assistants need one extra line. The FREE-AI report says people should be told when they are interacting with an AI system. If your website or app runs a chat assistant, the fixed-lines library holds the approved sentence that tells the customer they are talking to an AI assistant, and the brief tells the assistant to use it at the start of every conversation.

The same identity block also has a public job. AI answer engines work out who you are from your website, your profiles and other sites that mention you. Schema.org defines legalName as the official name of the organisation, such as the registered company name, and sameAs as a URL that identifies the organisation without doubt. Publishing your registered name in Organization markup, with sameAs links to your official profiles, gives engines the same naming rule your internal tools follow. The GEO and AEO FAQ for financial brands covers the markup questions teams ask most.

How long it takes and who signs off

In our practice, a first working version of brand guidelines for AI takes four to six weeks, and most of that time is compliance review of the fixed-lines library. The timings below are typical practice, not a regulator’s requirement, and a firm with a large product shelf will take longer.

Sign-off follows ownership. The marketing head signs the brief as a whole. Compliance signs the fixed-lines library and nothing else, which keeps its review narrow. Each product owner signs the register rows for that product. The tools owner signs the list of tools and the versions loaded. Keep every signature with a date, because the first question an auditor or a board committee asks is who approved the version in use on a given day.

If a regulator controls a line, the model copies it and never writes it.

How to tell your brand guidelines for AI are working

The brief is working when compliance edits fewer AI-assisted drafts and no fixed line ever reaches review in the wrong form. Four measures show it, and each is cheap to track.

First-pass approval rate = drafts approved without edits ÷ AI-assisted drafts submitted. This shows whether the voice rules and the library are doing their job. Track it from the first month so you have a baseline.

Fixed-line errors should be zero. Count every draft in which a disclaimer, tagline or naming rule was paraphrased, cut or misplaced. One error means a tool is not reading the library, or someone pasted an old version.

Prompt-set scores show drift. If the average score on the same twenty prompts falls after a model update, you know before a customer does. Keep the scores by tool, because two tools fed the same brief rarely behave the same way.

The public check is the slowest and matters most. Once a month, ask the main AI assistants the questions your customers ask about your brand and products, and record whether they name the right entity, quote current figures and describe what you are licensed to do. We publish our own record in the Yamm Labs AI Visibility Log, starting from a day one baseline. Attribution here is hard: an engine may change its answer for reasons that have nothing to do with your brief. Treat the trend over several months as the signal and ignore single answers.

What to do when a step fails

Most failures come from one of five causes, and each has a fix that does not require starting over.

• The tool ignores a rule. Move the rule higher in the instruction set, shorten the brief around it and add a rejected example. If the tool still fails the test prompt, take that task away from the tool.
• Compliance will not sign the library. Split it. Ask for sign-off on the entries already in approved use, and park new wording as a separate request with its own date.
• A rate changes in the middle of a campaign. Update the register row, re-run the prompts that mention the product and search scheduled content for the old figure. The register’s next-check date exists to catch this before it happens.
• Staff use a tool that is not on the approved list. Find out why. Usually the approved tool lacks a feature. Either add the brief to the new tool through the tools owner or give people a reason to stay on the approved one.
• Agency drafts drift from the brief. Ask which tool and which brief version produced the draft. If the agency cannot say, the contract clause from step 5 is the conversation to have.

When the brief itself proves wrong, change it in one place, give it a new version number, reload it everywhere and re-run the full prompt set. Partial reloads are how two tools end up following two different briefs.

What to tell your compliance team

Your compliance team will want the plan in a form it can file. A short note along these lines usually works. “Marketing is setting up written brand guidelines for AI, a brief for the AI tools on our approved list. Regulated lines, including disclaimers, taglines and naming rules, will come only from a library you approve, with your name and the date on each entry. Rates and fees will come only from a register owned by each product team, linked to the live product page. Every AI-assisted draft still comes to you for review, and we will log test results each month and after every change to a tool. We are asking you to approve the library once, and then only the changes.”

This is general guidance, not legal advice; your compliance team decides what each line must say and where it goes. For a brand that also wants its public pages read and cited correctly by AI engines, our AI visibility work for BFSI brands starts from the same identity block and facts register.

People also ask

What should brand guidelines for AI include?

They should include an identity block that fixes how the company is named, voice rules with approved and rejected examples, and a library of regulated lines to copy word for word. Add a facts register for rates and fees and the name of the person accountable for AI-assisted content.

Can an AI tool write a regulated disclaimer?

It should not write one; it should copy one. Disclaimers such as SEBI’s mutual fund standard warning must appear in their approved form, so keep them in a library the tool copies from and flag any draft where no approved line fits.

Who should own the AI brand brief?

The marketing or brand head should own the brief as a whole. Compliance owns the regulated lines, each product owner owns that product’s figures, and a tools owner decides which AI tools are approved and which version of the brief each one holds.

How often should the brief be updated?

Review it monthly and whenever a product, a rate, a regulation or a tool changes. Re-run the same test prompts after every update so you can see whether the change improved the drafts or broke something.