AI for B2B Marketing Content: ChatGPT or Jasper, Your Platform's AI, or Agents That Work From Approved Facts

Ask an AI tool for a case study and a fluent draft comes back quickly. It may also contain a saving that no client ever reported. There are three routes to using AI for B2B articles, posts, emails and sales pages. You can use a general assistant or writing tool, or the AI inside the marketing platform you already pay for. The third route is a set of agents that research, write and check from one base of approved company facts, with a person approving what goes out. The first or second route may be enough. On every route, the result depends on two things you set up yourself: a written list of approved facts and a named person who signs off.

What goes wrong when AI writes marketing content?

The first problem is a confident fact that nobody gave the model. Anthropic's documentation says “Even the most advanced language models, such as Claude, can sometimes generate text that is factually incorrect or inconsistent with the given context.” Google's guidance for site owners explains why: “generative models don't retrieve facts, but predict a likely sequence of words based on their training data.” In marketing copy, that means a percentage, a client name, a certificate or a feature that sounds right and has no source.

The same mechanism hurts your voice. Without your own material and rules, the likely wording tends to be everyone's wording, and your posts can start to sound like your competitors' posts. Google also warns that “using generative AI tools or other similar tools to generate many pages without adding value for users may violate Google's spam policy on scaled content abuse.”

Then there is the claim nobody approved. The UK code for non-broadcast advertising, the CAP Code, says “marketers must hold documentary evidence to prove claims that consumers are likely to regard as objective and that are capable of objective substantiation.” It also requires evidence “that a testimonial or endorsement used in a marketing communication is genuine”, and a client quote that a model made up fails that test. For US broker-dealers, FINRA states “Firms are responsible for their communications, regardless of whether they are generated by a human or AI technology.”

These problems reach readers if nobody owns the final check. When the writer trusts the editor and the editor trusts the tool, nobody has read the numbers.

What does a writing tool such as ChatGPT or Jasper give you?

The first route is a general assistant such as ChatGPT or Claude, or a writing tool for business, such as Jasper or Writer. Anthropic says Claude's projects “allow you to create self-contained workspaces with their own chat histories and knowledge bases”, with instructions that can set a tone.

A general assistant can even run a simple fact check. An example prompt in Anthropic's documentation asks for a press release “using only information from these product briefs and market reports”. It then adds: “If you can't find a supporting quote for a claim, remove that claim from the press release”. Any team can copy that pattern with its own fact sheet, and for a small team that may be all it needs.

The limits come from how people use these tools. Each writer keeps their own prompts and files, so one drafts from this quarter's fact sheet and another from last year's. Unless you set it up, nothing records where a claim came from or stops an unread draft from going out. Settings can also have no effect until someone connects them. Writer's help center warns: “In order for your style guide and terms to apply to outputs generated in WRITER Agent you must first connect them to the voice profile.” Plans cap what you can load, too: Jasper's Pro plan lists “2 Brand Voices, 5 Knowledge assets and 3 Audiences”.

What does your marketing platform's AI add, and where does it stop?

The second route uses AI that may already be part of a tool you pay for. HubSpot's help page says “Use HubSpot's AI tools to tailor generated or written content to your brand's specific voice.” Its product page lists a Content Agent with the line “Build blog posts, social content, and landing pages in your brand voice.” Adobe GenStudio for Performance Marketing gives each draft a “Percentage of compliance with your Brand”, based on how many of your guidelines it passes. For single image ads in Campaign Manager, LinkedIn says it “can generate multiple ad variants, including headlines, intro text, and suggested visuals.”

Drafting then happens where content is scheduled and measured, under user rights you already manage. The drawback is a narrow view. Each platform holds its own content and settings, so the facts in your HubSpot emails and your LinkedIn ads can drift apart. Approval controls may need a higher plan. HubSpot says “An Enterprise subscription is required to access Approval settings for specific content types”, and also that “Super Admins can skip approvals.”

A brand score counts the guidelines a draft follows. Whether a number in the copy is true is a separate check, so ask what the platform compares claims against. For features that platforms now call agents, ask where the facts come from and who approves the result.

What do agents working from approved facts add, and what do they cost?

An agent setup splits the work between several AI agents. Anthropic describes agents as “typically just LLMs using tools based on environmental feedback in a loop.” LLM means large language model, and tools are functions the model can call, such as a web search or a document store. One agent collects sources for a topic, and another drafts from the approved facts and voice rules. A checking agent then compares every claim with the facts and flags what it cannot match. A person approves or rejects the piece, and the system records who did.

The checking step follows a pattern Anthropic calls evaluator-optimizer: “one LLM call generates a response while another provides evaluation and feedback in a loop.” Anthropic also writes that agents “can then pause for human feedback at checkpoints or when encountering blockers.”

This setup can pay off when the volume of content is high. Every channel and language draws on one set of facts, and the reviewer sees which claims were checked. It also takes work. Anthropic, which sells models for such systems, advises “finding the simplest solution possible, and only increasing complexity when needed. This might mean not building agentic systems at all.” Someone must keep the fact base current, maintain the agents, pay for model use and act on the flags. If one writer and one editor cover your volume, that work gives you little in return.

What goes into a fact base of approved claims?

A fact base is the list of statements your company has agreed it may make in public, each with its evidence. Keep one list that every writer and every tool reads. For each statement, record:

• the exact wording, so writers and models copy it instead of rephrasing it;
• the evidence: a document, a figure with the date it was measured, or the person who confirmed it;
• where it may be used, for example the website and sales emails but not paid ads;
• the event that sends it back for review, such as a price change or a client withdrawing consent.

Add the words the company never uses, such as risk-free or fully secure. Client names, logos and quotes go in only with written consent.

Start with the statements your sales team repeats most often. When a draft contains a claim that is not on the list, the draft waits until the claim is approved or removed.

How do we write voice rules a model will follow?

Voice rules work when someone can check a draft against them. Adjectives such as friendly or professional cannot be checked. Rules like these can be checked:

• Address the reader as you and the company as we.
• One idea per sentence, and no sentence longer than two lines.
• A number with its unit instead of an adjective such as fast or powerful.
• Say what the product does not do as plainly as what it does.
• Banned words, each with the word to use instead.

Give each rule its reason and two lines from your own past content, one that follows it and one that breaks it. Anthropic's prompting guide says “Examples are one of the most reliable ways to steer Claude's output format, tone, and structure.” Keep the list short. For editorial rules in GenStudio, Adobe suggests 5-10 guidelines and asks for “clear and actionable guidance”, for example “Keep sentences under 20 words” instead of “Be concise”.

Test the rules before you rely on them. Ask a colleague to pick the AI draft out of a set that includes your own published pieces.

Who checks a draft before it goes out, and in what order?

Write the order down and keep it the same on every route. First comes a fact check against the approved list, by a person or a checking agent. Then an editor compares the draft with the voice rules. Claims about law, security, prices, client results or financial products go to legal or compliance. Last, one named person approves the final version, and the record shows who approved what and when. Give that person the claims with their sources and any flags, so the check is quick.

Some regulated firms already work this way. FINRA Rule 2210 says “An appropriately qualified registered principal of the member must approve each retail communication before the earlier of its use or filing with FINRA's Advertising Regulation Department”. That rule applies to FINRA's member broker-dealers and lists exceptions, but any company can copy its logic and publish nothing without approval.

What do we paste into AI tools, and what do vendors do with it?

Drafts can start from material that is not public yet, such as a price change or a client's results. Before anyone pastes it, check which plan they are on and what the vendor says it does with the data.

Terms differ by plan. Anthropic says that by default it will “not use your inputs or outputs from our commercial products” to train its models. Its consumer plans give users “the choice to allow their data to be used to improve Claude”, so a writer on a personal account may work under different terms. OpenAI says “data sent to the OpenAI API is not used to train or improve OpenAI models (unless you explicitly opt in to share data with us).” That line is about the API; for ChatGPT, check the terms of the plan each writer uses.

HubSpot says it “may use customer data, as defined in the HubSpot Customer Terms of Service, to train and improve HubSpot's own AI models”. Of its opt-out switch, it says “By default, the setting will be turned on.” It also says it does not let third-party AI providers train on customer data. Writer's help center says “We do not train our LLMs with customer data.”

Check these points in the contract and the data processing terms, and keep a dated copy. Give the team a one-page rule: which tools and plans are approved, and what never goes into any tool, such as unreleased financial results or a client's data.

How do vendors charge for AI that writes marketing content?

Vendors charge per seat, from a pool of credits, by usage, or by a mix. Jasper prices its Pro plan per seat per month and quotes custom prices for Business. Anthropic also prices its Team plan per seat per month. HubSpot says “Many AI features and agents run on HubSpot Credits”, bought “up front in packs for predictable spend or as they are used with pay-as-you-go.”

Usage pricing charges for the text a model reads and writes, counted in tokens. Tokens are small pieces of text that providers count and bill. Anthropic's API pricing page lists Claude Sonnet 5.5 at 2 dollars per million input tokens and 10 dollars per million output tokens. List prices change. Anthropic's Enterprise plan mixes seats and usage: “$20 per seat per month plus usage billed at API rates, so cost scales with the models and tasks your team runs.”

An agent setup calls a model several times for each piece. Request the measured model cost of one approved article from the team that would build the setup. Then add the hours of the people who keep the fact base. On any route, get the price in writing, including minimum seats and add-ons, and ask what happens when credits run out.

How do we measure results, including what AI assistants say about us?

Measure the process first. From the first week, track:

• time from brief to approved piece;
• factual corrections found in review, per draft;
• drafts rejected outright;
• editor hours per published piece.

If the number of corrections does not drop during the first months, fix the fact base and the voice rules before you change tools. Measure outcomes as you did before: search traffic, replies, booked calls and deals where the content played a part.

You may also want to know whether AI assistants name your company when buyers ask about your field. Google says “There are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary”. It adds that sites appearing there “are included in the overall search traffic in Search Console.” OpenAI says “OAI-SearchBot is used to surface websites in search results in ChatGPT's search features”, and that sites opted out of it “will not be shown in ChatGPT search answers”. Check that your robots.txt file does not block the crawlers of the assistants you care about.

To track mentions, write down twenty real buyer questions. Ask each in every assistant you care about, several times on a fixed schedule, and record whether you are named and which pages are cited. Count how often you are named each month, since one answer tells you little.

Which route fits our company?

Use a general assistant or writing tool when one or two people write a few pieces a week and an editor reads every one. Your platform's AI fits when content is created and published inside that platform, and its approval and voice settings cover your channels. Consider agents working from one fact base when several people publish across channels and languages, facts change often, regulated claims appear in the copy or one wrong number is expensive. If nobody can own the fact base, stay with the first two routes, because agents working from stale facts spread them to every channel at once.

What should we ask a vendor, or a team that would build an agent setup?

For a writing tool or a platform vendor:

• Which facts does the tool draft from, and can we see where each claim came from?
• Is there an approval step that nobody can skip, with a record of who approved which version?
• Is our data used to train any model, on which plan, and where does the contract say so?
• What does a month of our real volume cost, and what can we export if we leave?

For a team that would build an agent setup:

• Where will the fact base live, and who on our side can change it?
• How does the checking step decide that a claim is supported, and what happens to claims it cannot match?
• Which models and vendors will process our drafts, and under which data terms?
• What does one approved article cost in model use, and who maintains the agents after launch?
• What do we own at the end: code, prompts, the fact base and the logs?

Where does amBrain fit?

amBrain publishes this blog and has an interest in this topic. On its own product, it says: “Our own product, aigo, runs on AI agents that research, write, check and measure from one shared memory of company facts. A person approves everything that goes out.” That matches the third route above, so read this guide knowing that its publisher has a product of that kind.

This article is not a case study and describes no client work. It does not say that amBrain builds agent systems for clients, and it reports no results from aigo. The other companies, products and regulators named here are cited from their published pages. The article claims no relationship between amBrain and any of them and gives no amBrain prices or timelines.

amBrain has been building software since 2019. On how it works with clients, it says: “Three formats: full delivery, a dedicated team, or engineers embedded in your team.” Its line on ownership is “The client keeps full ownership of the product and the code, except our reusable components.” The questions above apply to any firm you consider.