4 habits that make or break your company's AI

A lot of AI advice these days is written about agents — AI that takes actions on your behalf, not just answers questions.

But most companies aren't using agents yet. They're still using chatbots: you ask, it answers, a person decides what happens next.

Here's the thing — whether you're using chatbots or agents, the advice is the same. What changes is what's at risk if you get it wrong.

A chatbot working from scattered or out-of-date information gives you a wrong answer. Annoying, but someone's still there to catch it. An agent working from the same information takes a wrong action — it just goes and does the wrong thing, often before anyone notices.

So the four habits below aren't just for companies rolling out agents.

They apply to any AI you're using today, chatbot or agent, and the new tools that will undoubtedly appear in the future. Get the foundations right now, and whatever AI tool you plug in next — including agents — will be working with clean structured data and knowledge.

1. Write like an AI is going to read it — because one will

Most of what gets written at a company is written for the next human who reads it. A Slack message, a half-filled-out CRM note, a file called final_v2.pdf — a colleague can usually fill in the gaps. An AI can't. It only has what's explicit.

This shows up constantly in brand and marketing work. A logo file named logo2.svg tells a person nothing and an AI even less. A campaign brief that lives only in someone's head isn't "written down" just because everyone on the team happens to know it.

The habit: name things, tag things, and write things assuming the reader has zero shared history with you. That's true whether the reader is a chatbot answering "what's our approved tagline for the EU launch?" or an agent trying to pull the right asset for a campaign.

2. Capture decisions, not just discussions

A meeting transcript is a record of people talking. It is not a record of what was decided. Someone restates the same point three ways, someone else half-agrees, and the actual decision — if there was one — is buried in there somewhere.

Whether you're asking a chatbot "what did we land on for the Q3 messaging?" or letting an agent act on the answer, both are stuck guessing if all you gave them is a transcript.

The fix doesn't need a tool. It needs one sentence at the end of every real conversation: we decided X, because Y, owner is Z. Write that down where it can be found later, and both chatbots and agents stop guessing.

3. Default to structured, not scattered

Direct messages, personal folders, and one-off documents are dark data — invisible to any AI, and often invisible to the next person on your team too.

Some of that should stay private, and that's fine. But brand and marketing content in particular tends to end up scattered by default: a guideline in someone's Drive, a logo variant in an email attachment, an approved case study buried three folders deep.

A chatbot answering "can I use this photo in a paid ad?" needs that answer to live somewhere findable. So does an agent that might one day be asked to build the ad itself.

The habit is boring but it works: put brand and marketing content in one place that's built to be searched, not a place it just happens to end up.

4. Keep your system of record current

This is the one that compounds. A chatbot pulling from an out-of-date brand guideline gives someone a wrong answer, stated with total confidence, no flag that anything's off. An agent doing the same thing does something wrong with that same confidence — and it might already be published before anyone notices.

Every AI you plug in — chatbot today, agent later — inherits whatever staleness is already sitting in your system of record. Clean it up once, at the source, and everything downstream gets better automatically. Don't clean it up, and every new AI tool just adds a faster way to be wrong.

What this means for MCP

MCP is the connector layer that lets a chatbot or an agent reach into your actual tools — your DAM, your CRM, your project tracker — instead of working from what you typed into a prompt.

It's a genuinely useful piece of infrastructure. But it doesn't change any of the habits above — if anything, it raises the stakes on them.

An MCP connection doesn't clean up your data on the way through. It just gives the AI faster, more direct access to whatever is already there — organized or not, current or not, tagged or not.

Connect an agent to a messy brand folder via MCP and you haven't fixed the mess. You've just given the mess a much faster way to reach the person asking the question — or the action the agent is about to take.

Which is really the whole point of having a proper system of record for your brand and marketing content in the first place. It's not just about you and your team being able to find things.

It's about having one clean, structured, current source that any chatbot, any agent, and any MCP connection can plug into and actually trust.

The real difference isn't the model

Within a year, most companies will have access to roughly the same models and the same connectors. The difference will be whether the company data/knowledge those models connect to — is actually usable.

Writing for the reader that isn't human, capturing decisions instead of discussions, keeping brand content structured instead of scattered, and treating your system of record as something that needs active upkeep — none of that requires new tools.

It requires deciding it matters, for chatbots now and agents later, and holding your team to it the way you'd hold them to any other standard.

That's the gap most companies haven't closed yet. It's also exactly the gap a proper system of record for your brand and marketing content is built to close.

Happy branding :)

P.S. You can now add knowledge content (such as blog posts, FAQs, Brand Guidelines) to your Brandkit account that is repurposed and (provided it is publicly viewable) served to AI chatbots and agents as both crawlable plain text data, and served via MCP to chatbots and agents. Learn more here.