A brain is worthless without an opinion

You’ve probably seen a company brain diagram showing off how well AI knows a business. Hundreds of glowing nodes. Connections everywhere. It looks like the machine has absorbed the whole company.

Fill a marketing agent’s brain with accounting context and the diagram will light up just as nicely as one filled with the context that agent needs.

As proof that AI understands your business, the picture is meaningless. It tells you nothing about whether the agent knows which acquisition metric you trust, why you changed your positioning, or when it should ask before recommending a budget change.

Loading post from X…

View @typesfast’s post on X ↗

A visualization shows structure. It doesn’t establish whether an agent uses the right knowledge in its work.

Someone has to decide what matters for the work you want the agent to do. That’s the opinion.

We do a lot of brain implementation at Bamboo for marketing teams. The teams bring the knowledge of their business; we work with them to make it useful to their agents. Five lessons keep coming up. They all come back to one belief: AI learning your business takes you.

You can connect Slack, Drive, email, meeting transcripts, and everything else. That doesn’t mean your AI knows what matters.

What matters is what happens when somebody asks Claude to do something:

1. 01

   Consult

   Did Claude know to consult the brain?

2. 02

   Retrieve

   Did the brain return the right context?

3. 03

   Use

   Could Claude use it to produce the right output?

A brain can contain the answer and never get consulted. It can return a document that stopped being true last quarter. It can supply exactly the right information and still leave the agent applying it badly. Each failure needs a different fix.

1. Curate your brain. Start with the work.

I’ve seen startups pitch turning your entire Slack history into a company brain. I don’t believe in that as the starting point today.

Your Slack history contains decisions, abandoned ideas, corrections, exceptions, and people thinking out loud. Connecting it gives an agent access to those things. Someone still has to distinguish what the team agreed from what somebody suggested.

Start with one marketing team and a job you want its agents to do well.

For a weekly campaign review, that might mean the current acquisition targets, your metric definitions, the reporting sources you trust, recent campaign changes, and examples of a good review. Find the useful documents you already have. Ask the team to fill in what those documents leave unexplained.

The goal is enough relevant context to do the job. A smaller brain built around that work gives you something you can inspect, question, and improve. You can expand it as you give agents more responsibility.

2. Have an opinion about what a great new hire needs to know.

Imagine a strong marketer joining your team on Monday. What would you need to explain before trusting their recommendations?

Strategy, goals, customers, brand, processes, examples of great work. Start there. Then get specific.

“Understands our goals” is a useful category. “Knows which acquisition target applies to this campaign, how we calculate it, and when an exception is allowed” gives you something to check.

Suppose an agent sees an $18 CPA against a $15 target and recommends pausing a campaign. The arithmetic is fine. But the team approved a seven-day audience test, and today is day three. A useful recommendation needs that decision and its budget limit.

Your team supplies the judgment about what the agent ought to consider. Turn it into questions the brain should help answer. Which target applies? What was approved? Who decides whether to stop early?

A checklist helps you start, but uploading a document called “Campaign goals” doesn’t prove those questions are answered. Read what it says. See whether an agent can use it. You may discover that the exception everyone on the team knows has never been written down.

You’ll miss things. That’s okay. You now have a way to find them.

3. Use session logs to find what’s missing.

Your team’s conversations with Claude or ChatGPT are probably the best roadmap for improving its brain.

Look at where somebody had to step in. What did they explain manually? Which document did they attach? What do they keep prompting over and over again?

In the campaign example, the marketer might reply: “We agreed to let this run for seven days. Check the brief.” That correction gives you a specific problem to investigate.

If the agreement was missing from the brain, capture it. If it was already there, check whether the agent looked for it and what came back. If the agent received the brief but ignored the exception, work on how it performs the review.

Adding another document won’t fix an agent that never consults the documents it already has.

Also, read the available tool history when checking what happened. An agent saying “I checked the brief” doesn’t establish that it did. Where you can’t observe the call, keep that uncertainty visible.

After a correction, ask the agent to do the work again and inspect the result. Saving the explanation is only part of the job.

4. Make learning part of how your team works.

Your company’s most important knowledge doesn’t only live in documents. It emerges in meetings, emails, Slack, project decisions, and the explanations teammates give each other every day.

Humans experience a lot, but remember selectively. Your company brain should do the same: experience broadly, learn selectively.

That is why connecting those sources can be useful. A meeting transcript may contain the reason you changed your acquisition target. An email may explain an exception for a particular customer. A Slack thread may settle which report the team should trust.

Preserve the decision with enough context to use it: what changed, why, who owns it, and when it applies. A suggestion in a meeting should remain distinguishable from an adopted policy.

Make this part of the work while the people involved still remember the reasoning. When you finish explaining a decision to an agent, decide whether the next person doing that job will need the explanation too. If they will, put it somewhere the team’s agents can find it again.

Otherwise, every conversation starts with somebody rebuilding the same context by hand.

5. Fight staleness, including what depends on the old information.

People update their understanding as the business changes. Your brain needs the same attention.

Suppose you change your positioning for Q3. Updating the positioning document leaves another question: what was built from the old version?

The campaign brief, sales deck, messaging guide, examples, and agent instructions may all carry the Q2 story. Some will still be valid. Others will need to change. Check them against the new decision.

An agent can retrieve an old brief perfectly and produce a confident answer that your team has already moved past. Search working correctly doesn’t make the source current.

Keep the current guidance clear, preserve the history as history, and review the material that depends on a changed source. Then inspect the next piece of work. Does it use the new positioning? Does it still repeat the claim you retired?

A brain confidently returning something that stopped being true six months ago can be worse than admitting it doesn’t know.

We’ve seen too many internal AI projects launch with excitement and slowly flicker out a month or two later. Keeping agents useful requires people to explain what matters, review the work, and update what changed. That participation has to survive the launch.

At Bamboo, we do the hands-on work with marketing teams to build that muscle in-house. It’s part of becoming an AI-native marketing organization: the team learns to teach, evaluate, and improve its agents as part of running the business.

Pick one workflow your team does every week. Write down what a great new hire would need to know to do it. Then open a recent agent conversation and find one thing somebody had to explain for the second time. Start there.