There is a very seductive AI workflow where you build yourself a little board of advisors.
You give one agent the personality of a ruthless CFO. Another becomes the customer whisperer. Another is the brand strategist who definitely owns linen pants. Another is the skeptical operator who says annoying, accurate things like, “What decision are we actually making?”
Then you ask them to review your idea.
And for a few minutes, it feels incredible.
The CFO finds the risk. The customer person finds the pain. The strategist finds the positioning gap. The operator politely removes the scented candle from your brain and asks for a next action.
Everyone is smart. Everyone has a take. Everyone is wearing the little name placard you made for them in the prompt.
And then, three hours later, you cannot remember what changed.
You have twelve paragraphs of impressive commentary, four conflicting recommendations, a new anxiety rash, and no record of the actual decision. The AI board did what most advisory boards do in bad startup mythology: it made the founder feel briefly less alone while creating a tasteful fog around accountability.
This is the problem with most “AI board of advisors” workflows.
They create perspective, not memory.
They create stimulation, not governance.
They create a very convincing little room where everyone talks and nobody keeps minutes.
And if nobody keeps minutes, it is not a board.
It is a group chat with better posture.
The advisor fantasy is not wrong
I do not want to dunk on the board-of-advisors idea too hard because, honestly, I get it.
Most of us are making too many decisions with too little context and too few people who can think with us at the exact moment the decision appears. You may have smart friends, mentors, coworkers, founders, editors, investors, or brutally honest group chats. You probably do not have all of them available at 11:43 p.m. when you are trying to decide whether a product idea is real or just caffeine wearing a blazer.
AI is genuinely useful here.
A good model can help you simulate perspectives you would otherwise forget to invite. It can stress-test a claim. It can ask the dumb question you were avoiding because the answer would create work. It can role-play the buyer, the skeptic, the editor, the CFO, the exhausted future version of you who has to maintain the thing you are excitedly inventing.
That is valuable.
But the value is not that the advisors sound smart.
The value is that the discussion changes the decision.
If your AI advisors do not leave behind a trace of what was decided, why it was decided, what risks were accepted, and what must be checked later, they are not helping you build judgment. They are helping you feel like judgment happened.
Those are different things.
One compounds.
The other produces a very nice transcript you will never open again.
The meeting is not the asset. The minutes are.
The mistake is treating the AI conversation as the work product.
You ask your advisory council a question. You get a long answer. You skim, nod, steal two phrases, maybe paste something into a doc, then move on. The transcript sits there like a hotel gym you technically had access to but blissfully ignored on the way to the bar.
But the transcript is not the useful part.
The useful part is the minutes.
Minutes are not a summary in the bland “here are the key takeaways” sense. Please do not let your AI become the person in the meeting who says “great discussion” and then sends a recap that somehow makes everyone sound employed by Deloitte. Don’t tell my wife I just made that joke.
Good minutes do something sharper:
They preserve the decision.
They name the argument.
They capture the dissent.
They assign the next action.
They create a future checkpoint where reality gets to vote.
That last part matters most. The point of an AI advisory board is not to manufacture certainty. It is to make your uncertainty more inspectable.
Without minutes, every session starts from scratch. Your AI council can give you a strong take on Monday, a different strong take on Thursday, and a third strong take two weeks later because nothing in the system remembers what was already considered, rejected, accepted, or proven wrong.
This is how people end up with “AI strategy” that is mostly vibes in a nicer font.
The work gets polished.
The judgment does not compound.
What most AI advisor workflows get wrong
Most AI board prompts are built around cast lists.
You are my board of advisors. Include a CEO, CMO, CFO, product leader, customer, investor, and contrarian. Debate this idea and give me recommendations.
This is fun. It is also incomplete.
The missing layer is governance.
Not governance in the “let’s create a 17-tab spreadsheet and slowly become the reason people hate operations” sense.
Governance in the simple, adult, emotionally inconvenient sense:
- What decision is being made?
- Who gets a vote?
- What evidence counts?
- What disagreement matters?
- What happens next?
- When do we revisit the call?
Without that layer, an AI board is just a mood board for competence.
The personas create texture, but the minutes create memory. The personas give you angles, but the minutes give you a trail. The personas help you think, but the minutes help you notice whether your thinking improved.
That is the part most AI workflows skip because it feels less magical.
Prompting a “world-class strategy board” feels futuristic. Writing down “we chose option B because option A requires a distribution channel we do not have” feels like eating your vegetables while the robot watches.
Unfortunately, just like when you were a kid, the vegetable is still the point.
The AI Board Minutes template
Here is the practical version.
If you are going to use an AI board of advisors, do not end the session with advice. End it with minutes.
Use this as the output format:
# AI Board Minutes
Date:
Decision / question:
Context:
Options considered:
## Advisors present
- Advisor 1: role and lens
- Advisor 2: role and lens
- Advisor 3: role and lens
## Recommendation by advisor
For each advisor:
- Recommendation
- Reasoning
- Evidence used
- Assumption they are making
- What would change their mind
## Points of agreement
- What did multiple advisors converge on?
## Real disagreements
- Where did advisors disagree in a way that changes the decision?
## Decision
- Chosen path:
- Why this path:
- What we are explicitly not doing:
- Risks accepted:
- Unknowns remaining:
## Next actions
- Owner:
- First concrete action:
- Deadline:
- What proof would make us continue:
- What proof would make us stop or change course:
## Review date
- When will this decision be revisited?
- What evidence should be available by then?
This looks boring in the way useful things often do.
But it changes the whole workflow.
Now your AI advisors cannot just sound smart. They have to expose their assumptions. They have to say what would change their mind. They have to separate agreement from meaningful disagreement. They have to translate commentary into a decision and a next action.
Most importantly, the output becomes reusable.
You can save it in Obsidian, Notion, a project folder, a CRM note, a decision log, a GitHub issue, a Linear ticket, or whatever small operational cave you have chosen to live inside this quarter.
The tool does not matter as much as the habit.
Every important AI-assisted decision needs minutes.
The smallest useful version
If that full template feels too heavy, use the five-line version.
Decision:
Why:
Dissent:
Next action:
Review trigger:
That is enough to stop the work from evaporating.
Decision: what are we actually doing?
Why: what argument convinced us?
Dissent: what objection are we accepting or watching?
Next action: what happens in the physical world because of this conversation?
Review trigger: what future signal would make us revisit the decision?
If your AI workflow cannot produce those five lines, it probably did not help you make a decision. It helped you rehearse one.
This has become one of my quiet tests for AI systems in general.
Not “did it produce a good answer?”
Did it leave behind better state?
Because better state is what lets future work get smarter. A good decision note means tomorrow’s AI has something to retrieve. Next week’s version of you has a reason to trust or question the call. Your team can see why the work moved in a particular direction instead of reverse-engineering the vibes from a Slack thread with 43 unread messages and one person saying “circling back” in a tone that should be illegal.
The meeting fades.
The minutes compound.
A fake example, because otherwise this gets too abstract
Let’s say you are deciding whether to launch a new lead magnet.
The bad AI board workflow looks like this:
“Act as my board of advisors and tell me if I should launch a lead magnet for startup founders about AI content strategy.”
The AI says yes, but also consider differentiation, distribution, quality, ICP, funnel fit, and follow-up. Congratulations. You now have a bowl of strategic oatmeal.
The better version asks the board to deliberate, then keep minutes.
The output might look like this:
Decision:
Launch a narrow lead magnet for seed-stage startup marketers, not a broad AI content strategy guide.
Why:
The broad guide is more shareable but less diagnostic. The narrow version attracts a clearer buyer and can be tied directly to a sales conversation or product workflow.
Dissent:
The brand advisor worries the narrow framing may reduce top-of-funnel reach. The growth advisor accepts that risk because list quality matters more than raw downloads for this experiment.
Next action:
Draft a one-page “AI Content Engine Audit” with 12 questions and one scoring rubric. Test it with 5 friendly operators before building a landing page.
Review trigger:
If fewer than 3 of 5 testers say they would forward it to a founder/marketing lead, revise the angle before launch.
That is different.
It is not just advice. It is a small decision artifact.
You can disagree with it. You can improve it. You can run the test. You can come back later and ask, “Were we right?”
That is where AI starts becoming useful for operators instead of just stimulating for brains that already have too many tabs open.
Minutes are how you make AI less amnesiac
A lot of people talk about AI memory like it is a product feature.
The model remembers your preferences. The workspace remembers your files. The assistant remembers your tone. The agent remembers your tasks.
Fine. Great. Useful.
But some memory has to be intentionally made.
If a decision matters, do not trust ambient memory. Do not assume the transcript will be enough. Do not let the only record of your reasoning be a conversational blob titled “new chat” that sits next to nine other conversational blobs titled “new chat,” each one containing a small fossil of your ambition.
Write the minutes.
Or better yet: make the AI write them before the session ends.
This is especially important when you are using multiple AI roles because multi-perspective output can trick you into feeling like you did due diligence. The CFO objected. The customer spoke. The contrarian waved a little sword. Surely the decision is stronger now.
Maybe.
But if you do not capture which objection changed the plan, what assumption survived, and what evidence you still need, the “board” becomes theater.
Very sophisticated theater.
But theater all the same.
The real job is not more advisors. It is better loops.
The temptation with AI is always to add more intelligence.
More agents. More roles. More critics. More perspectives. More prompts. More tabs. More little synthetic coworkers with aggressively helpful energy.
Sometimes that helps.
But often the bottleneck is not intelligence.
It is continuity.
Your work does not get better because a fake CFO yelled once. It gets better because the system remembers what the fake CFO warned about, whether that warning mattered, and what you learned when reality checked it.
That is a loop.
Decision → action → evidence → review → updated judgment.
That loop is the actual advisor.
The board is just a way to create sharper inputs.
So if you are experimenting with AI councils, advisory boards, debate agents, simulated customers, expert panels, or any other tiny conference room inside the machine, keep going. They can be useful. They can save you from obvious mistakes. They can give you language for a problem you were feeling but had not named yet.
Just do not confuse a good discussion with a working system.
The question after every AI board session should be:
**What minutes did this leave behind?**
If the answer is “none,” the board did not meet.
It just talked.
And honestly, we already had too many meetings for that.
Try this Monday
Run one AI board session on a real decision, not a fake productivity exercise.
Good candidates:
- Should we launch this offer?
- Which audience should this piece target?
- Is this feature worth building now?
- What should our homepage claim be?
- Which project should I kill this week?
- Should this idea become a post, a product, or a private note?
Give the board three to five roles max. More than that and you are basically hosting a wedding reception in your prompt.
Then end with the five-line minutes:
Decision:
Why:
Dissent:
Next action:
Review trigger:
Save it somewhere you will actually see again.
The win is not that AI helped you think once.
The win is that next time, you do not have to start from zero.
If you try it, reply and send me the five-line version. I am especially curious what shows up in the “dissent” line, because that is usually where the useful little monster lives.
-zc
P.S. If you want more essays like this — AI workflows with actual memory, taste, judgment, and fewer fake productivity fireworks — subscribe to Don’t Feed The Algorithm.
Rabbit hole for the overcaffeinated
- Wikipedia: Minutes — boring on purpose, which is exactly why it is useful. Minutes are a technology for organizational memory.
- Atul Gawande on checklists — a classic reminder that simple written structures can outperform heroic competence.
- Anthropic: Claude Skills — useful context for the broader move from one-off prompting toward reusable procedures and memory-bearing workflows.
- Simon Willison: Things we learned about LLMs in 2024 — especially useful for thinking about why tool use, context, and workflow design matter more than prompt theater.
Previously in this rabbit hole:
