I gave Grok Bot a job.
Then I gave it colleagues.
I ended up with a Chief of Staff, CFO and CMO sharing context in the same workspace.
I wanted to see whether this could actually work like a small AI leadership team.
Here’s what happened 👇
I changed 4 Claude settings once and got more value than I expected. Tiny setup choices matter: what it remembers, how it writes, what context it starts with. We spend a lot of time comparing models and surprisingly little time configuring the one we already pay for.
The real test of an AI automation is often month 2. I built one for recurring invoices, and the useful bit wasn't generating invoice #1. It was remembering what should stay the same, what should change, and what should disappear next month.
AI is very good at making messy data look finished. That is not the same as making it correct. I’d rather use it to clean, classify and flag weird rows, then keep the original data beside it. Faster cleanup, same source of truth.
I used AI as a second pair of eyes on a P&L. Not to “do finance” for me — just to flag what moved, what looked odd and what deserved a question. That is a much better role for AI than pretending the spreadsheet stopped needing a human.
An AI agent submitted a false murder tip because nobody explicitly told it not to submit forms. The useful lesson is not “AI went rogue.” It is that prompts describe the rules; permissions enforce them. If an agent should not press Send, don’t give it the button.
One sales-call transcript can become the follow-up email, CRM notes and next-step list. That is more useful to me than transcription alone. Recording the meeting is easy. Turning it into the admin you were going to do anyway is the better AI workflow.
I asked AI to diagnose a fictional 40-lead sales funnel. The useful finding wasn’t “get more leads.” It was that too many existing leads disappeared before a real sales conversation. Sometimes the cheapest growth strategy is fixing the funnel you already paid for.
A 3-email welcome sequence sounds like a writing task. It is really a sequencing task: what does someone need to understand first, second and third? AI can draft the emails quickly. Deciding the order is still the useful bit.
A 13-week cash flow forecast is a good AI use case because the hard part is not the spreadsheet. It is forcing yourself to make assumptions explicit. AI can build the model quickly. You still have to decide whether those assumptions are remotely believable.
Google may not need Gemini to win every model comparison. If Workspace becomes the place where AI agents get their identity, files, permissions and jobs, distribution may matter more than having the smartest model.
@NightOrders Interesting. Mine are sharing the same workspace, so they can see the same files and context. I hadn't considered separating ownership by folder like that. Have you found that it reduces agents stepping on each other or picking up context they shouldn't?
I gave Grok Bot a job.
Then I gave it colleagues.
I ended up with a Chief of Staff, CFO and CMO sharing context in the same workspace.
I wanted to see whether this could actually work like a small AI leadership team.
Here’s what happened 👇
I've started making AI argue with itself when I'm researching something.
One looks at the customer. One at the economics. One tries to kill the idea.
The last one is often the most useful.
Every leadership team needs a professional hater.
I gave AI an SOP and asked it to turn it into training.
Instead it basically told me my SOP was shit.
Missing exceptions. Contradictory steps. Things that only made sense if you already knew the job.
Slightly rude. Also correct.
Vibe coding has introduced a new business problem:
Before: “Do we really need to pay $30/month for this software?”
Now: “Do I really need to build this myself?”
I am not sure this is progress.
The funny thing about building an AI “team” is that giving every bot a fancy job title achieves basically nothing.
“You are the CFO” isn't a workflow.
Give it one boring job, correct it, then automate it. Much less sexy. Much more useful.
My AI trust hierarchy:
Draft an email → sure
Summarise a meeting → go ahead
Update my CRM → hmm
Send something to a client → hold on
Move money → absolutely fuck off
“Human in the loop” is doing a lot of work in that last category.
I spent a lot of time thinking AI automation meant elaborate agents.
Increasingly I think the killer use case is:
“Please remember how we did this annoying thing last month.”
Less Iron Man. More competent admin assistant.
Probably more useful.
AI found the most common complaint in some customer reviews: slow delivery.
Then I asked which complaints did the most damage. Different answer: billing errors and bad support.
50 “bit slow” reviews ≠ 5 “you charged me twice” reviews.
Loud and expensive aren't the same thing.