I spent $1,200 in two days on GrokBot (and it was worth it)

I spent $1,200 in two days having GrokBot click around a browser.

It was slow, brittle, and comically expensive.

I would do it again.

One of the tools I needed did not have an API I could use, so I gave GrokBot computer use and let it brute-force the workflow. Two days later, I knew the workflow worked.

The $1,200 bought the proof. Now we can make it cheap.

This fits something I have been talking about for years: Matzner’s Hierarchy of Leverage.

1. Owner

2. Expensive talent

3. Less expensive talent (often, not always global)

4. AI and automation (tech)

Systems push work down the pyramid without letting quality fall apart. I wrote in Spaghetti Don’t Scale that you cannot push work down until you define it first.

Something Zaki (Sagan’s President of AI) said in a meeting today made me realize the bottom rung (tech) has its own hierarchy:

1. Powerful model with computer use

2. Cheaper model

3. API call

4. Deterministic code

The same rule applies. Keep pushing the work down.

Say you run 20 service vans.

Every Friday, someone logs into the fuel-card portal, downloads the transactions, hunts down missing receipts, and updates a spreadsheet for the bookkeeper.

To prove the workflow, give the whole mess to the smartest model available (think fable).

Let it click around. Let it read the receipts, match the charges, and flag anything weird. It will cost too much and occasionally do something stupid, but you will learn what the finished system needs to do.

I have written about global talent as pickup trucks and automation as train tracks. During problem discovery, you do not know where the tracks belong yet. Use the truck first! People or expensive agents.

Once the workflow works, CLEAN IT UP.

Pull the fuel transactions through an API. Add rules for missing receipts and duplicate charges. Use a smaller model for the handful of receipts the rules cannot read. Send the strange stuff to a human.

You do not need the smartest model in the world deciding whether a $74.18 Shell receipt matches a $74.18 fuel-card charge.

I once wrote that I wanted our API costs to go up. Every dollar of useful API spend meant more repeatable work had moved out of someone’s hands.

The same cleanup applies to model costs. Once the job stops requiring judgment, stop paying for judgment!

During DISCOVERY, start expensive and work your way down.

In PRODUCTION, reverse the order.

Start with the least expensive model that has a reasonable chance of completing the task correctly. Check its work. Escalate to a stronger “model” only when it fails. Escalation could be to a more expensive AI agent… or a person!

A lot of companies do the expensive part forever. They prove a workflow with the smartest model available (or human!!), then keep paying that model to perform steps that should have become API calls six weeks ago.

For every step in an workflow, ask:

Yallah Habibi,

Jon