The Beanz Brief
The day I switched the whole crew off
A crew of AI workers, a credit card bill, and no visible output. What we changed so it never happens again, and two rule changes you need before December.
Morning.
Last week’s issue was a win. This one is the failure that made the win possible, because if I only send you the wins you should stop reading.
26 August
By late August I had an AI crew running most of the operations of a small membership business: content, a directory, bots, bookkeeping prep. I was proud of it. Then I looked at the bill and the output side by side and could not match them. Credits were being spent, in volume, and I could not point to what they had produced that week.
So I switched it all off. Every worker. Reset to zero.
That is the sentence the vendor case studies never contain, and it is the most useful thing I can tell you about running AI in a business: the failure mode is not the machine doing something wrong. It is the machine doing a great deal of something, invisibly, on your card.
What was actually wrong
Not the models. Three plain operating problems:
- No definition of done. Workers were given goals, not deliverables. A goal can be worked on forever.
- No cost line per job. We tracked spend per month, not per task. A month is too late.
- No “stop and ask” rule. When a worker hit an ambiguous decision, it tried harder instead of asking. Trying harder is expensive.
What we changed
- Every job now ends in a named file or a named change, or it is not a job.
- Every job writes a one-line status when it changes state: working, needs a decision, blocked, done. A human reads those lines, not the transcripts.
- A decision that belongs to a human stops the worker until the human answers, and the answer is recorded in the human’s own words. No “always allow”.
- Spend is tracked per job, not per month, and it sits next to the thing the job produced.
- Nothing is published, sent, filed or paid unless a person approved it or it is covered by a standing rule they wrote and can switch off. The crew drafts; a person remains accountable.
Since the reset the same crew has done, among other things, a year of GST prep in a day (last issue) and the research behind this newsletter. The difference is not smarter AI. It is that I can now tell you what each job cost and what it produced.
The rule for you
If you have anyone in your business using AI tools in anger, ask one question this week: what did it produce, and what did it cost, last Tuesday? If the answer takes more than a minute, you have the August problem. It is fixable in an afternoon and it is mostly a definition-of-done problem, not a technology one.
Two things that changed (verified 21 October 2026)
1. Australia: from 10 December 2026 some APP entities must describe automated decisions in their privacy policy. The duty applies where an APP entity arranges for a computer program to use a person’s personal information to make a decision, or do something substantially and directly related to one, that could reasonably be expected to significantly affect the person’s rights or interests (credit, access to a service, hiring). The policy must say what information is used and which decisions or actions are involved. Infringement notices apply. Who it affects: APP entities using personal information in automated customer eligibility, booking triage, or recruitment decisions that could significantly affect a person. What to do: confirm whether the duty applies, then list the decisions your tools touch and update the policy. Cost: an hour with the list, a lawyer’s half hour if you want it checked.
2. The vendors moved into “small business” mode this year, in US dollars. Anthropic’s small-business offer in May, Meta’s business agent in WhatsApp and Instagram in June, Xero’s agents in July, MYOB’s AI tools and assistant connector in August. What it means: you can buy every one of these by card today. None of them comes with anyone who sets it up, connects it to your actual systems, or watches it. That last-mile gap is the whole reason this newsletter exists. What to do: nothing yet, unless one of them connects to a system you already pay for; then test it on one workflow with a definition of done (see above).
One number
Days since the reset with every job carrying a cost and a deliverable: 55.
Next week
What a clinic’s front desk looks like when the phone answers itself and the humans stay in charge. A teardown, with the parts that didn’t work.
Beanz
