What counts as busywork
Busywork is any task where the thinking is done but the typing is not. Transcribing a meeting you attended. Renaming a hundred files to match a pattern you already decided on. Copying form submissions into a spreadsheet. Summarising a document you will still need to read properly later. These are the jobs where machines earn their keep, because a human can check the output faster than they could produce it.
The AI desk files tools against that test. If verifying the result takes longer than doing the task by hand, the tool does not make the issue, however impressive the demo looked.
Logged this season
Dispatch 118-05 · Transcription
Whisper
OpenAI's open-source speech recognition model, released in 2022, now sits inside many transcription apps and can also run on your own machine. Strong on accented English and many other languages. Catch: it can confidently invent words during long silences, so skim the quiet bits.
Dispatch 117-02 · Local models
Ollama
A simple way to download and run open language models on your own computer. Nothing leaves the machine. Catch: smaller models are noticeably less capable, and a laptop fan will tell you when it is working hard.
Dispatch 116-07 · Workflows
n8n
A visual workflow builder you can self-host, connecting apps through triggers and actions. Good for "when a form arrives, add a row and send a message". Catch: self-hosting means you are now the IT department.
Dispatch 116-01 · Workflows
Zapier
The hosted veteran of app-to-app automation, with thousands of integrations and AI steps added over time. Catch: task-based pricing climbs quickly once a workflow runs often.
Dispatch 115-04 · Phone
Apple Shortcuts
Built into iPhone, iPad and Mac. Chain actions such as resizing photos, logging water intake or texting your ETA. Free and already installed. Catch: sharing shortcuts with Android-using relatives is not happening.
Dispatch 114-06 · Models
Hugging Face
The largest public hub for open models and datasets, with browser demos called Spaces. Useful for trying a niche model before installing anything. Catch: quality varies wildly between uploads.
The verification habit
Language models produce fluent text whether or not it is correct. That is the single most important fact about them. The working habit the desk recommends is simple: use AI output as a first draft, never as a source. If a summary cites a figure, find the figure in the original. If it names a law, a person or a date, check one of them. If the first check fails, assume the rest might too.
Transcription and translation deserve the same treatment, especially for names, numbers and anything with legal weight. A machine-made transcript is an excellent index to the recording; it is not the recording.
Data in, data out
Before pasting anything into a hosted AI tool, ask three questions. Does the service keep my input? Does it use my input to train future models? Can I switch that off? Most major providers now publish answers and offer settings, but defaults differ between free and paid tiers and between consumer and business accounts. Contracts, medical notes, customer lists and passwords should not be pasted into a consumer chatbot at all. If a task needs sensitive data, a local option such as Ollama or an on-device Whisper build keeps it on your hardware.
A small automation checklist
- Write the task down in one sentence before choosing a tool.
- Do it by hand three times and time it. If it takes under a minute, leave it alone.
- Start with what you already own: Shortcuts, your email client's rules, spreadsheet formulas.
- Add a failure alert. Silent automations fail silently.
- Review every automation each quarter and delete the ones you forgot existed.