How to actually use AI in your business (without getting burned)
A plain-English walk-through for people who run a business, not a server room. About 15 minutes, start to finish. Diagrams first, words second.
You don't need a keyboard. Talk it through, let one AI write the job down, let another one build it.
See it in action
Brad talks to Claude (an AI assistant) by voice in the car. It reads his existing notes, writes a brief (a plain text file describing the job) and saves it to his Google Drive. Later, a dev agent (an AI given a specific job and instruction file) on a Raspberry Pi at home picks the brief up, hands it to the agent that builds web pages, and that agent builds and publishes the page. This page was made exactly that way.
- 🎙️Brad speaksin the car, Claude app
- 📄Brief in DriveClaude checks old briefs, writes a .md file
- 🖥️Dev Pi picks it up“scan the dev folder”
- ↪️Routed to the apps buildera request file is the hand-over
- 🔨BuiltHTML, diagrams, tests
- 🌐Livehowithinkitworks.bdotrad.me
Part 1: In the car voice · Claude app
Part 2: At the Pi typed · Claude Code on the dev Raspberry Pi
Later, at home. Brad's typos left in on purpose: the agent works out what he means.
Both conversations are trimmed and tidied from the real thing. Tool-call noise and private file IDs are removed.
The model never changes and never remembers you. What you control is the text you put in front of it.
How AI works
The model is the big trained neural network that does the “thinking”. You don't change it. Every time you ask it something, a bundle of text goes in, called the context, and text comes out. The thing you type is the prompt. That's the whole interface.
Swipe the diagram sideways to see it all →
Tap or hover any block in the diagram for a one-line explanation.
One manager agent and a few narrow specialists beat one do-everything assistant.
Agents and a PM
An agent is the same model, plus a specific instruction file, plus a specific job. “The summariser.” “The web developer.” “The reviewer.” A PM agent takes a request, breaks it into tasks, hands each to a specialist, then checks and assembles the results. A person stays in charge of the goal and signs off the result.
Click an agent to see the kind of instruction file it runs on.
Have one agent read each PDF once and write a short summary. Every other agent reads the summaries.
The PDF trick: summarise once, read many times
Swipe the diagram sideways to see it all →
How much text gets read? illustrative only
Assumes a PDF is about 15,000 words, a summary about 600, every task reads the whole folder (without) or all the summaries plus two source PDFs for exact detail (with), and the summaries are written once. Real numbers vary a lot. The shape doesn't.
What a summary file looks like
A made-up example. The point is the structure, and the page references back to the source.
# Geotechnical Investigation – Lot 12 Example Rd (made-up) - **Source:** `source/geotech-lot12-rev-B.pdf` (42 pages) - **Summarised:** 2026-11-19 by the summariser agent. Check before relying on it. ## Purpose Site classification and footing advice for a two-storey dwelling. (p. 3) ## Key facts - Site class **M** to AS 2870 (moderately reactive clay). (p. 11) - 6 boreholes, 3.0 to 4.5 m deep. Groundwater not met. (p. 7, Table 2) - Fill up to 0.8 m near BH4 – uncontrolled. (p. 9) ## Numbers | Item | Value | Page | |---|---|---| | Characteristic surface movement ys | 30 mm | 11 | | Allowable bearing, natural clay | 100 kPa at 0.6 m | 14 | ## Recommendations / decisions - Stiffened raft, or piers through fill near BH4. (p. 15) - Geotech to inspect footing excavations. (p. 16) ## Open questions - Retaining wall on east boundary not covered. (p. 2, Limitations) ## Where to look for exact detail Borehole logs: Appendix A (p. 22–33). Lab results: Appendix B (p. 34–40).
AI drafts, a person checks. Keep sensitive data out of free tools.
Doing it safely
One-page AI policy (template)
[Company] AI use policy – version 1, [date] 1. Approved tools Staff may use: [tool, business plan] with their work account. Anything else needs approval from [name/role] first. 2. Never put in - Client or customer personal information - Financial records, pricing, tenders, contracts - Anything under a confidentiality agreement - Passwords, access codes, security details 3. Always check AI output is a draft. The person who uses it is responsible for it. Calculations, designs, advice and certifications are checked and signed off by a qualified person, as they are now. 4. Be open Say when AI helped produce a document that goes to a client if the client expects to know. 5. Keep records Save the instructions and outputs for work that matters. 6. Report problems Data put in by mistake, or wrong output that got out: tell [name/role] the same day. Reviewed every [6] months by [name/role].
General information only, not legal advice. Adapt it to your own obligations.
The value compounds in your reusable instruction files and notes, not in clever one-off prompts.
Getting the best out of it
- Write clear briefs. Purpose, audience, constraints, what “done” looks like, and an example.
- Give it good source material. Clean text beats scanned images.
- Build a small library. Reusable instruction files and notes. This is where the compounding value is.
- Iterate. First draft, critique, second draft. Ask it to critique its own work.
- Split the jobs. Separate agents for separate jobs, not one giant prompt.
- Let it ask first. When the task is vague, tell it to ask you questions before it starts.
- Make corrections stick. Correct it, then save the correction into its instruction file so it doesn't happen again.
- Use your dead time. Dictating a brief while driving or walking turns travel into progress. See Tab 1.
Where to start on Monday
- Pick one repetitive, document-heavy job: summarising incoming reports, drafting meeting minutes, tidying specs.
- Write a one-page policy (template in Tab 5).
- Run one summariser pass over a small folder of non-sensitive PDFs, and review the output.
- Add a second agent that uses those summaries.
- Review after two weeks: time saved, errors found, next step.
Got a question? Ask it here. Brad answers on the page.
Three agents, three plain text files handed between them, and nobody typed any code.
How this page was built
Written by the agent that built it (Claude Code, working the apps project on the dev Raspberry Pi), from its own request files and git history.
- 🎙️Brief dictated19 Nov, morning · in the car
- 🖥️Dispatched on the Pi10:07 · request file written
- 🤖Picked up10:07 · apps builder claims the project
- ➕Part 2 arrives10:10 · “just send the lot”
- 🔨Built10:12 – 10:19 · first commit
- 🌐Published10:19 · every tab link checked
- 🏠Own address10:50 · part 3: howithinkitworks.bdotrad.me
- ❓Ask a question11:11 · part 4: question tab, static transcripts
What it read first
Decisions it made by itself
index.html for the words, style.css, app.js. No libraries, no web fonts, no tracking. Nothing to install, nothing to break.#hash says which tab is open, so howithinkitworks.bdotrad.me/#pdf-trick opens straight to Tab 4. Arrow keys move between tabs.What it couldn't do
Files
public/ai_skool/ ├── index.html all eight tabs' words and the diagram frames ├── style.css light/dark colours, tabs, bubbles, print layout ├── app.js tabs + hash links, transcripts, diagrams, calculator ├── ask.js the Ask a question tab: form, answers, Brad's admin view ├── header.css the shared logo strip, copied in so the page stands alone ├── rat-logo.png the logo └── brief.txt the brief this page was built from
Commits
Mistakes and fixes
If you only watch one: the 8-minute 3Blue1Brown explainer.
Watch and read next
Plain links, nothing embedded. Each one was checked on YouTube on 19 Nov 2026. Lengths are rounded.
Provider and regulator pages on data privacy
Stuck on something, or not sure it applies to your business? Ask. Brad answers here.
Ask a question
No login needed. Questions aren't shown on the page until Brad has answered them, and your name and contact details are never shown.
Answered questions
Loading…
