Skip to main content
← All guides

Guide 9

Which processes in your business could AI take over?

The guide from the video, with the full AI instruction ready to copy.

1. What you'll end up with

A list of the recurring tasks in your business that an AI could take over or prepare for you, ranked by payoff and risk. For the best ones you get a ready project brief: what exactly should happen, who checks it, what could go wrong. And for your number one, a first test you can run this week.

  • Who it's for: business owners and the self-employed who feel that too much runs across their desk. No coding needed.
  • Time: set aside a quiet half hour. The AI asks you 25 questions at most, one at a time. Short answers are fine.
  • What you need: an AI chat such as Claude (claude.ai) or ChatGPT (chatgpt.com). Both have a free plan. Claude's free plan has usage limits that, according to its pricing page, reset on a rolling five-hour window. If the conversation stops halfway, pick it up later in the same chat.

2. Step by step

  1. Open your AI chat and sign in. In the browser on your computer (claude.ai or chatgpt.com) or in the app on your phone.
  2. Before you start, stop your chats from being used for training.
    • Claude: go to claude.ai/settings/data-privacy-controls and switch off "Model Improvement". Or use an Incognito chat: Anthropic says those aren't used for training.
    • ChatGPT: your profile → Settings → Data controls → switch off "Improve the model for everyone". Or use a Temporary Chat: OpenAI says those aren't used for training and are kept for up to 30 days.
  3. Copy the AI instruction from section 3 into your notes app and fill in the "What you already know about me" block. Replace everything in [SQUARE BRACKETS] with your own details. Leave out what you don't know; the AI will ask.
  4. Open a new chat and paste the whole instruction as your first message. Send it.
  5. Answer the questions. Keep it short and honest. "Not sure" is a perfectly good answer. On your phone, you can use your keyboard's dictation and just talk.
    • "skip" skips a question.
    • "done" ends the interview and you get the results with whatever is there so far.
  6. Read the results. Check the table first: are your time figures right? Is a task missing? Tell the AI in the same chat and it will redo the numbers.
  7. Save the result. Paste it into a document. You'll need it for the next steps.
  8. Carry on with number one using the follow-up instructions in section 4, in the same chat.

3. The AI instruction, ready to copy

Copy everything inside the box, fill in the placeholders, and paste it as the first message in a new chat.

You're my advisor on how small businesses run day to day. Together we'll work out which recurring tasks in my business an AI could take over or prepare. Then you rank them by payoff and risk and write a project brief for the best ones, so I can get started.

== WHAT YOU ALREADY KNOW ABOUT ME ==
(Ask about anything left empty. Don't ask again about anything filled in.)
- Industry and what we sell: [FOR EXAMPLE: ACCOUNTING PRACTICE, BIKE SHOP WITH A WORKSHOP, AGENCY]
- Who works here (number and roles): [FOR EXAMPLE: ME, 2 FULL-TIME, 1 PART-TIME]
- How customers find us: [FOR EXAMPLE: REFERRALS, GOOGLE, WALK-INS]
- Software we use every day: [FOR EXAMPLE: EMAIL, POINT-OF-SALE SYSTEM, EXCEL, WHATSAPP]
- Working days per week: [FOR EXAMPLE: 5]
- What eats most of my time right now: [IN YOUR OWN WORDS]

== HOW TO RUN THE INTERVIEW ==
1. Explain in two sentences what we're about to do. Tell me: 25 questions at most, short answers are fine, "not sure" is allowed, "skip" skips a question, "done" takes us straight to the results. Then ask the first question.
2. Always ask exactly ONE question per message. No forms, no lists of questions. With each question, give two or three short example answers from my industry so I can answer faster.
3. Order:
   a) Ask about whatever is missing in the block above.
   b) Ask me: "Walk me through a normal week: which tasks keep coming back?"
   c) Then ask about at most three areas I haven't mentioned yet, from this list: enquiries and sales · quotes, orders, invoices · customer contact afterwards (appointments, reminders, reviews) · marketing · purchasing and stock · team and coordination · bookkeeping and admin.
   d) Pick up to 6 tasks that take the most time or cause the most frustration. Tell me which ones you picked and ask whether one is missing.
   e) For each of these tasks, ask exactly two things, each in its own message:
      - "How often, and how long each time?" (for example "daily, 15 minutes")
      - "Where does the information for it come from, and what happens if a mistake slips through and nobody notices?"
4. If an answer is unclear, ask once more. If I contradict myself, point it out. Never guess.
5. Keep it short. No summary after every answer, no compliments.

== HOW TO SCORE ==
For each task:
- Hours per month = how often x how long, calculated only from my answers. Use 4.3 weeks per month; "daily" means on each of my working days. Show the calculation. If a figure is missing, write "missing", never an estimate.
- Time (1-3): under 2 hrs/month = 1 · 2 to 8 hrs = 2 · over 8 hrs = 3
- Same rules every time (1-3): 3 = almost always done the same way · 2 = mostly the same, with exceptions · 1 = different every time, needs experience and gut feeling
- Information in writing (1-3): 3 = already in writing (email, form, spreadsheet) · 2 = partly · 1 = only in someone's head, on paper or on the phone
- Cost of a mistake (1-3): 1 = harmless, easy to fix · 2 = annoying (unhappy customer, rework) · 3 = expensive or legally sensitive (money, contracts, health, staff, tax)
- Score = Time + Rules + Information + (4 minus Cost of a mistake). Maximum 12.
- Personal data (names, phone numbers, addresses, health, bank details, staff records): mark the task "⚠️ data". It doesn't lower the score, but it goes into the risks.
Category:
- "Start here" = score 9 or more AND cost of a mistake 2 or lower
- "Later" = score 6 to 8, or 9 and more with cost of a mistake 3
- "Not for AI" = under 6, or Rules 1 together with Cost of a mistake 3
On a tie, the task with more hours per month comes first.
How often and how long must come from me. You rate Rules, Information and Cost of a mistake yourself from my answers, and in the last section you say where that is your own judgement, so I can correct it.
Every task starts like this: the AI prepares, a person checks and approves. Nothing goes out unchecked to customers, authorities or money.

== FIXED RULES ==
- Don't invent numbers. Every number comes from me or is calculated from my answers. Never state time saved, costs or prices yourself: put a placeholder there for me to fill in.
- Only tasks I mentioned go into the table. At most two ideas of your own, listed separately underneath, labelled "Your idea, not from our conversation".
- Don't claim that a specific piece of software can do something. Describe what's needed (for example "a form that collects the details") and mark open points "to be checked".
- No promises like "fully automatic" or "never again". The AI prepares, a person decides.
- Don't answer legal questions (data protection, contracts, tax, employment law). Mark them "check with an expert".
- Write so that someone without any tech background understands every word. Use a technical term only when you have to, with an explanation of five words or fewer.
- If I paste personal data about customers or staff, tell me once that it doesn't belong here.

== WHAT THE RESULTS LOOK LIKE ==
# Your AI roadmap
## 1. Your business in three sentences
## 2. All tasks at a glance
Table with these columns: No. | Task | How often x how long | Hrs/month (calculation) | Time | Rules | Information | Cost of a mistake | Score | Data | Category
Sorted by score.
## 3. Your ranking
The top five, one sentence each on why they're there.
## 4. Project briefs
For the top three in the ranking, in full. If one of them is in "Later", say what extra checking it needs. Each with exactly these points:
- Goal: (one sentence)
- Trigger: (what starts the process)
- Inputs: (what the AI needs, and where it comes from)
- Steps: (3 to 6 steps, each saying who does it: AI, you or the team)
- Output: (what's ready at the end)
- Who checks and approves:
- Time today: (my figure, hrs/month) · Time saved: [YOUR ESTIMATE AFTER A TWO-WEEK TEST]
- Risks and data protection:
- How you'll know it works: (2 or 3 points that can be checked)
- What you need: Level A = just an AI chat · Level B = an AI chat plus your templates and examples · Level C = software that has to be connected (someone has to set that up)
For places 4 and 5, a short profile of five lines each: Goal · Trigger · Who checks · Biggest risk · What you need.
## 5. Your first step this week
For number one: a ready instruction for the AI that lets me run the process once by hand in the chat. Placeholders in [SQUARE BRACKETS] for my examples. It has to work today, with just an AI chat and nothing to install.
## 6. What the AI assumed and what's still open

== BEFORE YOU SEND THE RESULTS ==
Check every point. If one isn't met, fix it first:
- Every task in the table comes from my answers.
- Every number comes from me or is calculated from my answers, with the calculation shown.
- No time saved, costs or prices from you.
- Scores and categories follow the rules above exactly.
- Every task involving personal data is marked and has a data protection point in its project brief.
- The top three have all ten points, places 4 and 5 a short profile.
- The first step works today with just an AI chat.
- No technical term without an explanation.
End with one line: "Self-check: all points met", or which point is open and why.

4. Follow-up instructions

Keep writing in the same chat, so the AI already knows your business.

a) Turn number one into a build plan

Turn number one into a detailed build plan. Describe each step so that someone on my team could do it without asking: what comes in, what the AI does with it, what a person checks, what goes out. Write the templates and texts the AI needs as well, with [PLACEHOLDERS] for my details. Set up a two-week test: what I write down each day, and how I decide at the end whether it's worth it. Don't invent numbers.

b) What do I need to connect?

For number one: which information has to move from where to where for this to work? Make a list: where it comes from, where it goes, who does it today. For each point, tell me whether an AI chat alone can handle it or whether software needs to be connected. Don't claim a specific product can do it. Instead, write down the questions I should ask my software provider.

c) The instructions for your team

Write one page of instructions for my team for number one: what it's for, when to use it, what to do step by step, what the AI does and what a person always checks themselves, what never goes into the AI chat, and who to ask when something looks off. Plain language, speaking directly to the reader.

d) Have a second AI check the result (the checking rule from the video) Open a new chat and paste this:

You're checking another AI's work.
My task was: [TASK].
The result: [FOR EXAMPLE, THE AI ROADMAP OR THE BUILD PLAN].
Check:
1) What's in the result that didn't come from my task, like numbers, names or promises?
2) What from my task is missing?
3) What wouldn't someone without a tech background understand?
Give me the list first, then an improved version. Don't add anything new.

5. Example: what it looks like

FICTIONAL EXAMPLE. This business doesn't exist. All names and figures are made up, only to show how the instruction works.

The filled-in block:

- Industry and what we sell: [LEFT EMPTY ON PURPOSE, SO THE AI HAS TO ASK]
- Who works here (number and roles): me (owner, sales and office), 2 mechanics, 1 Saturday helper
- How customers find us: walk-ins, Google, referrals, companies with a bike-to-work scheme
- Software we use every day: email, point-of-sale system with a workshop calendar, WhatsApp Business, an Excel stock list
- Working days per week: 6
- What eats most of my time right now: the phone never stops ringing about repair bookings, and I'm doing paperwork in the evenings

Excerpt from the results. This is what came out of our test run with Claude on 15 October 2026. A second AI played the fictional shop owner, working from a fact sheet with made-up figures. Like a real owner, it was sometimes vague and contradicted itself, and the AI caught it. After 25 questions, this was the result. Shortened, otherwise unchanged.

2. All tasks at a glance

No. Task How often x how long Hrs/month (calculation) Time Rules Information Cost of a mistake Score Data Category
2 Customer messages ("your bike's ready", routine replies) 6 messages a day x 3 min, 6 working days 6 x 3 = 18 min/day; x 6 days = 108 min/week; x 4.3 = 464 min = 7.7 hrs (evening part counted inside this, to be confirmed) 2 3 3 2 10 ⚠️ data Start here
5 Scheme quotes and forms 4 a month x 45 min 4 x 45 = 180 min = 3.0 hrs 2 3 3 3 9 ⚠️ data Later
1 Repair booking requests Phone: 5 calls a day x 4 min, Mon–Fri 5 x 4 = 20 min/day; x 5 days = 100 min/week; x 4.3 = 430 min = 7.2 hrs (phone only; Saturday and email/WhatsApp requests: missing) 2 2 2 2 8 ⚠️ data Later
3 Weekly parts ordering Weekly, 1.5 hrs 1.5 x 4.3 = 6.5 hrs 2 2 2 2 8 none mentioned Later
4 Repair quotes 3 a week x 20 min 3 x 20 = 60 min/week; x 4.3 = 258 min = 4.3 hrs 2 2 2 3 7 ⚠️ data Later
6 Month-end receipts and invoices Monthly, 3 hrs 3.0 hrs 2 2 2 3 7 ⚠️ data Later

(scoring notes and ranking cut)

Task 2: Customer messages

  • Goal: The AI drafts the "your bike's ready" messages and replies to routine questions, so you only check and send.
  • Trigger: A repair job is marked finished in the workshop calendar, or a routine customer question arrives.
  • Inputs: For each finished job: first name, bike type, job number and pick-up hours, which you copy from the POS calendar into the chat. Also 3 to 5 of your usual messages as examples, with names removed.
  • Steps:
    1. You: copy the finished jobs (first name, bike, job number) into the chat once or twice a day.
    2. AI: writes one draft message per job in your tone.
    3. You: compare each draft with the calendar (right name, right bike).
    4. You: send the message yourself by WhatsApp or email.
    5. AI: for routine questions, drafts a reply from the facts you give it (opening hours, prices you supply).
    6. You: check the reply and send it.
  • Output: A list of ready-to-send drafts, one per job, each showing the job number.
  • Who checks and approves: You. Later, the Saturday helper could do this after you've shown them how. Nothing is sent by the AI.
  • Time today: 7.7 hrs/month (18 min a day, to be confirmed) · Time saved: [YOUR ESTIMATE AFTER A TWO-WEEK TEST]
  • Risks and data protection: A wrong bike or customer in a message is annoying (you rated it that way). Check each draft against the job number. ⚠️ Customer names and phone numbers are personal data. Use first names and job numbers only, never phone numbers or addresses. Whether and how you may use customer data in an AI chat: check with an expert. What the AI provider does with your entries: to be checked.
  • How you'll know it works:
    • No message goes to the wrong customer during the two-week test.
    • Most drafts need only small changes (keep a tally, your target: [YOUR TARGET]).
    • The daily time for this task is measurably below today's 18 minutes.
  • What you need: Level B (an AI chat plus your example messages).

5. Your first step this week

Copy this into an AI chat, fill in the square brackets, and run it once with 2 or 3 finished jobs. Please don't paste surnames, phone numbers or addresses.

You are my assistant at [SHOP NAME], a bike shop and repair workshop. You write short, friendly "your bike is ready" messages for WhatsApp. You only prepare drafts. I check and send them myself.

Here are 3 examples of how I normally write:
1. [EXAMPLE MESSAGE 1, first name only]
2. [EXAMPLE MESSAGE 2]
3. [EXAMPLE MESSAGE 3]

Our pick-up hours: [PICK-UP HOURS]
Anything we always mention: [E.G. "please bring your lock key", "payment by card or cash"]

Here are today's finished jobs:
- Job [NUMBER 1]: first name [FIRST NAME 1], bike [BIKE TYPE 1], work done [SHORT NOTE 1]
- Job [NUMBER 2]: first name [FIRST NAME 2], bike [BIKE TYPE 2], work done [SHORT NOTE 2]
- Job [NUMBER 3]: first name [FIRST NAME 3], bike [BIKE TYPE 3], work done [SHORT NOTE 3]

Please:
1. Write one draft message per job, in my tone, no longer than [NUMBER] sentences.
2. Start each draft with the job number so I can check it against my calendar.
3. Don't invent prices, dates or details I haven't given you. If something is missing, write [MISSING: ...] instead.
4. After the drafts, list any job where you were unsure.

Then compare each draft with your POS calendar, change what you want, and send. Write down how many minutes it took, so you have a figure for the two-week test.


6. How you'll know it's working

  • The table lists your tasks, in your words, and nothing you never said.
  • Every hours figure comes with the calculation from your answers.
  • Time saved shows a placeholder, not a number from the AI.
  • Number one is something you can try this week without hiring anyone.

Common problems and what to do:

What happens What you type
The AI sends you ten questions at once. "One question per message, please, as the instruction says."
It writes "this saves you 10 hours a week". "Where does that number come from? Take it out and use the placeholder."
The table has a task you never mentioned. "I didn't mention number X. Move it to your own ideas."
The conversation stops halfway. Come back later and type "continue" in the same chat. On a free plan, that's often the usage limit.
You've had enough questions. Type "done". You get the results with what's there, and open points go into the last section.
Everything lands in "Later". Not a mistake. Then ask: "Which of these tasks could I change to make it less risky?"

7. Limits and data protection

  • The AI only knows what you tell it. The ranking is only as good as your time figures. Honest beats flattering.
  • The scores help you sort, they're not a guarantee. Whether a process really saves time only shows up when you test it on real cases.
  • This conversation doesn't need any customer data. Describe how you work, not who you work for. What stays out of the chat:
    • names, addresses, phone numbers and email addresses of customers or staff
    • health data, bank details, ID numbers, salaries
    • contracts, passwords, logins
    • anything covered by professional confidentiality (accountants, medical practices, law firms)
  • If you later want to work with real customer data, check first with your data protection officer or an expert whether and how that's allowed with the AI provider you've picked. If you're in the EU or UK, that means GDPR. This isn't legal advice.
  • Anonymise your examples: instead of "Mr Smith, 5 High Street", write "Customer A". That's all the AI needs.

8. Next step

If you want to build your number one with Claude yourself, our AI Blueprint walks you through the setup step by step.

Go to the AI Blueprint
Which processes in your business could AI take over? | Guide by AILoopwise