We’ll send a password for the resource area and regular UK-focused AI updates. Unsubscribe anytime.

The difference between “thinking” jobs and “recipe” jobs

This is the distinction that matters most.

Thinking jobs need judgement. The input is different every time.

• “Read this angry email and suggest a reply that calms things down”
• “Look at these customer reviews and tell me what people are most unhappy about”
• “Summarise this contract and flag anything unusual”

AI is genuinely good at these. Code usually isn’t.

Recipe jobs follow the same steps every time. Only the names and numbers change.

• “Take this spreadsheet and remove duplicate rows”
• “Add up all sales by product and create a summary”
• “Find every customer who hasn’t ordered in three months”

These are exactly the kind of jobs computer code has always been good at — long before AI existed.

Ask yourself:

“If I did this task 50 times, would the instructions change each time, or just the details?”

If it’s just the details, code might save you money.

Where does the saving actually come from?

There are two quite different ways you might use AI, and they cost very differently.

The first is a subscription service like ChatGPT. You pay a flat monthly fee and ask questions. You don’t see a separate charge for each one — but heavy repetitive use can push you towards a higher tier with “fair use” limits.

The second is using an AI model through an API. That’s where your own software sends information to an AI service and receives an answer. API usage is normally metered — you pay according to how much you use. OpenAI, for example, currently publishes API prices based on tokens, with different rates for different models. Its API is billed separately from ChatGPT subscriptions.

That distinction matters.

Imagine you have a piece of work you need to perform 500 times a month. If every job requires an AI API call, you’re potentially paying for 500 AI interactions.

But suppose the task is simply:

• open a CSV file
• remove duplicate rows
• sort the information
• calculate totals
• produce a new spreadsheet

There’s no real reasoning involved once you’ve decided what the rules should be.

You could ask AI to write the Python code that does all of that. After you’ve tested it, you can run the Python program 500 times without asking an AI model to work out how to perform the task each time.

That’s where the saving comes from.

How do you actually get AI to write the code?

This is the part that surprises most people: you don’t need to learn to code to get this benefit.

You describe what you want in plain English. AI writes the code. You test it. If it doesn’t work, you show AI the error and ask it to fix the problem.

Python is particularly useful for this sort of small business automation. It’s free, it’s widely used, and AI tools have had a huge amount of practice writing it, so the code tends to be more reliable than in less common languages. A Python script can simply be run from your computer.

Don’t start by saying “Write me a Python program.” That’s too vague.

Start with the job. Tell the AI:

• what you currently do
• what files you use
• what the input looks like
• what you want the output to look like
• how often you perform the task
• what should happen when something goes wrong
• what you don’t want the program to do

For example:

“Every Friday I download a CSV file from my booking system. It contains booking date, customer name, service, price and booking status. I want a Python program that lets me select the CSV file, removes cancelled bookings, calculates total revenue by service and saves an Excel report. Don’t change the original file. If a required column is missing, stop and tell me exactly what’s wrong.”

That’s a much better starting point.

Then ask AI to explain the code in plain English:

“What does each part of this program do?”

“What could go wrong?”

“What information does this program send outside my computer?”

That last question matters particularly if the program connects to an AI service.

The clever bit: using AI and code together

You don’t have to choose between “code” or “AI”. You can use both.

Take a small clothing shop. Every Monday morning, the owner downloads a sales CSV from the till system.

A Python program could do all the predictable work:

• remove duplicate transactions
• calculate total sales
• work out percentage changes
• find best-selling and worst-selling products
• calculate average transaction value
• break down sales by day and category

Then — instead of you taking those results and manually putting them into ChatGPT — the Python program itself can send a short summary to an AI service through an API.

The AI could be asked:

“Look at these figures and explain the three most significant changes in plain English. Don’t invent reasons that aren’t supported by the data.”

The AI’s response can then be automatically added to the finished report.

So the whole process becomes:

Sales file → Python processes the data → Python sends the results to AI → AI explains the results → Python creates the finished report.

If you have 10,000 sales transactions, there’s little reason to send all 10,000 to an AI model simply to calculate the total. Python can calculate the total locally. You then send a much smaller set of figures to AI for interpretation.

Depending on the AI service and model you’re using, API calls are charged according to usage. So reducing the amount of information sent to AI can reduce the amount of AI processing you’re paying for.

The important idea:

Don’t think of this as:

“Python instead of AI.”

Think of it as:

“Python first, AI when needed.”

That’s a much more useful way to design an automated process. Let ordinary software do everything it can, then use AI for the part that actually needs AI.

Do you need to be technical?

No, though it does help to have someone in your corner who’s comfortable with a computer.

The simplest option is running a script on a desktop computer — either by double-clicking it or by scheduling it to run automatically at a set time. This suits internal tools nobody outside the business needs to touch.

For anything you or your staff need to interact with, a small web app is often more useful — and easier to build than most people expect. Tools like Streamlit let you turn a Python script into a simple web page with buttons, boxes to upload files, and results shown on screen, without needing to know web design. A Streamlit application can run locally on your computer and open in your web browser.

Instead of opening a command prompt and typing a Python command, you could have a little web page on your computer saying:

Weekly Sales Report

[Choose CSV file]

[Create Report]

You, or whoever helps you set it up, describe what you want the screen to look like, and AI can write that too.

What about putting the tool online?

You can do that too — Streamlit provides options for deploying applications, including its Community Cloud service.

But don’t rush into this.

If a program only needs to be used by you on your own computer, keeping it local may be simpler and safer.

Once you put an application online, you’ve introduced additional considerations around:

• accounts
• passwords
• access controls
• hosting
• backups
• software updates
• security
• personal information
• who can access the application

Streamlit itself points out that security is a shared responsibility when applications are deployed, and its deployment guidance includes handling secrets and private information securely.

So don’t put something online simply because you can.

What about customer and employee information?

This applies whether you’re using AI directly or asking it to write code for you. Before you paste customer information, financial data or anything confidential into an AI tool, check what that tool does with what you type. Some tools use your input to improve their models unless you turn that setting off. Some store it for a set period. Some don’t use it for training at all on paid business plans.

Ask the provider directly, or check their documentation, rather than assuming.

If you’re building code that will process customer data, that data protection responsibility doesn’t go away just because a machine, rather than a person, is doing the processing. The ICO says data protection law can apply to small businesses, including sole traders, because they commonly process personal information about customers and staff.

Under UK GDPR and the Data Protection Act 2018, you’re still the data controller, and you’re still accountable for how that information is handled, stored and secured — whether a human or a script is touching it.

For most small businesses this doesn’t rule out using AI to write code. It does mean thinking about where the data lives once the script is running, whether it needs to leave your computer at all, and whether anyone else needs access to it.

When is repeat AI use actually the right choice?

It’s worth being honest here: sometimes paying for AI every time is genuinely the better decision, and building code would waste your money rather than save it.

If a task happens rarely — say, drafting a redundancy letter once a quarter — there’s no meaningful saving from building code you’ll barely use.

If the task changes shape every time — like responding to unpredictable customer complaints — code can’t easily capture “the rules” because there aren’t fixed rules to capture.

If nobody in or around your business can check that the code is doing its job correctly — an unchecked script quietly making mistakes can cost you far more than the AI subscription ever would. The FSB’s most recent figures show that 55% of small firms now use AI, up from 20% in 2023, and 54% of AI-using small businesses point to AI hallucinations as a specific worry, alongside security concerns and unclear responsibility when things go wrong. Those same worries apply to code AI has written for you.

Untested code that quietly gets something wrong — say, a stock reorder script that miscounts — can cost you more in wasted stock or missed sales than the AI subscription fee you were trying to avoid.

Don’t build code for the sake of it.

A realistic look at the costs and savings

None of this is completely free, and it’s worth being clear-eyed about that before you start.

Getting AI to write your first script will likely cost you the price of whatever AI subscription or pay-as-you-go access you already have, plus some time — either yours or someone else’s — to describe what you want clearly and test that it works. If you need someone with more technical confidence to check or tidy up the code, that might mean an hour or two of a freelancer’s or IT support firm’s time.

Against that, weigh up what you’re currently spending, in subscription costs and in your own time, on the repetitive version of the task. A task that takes fifteen minutes of AI-assisted work five days a week adds up to well over an hour a day across a small team, every week, indefinitely. A script that does the same job in seconds, built once, keeps paying that time back.

Suppose a task takes you 20 minutes every day. That’s more than 120 hours a year. If it’s genuinely repetitive, a small program could be worth building.

But if you only perform the task twice a year, spending six hours creating and testing software probably isn’t sensible.

It depends entirely on your situation. A one-person business doing a task once a month probably won’t see a worthwhile return. A team running the same process daily very often will.

A simple decision rule

Use AI directly when:

• the task requires interpretation or judgement
• the information changes substantially each time
• you need a draft written or a document summarised
• you need language translated or adapted
• it only happens occasionally

Consider code when:

• the task is repeated at least once a week
• the rules are predictable and don’t change
• the same inputs produce the same type of output
• you’re moving or cleaning data, calculating figures, or producing standard reports

Consider a hybrid when:

• Python can prepare the information
• AI can interpret the results
• the AI doesn’t need the entire original dataset
• you want to reduce the number or size of AI requests
• you need both consistent calculations and flexible language

That last category is where many small businesses find the sweet spot.

The bigger point

AI doesn’t have to be the thing doing everything. In fact, that’s often a poor way to use it.

A business owner might discover AI and start asking it to perform every task because it’s convenient. But convenience isn’t the same as good decision-making.

Sometimes the best solution is a spreadsheet formula. Sometimes it’s a Python script. Sometimes it’s an ordinary piece of software you’ve already got. Sometimes it’s AI. Sometimes it’s a mix.

The useful skill isn’t learning how to make AI do everything. It’s learning to recognise which part of a job actually needs a brain.

If the answer is “none of it”, don’t pay an AI to do it.

If the answer is “some of it”, let ordinary software handle the predictable parts and reserve AI for the parts where interpretation is genuinely useful.

Try this next

Pick one task you do on your computer at least once a week — something repetitive, like sorting a file or adding up numbers.

Open your AI tool and type:

“I do this task every week: [describe it in plain English]. Could a simple Python script do it instead? If so, write me one. Explain what each part does, and tell me what could go wrong.”

See what comes back.

You don’t have to use it. You don’t have to understand every line.

Just see if it looks like it could save you an hour a week.

That’s worth finding out.