AI on the busywork, not a bet on the wrong tool.
I build AI workflows and automations for businesses, online stores included: customer emails sorted and drafted for a person to approve, product data merged from several sources and cleaned before it is published, documents read and keyed in, orders moved between systems without anyone retyping them. Sometimes that means AI. Often it means a plain piece of code or an app you can buy this afternoon, and I’ll point you to it.
Three tools for the job
AI, a plain script, or an app you can buy
Plain code does most of the work in most automations, and AI handles the one step that involves reading or writing words. Before recommending a build I check whether an existing app already does the job well enough.
A plain script
Rules you could write on a card: if this, then that. It fails when the input doesn’t match the rules.
- Orders to a courier system
- Nightly stock from a supplier file
- Weekly figures from four exports
Start with the task
Real AI workflows, and what each one does
AI email triage
Customer emails sorted by what they want, with a reply drafted for a person to approve.
Returns, refunds, delivery questions and complaints arrive in one inbox. The workflow reads each email, pulls the order it refers to, labels it (return, refund, where is my order, complaint), drafts a reply using your policies, and flags anything unusual, angry or high-value for a person. Nothing is sent without someone approving it until the error rate has been measured on your own mail.
Product data from many sources
Supplier feeds, spreadsheets and PDFs merged into one clean record per product.
Three suppliers describe the same product three ways. The workflow collects every source, matches the records, fills gaps, standardises units and attributes, writes first-draft titles and descriptions, and scores each product for completeness. Clean records go to your PIM or straight to the store; anything it isn’t sure about waits for a person.
AI workflow automation
Multi-step workflows that connect your tools and hand only the judgement calls to a model.
Built in n8n, Make or plain code, depending on what you already use and how much it has to carry. The workflow does the routine steps itself, calls a model only where the input is messy language, logs every decision, and stops for a person at the points you choose. I’ll say when a job has outgrown a no-code tool.
Document processing
Invoices, purchase orders and delivery notes read and entered for you.
PDFs and scans arrive by email. The workflow reads the fields that matter, such as supplier, lines, quantities, prices and dates, checks them against the purchase order or the stock system, and enters them. Mismatches and low-confidence reads are queued for a person rather than guessed.
Orders between systems
Web and marketplace orders passed to accounts, couriers and suppliers without retyping.
Orders from the store, Amazon or a trade portal are sent on to the accounts package, the courier system or a drop-ship supplier in the format each one expects, with tracking and invoices brought back. Usually plain code, not AI, and the most reliable win on this list.
Supplier stock and price files
The spreadsheet that arrives every morning, applied to the store automatically.
The supplier’s file is picked up from email or a shared folder, checked for obvious errors, such as a price that dropped 90% overnight, and applied to stock and prices. A summary of what changed goes to a person, with anything suspicious held back.
Internal knowledge assistant
Staff questions answered from your own policies, manuals and past tickets.
An assistant that answers from your documents, not the open internet: returns policy, product specs, supplier terms, how-to guides. Every answer links to the page it came from, so staff can check it, and questions it can’t answer are logged so the gaps get written up.
Reporting
The weekly figures someone assembles from four exports, built once and sent every Monday.
Sales, ad spend, stock and customer service numbers pulled from each system, reconciled, and written up as a short summary with the figures that moved and a likely reason. The numbers come from code; AI only writes the words around them.
Stock planning
Sales history used to flag what will run out before it does.
A forecast built from your own sales, seasonality and lead times, flagging lines that will run short and suggesting order quantities. It supports the buyer’s decision rather than making it.
What happens to your customers’ data
Before anything is built, the task is written down: the steps as the person doing it would describe them, how often it happens, how long each one takes, what a mistake costs, and where the data lives. Some tasks turn out not to be worth automating, and the ones that are usually point clearly at the right tool.
Sending customer names, addresses or order histories to an AI provider counts as sharing personal data with a processor under UK GDPR. That is allowed, with the right agreement in place and a note in your privacy information, but it should be a decision rather than an accident. Where it makes sense the automation sends only what the task needs, strips personal details before anything leaves your systems, or uses a model on infrastructure you control.
Then it is built against real examples from your own data, tested on them before anything goes live, and run with a person checking until the results earn less checking.
- This isn’t legal advice. If your data is sensitive, your data protection adviser should see the design.
- Two costs are stated before you agree: building it, and running it, which depends on volume and on whether a paid model is involved.
AI work from before it was a product category
My first AI work predates ChatGPT, when a project meant collecting the data and training the model yourself.
That background is useful now mostly for what it rules out: it makes it easier to see when a task needs a model and when a few lines of ordinary code would be more reliable.
The rest of eighteen years has been in the systems these automations plug into: ecommerce platforms, ERPs, email tools and the servers under them.
What this is for
Automation on a named task, not an AI programme
What it covers
- A task your team does by hand every day, written down with the person who does it
- A recommendation: an app you can buy, a plain script, AI, or a mix
- Build and running costs stated separately before you agree
- Built and tested on real examples from your own data
- Run with a person checking until the results earn less checking
What it does not
- A chatbot on the home page for the sake of having one
- A company-wide AI transformation programme
- A monthly retainer for workflows nobody has defined
- Training your team to use AI tools, which sits under Learning
AI and automation questions
- How much does AI automation cost?
- Two costs: building it, priced on the task once it is written down, and running it, which depends on volume and on whether a paid model is involved. Both are stated before you agree.
- Do I need AI, or would a script do?
- If you could explain the rules to a new starter on one page, a script will usually do, and it will be cheaper and more predictable. AI is for the steps where the input is messy language.
- Will it replace someone on my team?
- In most small businesses it takes the dullest hours out of someone’s week rather than their job. What happens to those hours is your decision.
- Is it safe to send customer data to an AI model?
- It can be, with the right processor terms and only the data the task needs. Where it makes sense, personal details are stripped before anything leaves your systems.
- What happens when the AI gets it wrong?
- The design assumes it sometimes will: a person checks customer-facing output, uncertain cases go to a person, and every decision is logged.
- Can you work with n8n, Make or Zapier?
- Yes, where they fit the task. I’ll say if a job has outgrown them.