Start with time, not technology.
Every conversation I have about AI automation starts in the wrong place. Someone has read about a tool — Make, Zapier, ChatGPT, n8n, whatever it is this month — and they want to know if they should be using it. My first question is always the same: what are you doing today that takes longer than it should?
That sounds obvious, but most people have never actually clocked how long their repetitive tasks take. A Deal-based plumber I worked with last year genuinely didn't know he was spending around four hours a week copy-pasting job details between his booking form, his invoicing software, and a WhatsApp group for his two part-time lads. He thought it was "just admin". It was half a day, every week, that he could have spent on the tools or quoting new jobs.
Before you look at any tool, spend one week writing down every task you do more than twice. Note roughly how long each one takes. That list is your automation backlog. Everything else is distraction.
The tasks where automation earns its keep.
There's a clear pattern in the places where automation genuinely helps a small business. The task needs to tick most of these boxes: it happens regularly (at least weekly), it follows a predictable sequence of steps, it involves moving information from one place to another, and getting it slightly wrong doesn't cause a disaster.
Booking confirmations are a classic. A Kent holiday cottage owner I helped last autumn was manually sending a confirmation email, then a check-in instructions email three days before arrival, then a review request two days after checkout. All from memory, occasionally forgetting the middle one. We set up a simple automation in Make that triggers off her Stripe payments — the confirmation goes instantly, the instructions email fires automatically at the right time, the review request follows. Took about a day to build. She reckons it saves her forty-five minutes a week and she no longer lies awake wondering if she sent the instructions.
Other tasks where I've seen it work well: generating first-draft responses to common customer enquiries, pulling data from an order form into a spreadsheet or a CRM, sending a weekly summary report to a client without touching it, flagging low stock levels in a Shopify store, routing incoming enquiries to the right person based on keywords.
The tasks where it tends to go wrong.
Automation breaks down when the task requires judgement, relationship, or context that changes every time. I'd be cautious about automating anything where a wrong answer genuinely damages the customer relationship — and even more cautious about anything customer-facing that the customer thinks is human.
I've seen a Canterbury retailer build a chatbot that was supposed to answer product questions. It answered them confidently and incorrectly about half the time. The customers who got bad information didn't complain — they just didn't come back. That's the worst outcome. A polite "I'll get back to you within a few hours" is almost always better than a fast wrong answer.
Complaints handling, bespoke quoting, anything involving a nuanced back-and-forth — these are not automation territory yet for most SMBs. The tools are getting better, but the failure modes are still too unpredictable for anything where trust is the product.
What it actually costs to set up.
This is where a lot of the hype falls apart. The tools themselves are often cheap — Make's free tier handles a reasonable number of automations, Zapier starts at around £16 a month, and many AI writing tools are under £20 a month. But the cost isn't the subscription. The cost is the time to build, test, and maintain it.
A simple two-step automation (form submission triggers an email) takes an hour or two if you know what you're doing. A more involved workflow — say, a new Shopify order creates a delivery note PDF, sends it to Royal Mail Click & Drop via their API, and updates a Google Sheet — can easily take a day to build properly, and then another half-day when one of the services changes its API or the spreadsheet columns shift.
The honest calculation is: how many hours will this save per month, and how many hours will it take to build and keep running? For the plumber I mentioned, four hours a week saved versus about six hours to build and one hour a month to maintain. Clear winner. For a task that saves twenty minutes a week and takes two days to build, you're looking at nearly a year before it breaks even — and that assumes nothing breaks in the meantime.
AI specifically — beyond just connecting apps together.
There's a distinction worth making between automation in the traditional sense (if this, then that, move data around) and AI-powered automation, where a language model is actually generating or interpreting something. Both have their place, but they have different risk profiles.
The AI part — using something like GPT-4o or Claude — adds real value when the output needs to sound like a human wrote it, or when the input is unstructured. A service business receiving enquiry emails in all shapes and sizes can use an AI step to classify the enquiry (new client, existing client, complaint, supplier) and draft a suggested response for a human to review. The human still sends it. But instead of starting from a blank page, they're editing a reasonable first draft. That's a genuine time saving without the risk of fully unsupervised output.
I built something similar for a Faversham-based events company last year — incoming enquiries were being missed because the inbox was shared between three people and nobody was sure who was responding. The automation reads the email, classifies it, drafts a reply suggestion, and posts it to a Slack channel where the right person picks it up. Enquiry response time went from occasionally never to usually within a couple of hours. That's a real business improvement.
The three questions to ask before you build anything.
Before I start any automation project, I ask three things. First: if this automation produces a bad output once a month, what's the worst that happens? If the answer is "a slightly weird email gets sent internally", proceed. If the answer is "a customer gets billed the wrong amount", add a human review step or don't automate it yet.
Second: who will maintain this when it breaks? Automations do break. APIs change, free tiers get paywalled, someone renames a column in a spreadsheet. If the honest answer is "nobody", either keep it very simple or factor in a maintenance budget.
Third: are you automating because the process is working and you want to do it faster, or are you automating because the process is a mess and you hope the tool will sort it out? Tools don't fix broken processes. They make broken processes break faster and at scale. Straighten the workflow out first, then automate it.
Where to start if you're new to this.
Pick the most boring, predictable task on your list. Something that happens every week without fail, follows exactly the same steps, and involves no customer-facing output. Build the simplest possible version — don't try to connect five systems at once. Use Make or Zapier with their built-in templates to get a feel for how it works before writing anything custom. Run it in parallel with the manual process for a fortnight before you trust it fully.
For most Kent small businesses, the first worthwhile automation is something in the enquiry-to-confirmation pipeline, or in the order-to-fulfilment pipeline for anyone running a small shop. Those areas have the most repetition and the most obvious time drain. They're also forgiving enough that a minor error doesn't end a customer relationship.
Honestly, the best ROI I've seen from AI automation for SMBs isn't some grand transformation. It's getting forty-five minutes back every day from tasks that didn't need a human doing them in the first place — and then using that time on the work that actually does.