AI and automation
The most common AI automation mistakes are automating a messy process, letting AI talk to customers with no human check, and switching it on and forgetting it. Each one has a simple fix that costs little more than an hour of planning.
By Look North Digital · Published
Most problems aren't caused by the technology. They come from how it was set up and looked after. A tool that works well in a demo meets real customers, odd requests and busy weeks, and small gaps start to show.
The good news is that the same few mistakes come up again and again. If you know them in advance, you can avoid most of them before you spend any money.
| Mistake | What it looks like | The fix |
|---|---|---|
| Automating a messy process | The automation copies the muddle, only faster | Write the process down and tidy it first |
| No human check | A wrong price or promise reaches a customer | A person approves anything customer-facing at first |
| Wrong account for customer data | Customer emails pasted into a personal AI account | Use business plans and a simple data rule |
| Faking it | AI-written reviews or posts pretending to be customers | Use AI only for your own genuine content |
| Set and forget | An automation quietly fails for weeks | Check the run history every week |
| No owner | Nobody knows how it works when it breaks | Name an owner and write a one-page note |
| Too big, too soon | Months of set-up before anything useful happens | Start with one small task and build up |
If your enquiry handling or booking process is muddled, automation makes the muddle happen faster. A common example is an automatic reply that promises a call back 'within the hour' when nobody is free to make it.
Before you automate anything, write the process down as a list of steps on paper. Note who does each step, what information they need and what happens when something unusual comes in. You'll often spot a step you can simply remove.
This is the most damaging mistake, because the errors are public. The NCSC says AI tools can 'hallucinate' incorrect facts, can be biased and are often gullible when given leading questions. Google's own help pages warn that Gemini features may suggest inaccurate or inappropriate information.
Your business is still responsible for what goes out in its name. So for anything customer-facing, such as review replies, emails, quotes or social posts, start with a person reading and approving every message. Once you've seen weeks of good results, you can relax the checks on the simplest, lowest-risk messages.
A useful test: if you'd be embarrassed to see a message screenshotted and shared, a person should approve it before it's sent.
It's easy for staff to paste a customer's email into a free AI tool to get a quick reply. The NCSC advises not to include sensitive information in queries to public AI tools, and points out that the company running the tool can see what you type.
Plan type matters. ChatGPT's pricing page says content on its Free, Go, Plus and Pro plans is used to train its models, with an opt-out available. Google says Gemini in Google Workspace doesn't use customer data to train models without permission. Choose business accounts and tell your team which tools they may use.
UK data protection law still applies when you use AI. The ICO (Information Commissioner's Office, the UK's data protection regulator) says your organisation stays accountable and that responsibility can't simply be handed to technical teams. Its guidance on AI and data protection is the place to go for detail.
It can be tempting to have AI write a few glowing reviews, or to post comments that look like they came from customers. Don't. Google's policy says fake engagement is not allowed and will be removed. Reviews must reflect a genuine experience, and reviews from staff or anyone with a conflict of interest are not allowed.
Be open about AI in your content too. Google's guidance on helpful content asks whether the use of automation, including AI, is clear to visitors. Using AI to help draft your own blog posts or social posts is fine. Pretending AI content came from real customers is not.
Automations break quietly. A password changes, a form gets a new field or a connected app updates, and runs start failing without anyone noticing. Meanwhile, enquiries sit unanswered.
Most automation tools keep a log. Zapier, for example, keeps a run history that shows whether each run succeeded, and lets you filter by status and replay failed runs. Put a short weekly check in the diary, and look at a sample of what the AI actually sent.
In a small business, automations are often set up by one person on a busy afternoon. When that person is on holiday or leaves, nobody knows how it works, what it's connected to or how to stop it.
Give every automation a named owner. Then write a one-page note covering what it does, which tools and accounts it uses, who it contacts and how to switch it off. Keep the prompts you use in that note too, so they don't live only inside one person's account.
Trying to automate enquiries, bookings, invoicing and marketing at once usually means months of set-up and nothing finished. It also makes it hard to tell which part is causing a problem.
Start with one small, frequent task that has a clear right answer, such as acknowledging new enquiries or drafting review replies for you to approve. Run it for a month with checks in place. Add the next task only when the first is working well.
A short check before launch catches most of the mistakes above. Run through it with whoever will own the automation.
| Question | Good answer |
|---|---|
| Is the process written down? | Yes, step by step, including unusual cases. |
| Who approves customer-facing messages? | A named person, at least for the first few weeks. |
| What data does it use? | Only what it needs, through business accounts. |
| What happens when it can't handle something? | It passes the job to a person, who is told straight away. |
| Who owns it? | A named person, with a one-page note on how it works. |
| How will you know if it fails? | A weekly check of the run history, in the diary. |
| How do you switch it off? | Written down, and more than one person knows how. |
Our AI training is a set of live remote sessions built on your own work, so your team learns to use AI safely on real tasks. You keep the prompt library afterwards. Our custom software and integrations work connects your website, CRM, booking system and supplier APIs (the links that let systems share data) with checks and hand-offs built in.
Our Automated Social Media service plans and writes your Facebook and Instagram posts, but nothing is published without your approval. To talk it through, use the 'Tell us about your project' form.
Letting AI send messages to customers without a person checking them is the most damaging. AI can give convincing answers that are wrong, and the mistakes are public. Start with a person approving every customer-facing message.
For simple, low-risk messages, such as an acknowledgement that an enquiry has arrived, it can be fine once you've tested it well. Anything involving prices, promises, complaints or personal matters should have a person in the loop. Review samples regularly even when it's running on its own.
For general tasks with no customer data, often yes. But on ChatGPT's personal plans content may be used to train OpenAI's models unless you turn this off in settings, and the NCSC advises against putting sensitive information into public AI tools. Give staff clear rules and business accounts for customer work.
Look at the run history or log once a week, and read a sample of what the AI sent. Review whether each automation is still needed every few months. Automations can fail quietly when connected apps change.
No. Google's policy says fake engagement is not allowed and will be removed, and reviews must reflect a genuine experience. Use AI to draft a polite request asking real customers for an honest review instead.
Your business is. The ICO says organisations stay accountable for how they use personal data in AI, and customers will hold you responsible for what's sent in your name. That's why human checks and a named owner matter.
Pick one small task, write the process down, and run it for a month with a person checking the results. Name an owner and keep a short note on how it works. Add more only once the first task is running well.
For content, Google's helpful content guidance asks whether the use of AI is clear to visitors. Being open is a good habit generally. If AI handles personal data, check the ICO's guidance on being transparent with people.
Keep reading
More on the same subject, and where we can help with it.
Next step
Tell us what eats your time or where enquiries slip away. We will suggest the right mix of website chatbot and automation, with a fixed price.
We use cookies that are essential for this site to work. With your consent, we may also use analytics cookies to understand how the site is used. See our cookie policy for details.