Most small and medium businesses do not need a new AI product. They need the work that already arrives every day, such as enquiry forms, supplier invoices, customer emails and CVs, handled faster and more consistently. AI automation means placing a large language model at one or more steps of an automated workflow, where it does the reading and writing that used to need a person, then passes the result back to the systems you already use.
What AI automation fixes, and what changes for you
Ordinary automation is good at fixed rules: when a form is submitted, copy these fields into the CRM. It struggles the moment the input is free text. A customer writes three paragraphs about a problem, an invoice arrives as a scanned PDF in a different layout from last month’s, or an enquiry mentions a postcode, a budget and a deadline in no particular order. Until recently, someone had to read each one.
A language model can read that text and turn it into something a workflow can act on: a category, a set of fields, a short summary or a first draft of a reply. The practical change is that whoever used to triage the inbox spends their time on the cases that need judgement, and routine items move on as soon as they arrive instead of waiting until someone gets to them.
This is a different job from being found in AI search. If your question is how to get your business mentioned in ChatGPT answers or Google’s AI Overviews, that is AI search optimisation, which is marketing work rather than a workflow build.
Who this suits, and who it does not
It tends to pay off for UK small and medium businesses where the same kind of text arrives in volume:
- service businesses receiving dozens of enquiries a week through forms, email and messaging, where the speed of the first reply affects whether the job is won;
- firms that still key in documents by hand, such as purchase invoices, delivery notes, application forms or signed paperwork;
- teams answering similar customer emails all day, where most replies follow a pattern but still need a human tone;
- businesses whose CRM is half empty because nobody has time to copy details out of emails.
It is the wrong fit if one person can handle the volume comfortably in an hour a day, if the underlying process still changes every month, or if every case genuinely needs expert judgement. In those situations I would usually suggest tidying the process first, or a simple rule-based automation that costs less to run and maintain.
What I automate with AI
Most AI steps in a business workflow do one of four jobs, and a single workflow often combines two or three of them.
Classifying
The model reads an incoming message and assigns it a category: new enquiry, existing customer, complaint, supplier or spam. The category then decides what happens next, whether that is lead routing to the right person, a priority flag or a holding reply. Classification is the safest place to start, because the output is a single label that is easy to check.
Extracting
The model pulls named fields out of unstructured text or documents: a customer’s name, postcode, service and budget from an enquiry, or the supplier, invoice number, date, net amount and VAT from an invoice. Each field is checked against simple rules, for example that the VAT is a plausible share of the net amount, before it is written into your CRM, accounts software or spreadsheet.
Drafting
The model writes a first version of a reply, a quote cover note, a call summary or an internal handover, using your wording guidelines and the facts already held in the record. Drafts wait in a queue or in the email client for someone to edit and send. I do not set AI up to send customer-facing messages unreviewed, unless the content is tightly limited and the risk is low, such as a booking acknowledgement.
Summarising
The model condenses a long email thread, a call transcript or a form with many fields into a few lines a manager can act on, and the workflow posts that summary where the team already works: a shared inbox, Slack, Microsoft Teams or a note on the customer record.
What an engagement includes
A project usually covers one workflow from start to finish, rather than a scattering of experiments. That means:
- a short map of the current process, showing where the text comes from and where the result needs to land;
- the choice of model and automation platform, with the reasons written down;
- the prompt and instructions the model works from, kept in a versioned document so every change can be traced;
- a test set of real, anonymised examples from your business, used to measure accuracy before launch;
- error handling, logging and an alert when something fails;
- written documentation and a handover session for whoever will own the workflow.
How I build an AI workflow
- Collect examples. I ask for a few weeks’ worth of real items, with personal details removed where possible. They show the variety the model will face, including the awkward cases.
- Define the output. We agree exactly what a correct answer looks like: the list of categories, the fields and their formats, the tone of a draft. Vague outputs are the most common reason AI steps disappoint.
- Build and test. I run the model over the examples, compare its answers with what your team would have done, and adjust the instructions until the error rate is acceptable for that particular task.
- Run in shadow mode. For an agreed trial period the workflow runs alongside your existing process, producing results that are checked but not acted on.
- Go live with a person in charge. Items the model is unsure about, and anything above an agreed level of risk, go to a person for approval. That human-in-the-loop step stays in place permanently for decisions that affect customers.
- Review after launch. Once enough live items have passed through, I go through the logs, look at what was corrected by hand and tighten the instructions.
What to decide before automating with AI
A few decisions shape the whole build, and they are far easier to settle before any work starts.
Personal data and UK GDPR
If enquiries, CVs or customer emails pass through an AI model, you are processing personal data and UK GDPR applies. You need a lawful basis, a privacy notice that mentions the processing, a contract with the AI provider covering how it handles the data, and a clear answer on where the data is processed and whether it is used to train the provider’s models. Those terms differ between providers and between their consumer and business plans, so I check them for each one rather than assume.
Where a workflow makes decisions about people that have significant effects, such as screening job applicants, the law expects extra safeguards, including a way for a person to review the outcome. I flag these points and design around them, but I am not a lawyer, and anything sensitive should be signed off by whoever advises you on data protection.
What happens when the model is wrong
Language models sometimes give confident answers that are simply wrong, which is usually called hallucination. The useful question is what a mistake costs. A misfiled enquiry that someone spots the next morning is cheap; a wrong price in a sent email is not. I set the level of human checking according to that cost.
Running costs
AI providers charge by use, usually according to the amount of text processed, on top of any automation platform subscription. The cost rises with your volumes and with the model chosen, so I estimate it from your real numbers before building, and that estimate is part of deciding whether to go ahead. The build itself is quoted in GBP after a first call about the workflow, and the quote is agreed in writing before work starts.
What breaks, and how I prevent it
- Inputs change. A supplier redesigns its invoice or a new field appears on a form, and extraction quietly gets worse. Validation rules and a regular sample check catch this early.
- Silent failures. An access key expires or a provider has an outage, and items stop flowing without anyone noticing. Every workflow I build alerts someone on failure and holds unprocessed items so they can be run again.
- Model updates. Providers retire and replace models, and the same instructions can behave differently on a newer version. I fix the model version where the provider allows it and re-run the test set before any switch.
- Nobody owns it. The person who asked for the workflow leaves, and no one knows how it works. The documentation and handover exist for exactly that reason.
What you receive
At the end of a project you have a working workflow running in your own accounts, so you are not tied to me. You also get the instructions, the test set and the accuracy results; a short operating guide covering what to check, how to correct an item and what to do when an alert arrives; and notes on the data protection points for your records. Whether monitoring and adjustments continue after handover, and on what terms, is agreed before work starts.
Next step
The quickest way to find out whether AI is worth adding is to look at one process together. Book a free 30-minute call and bring a description of the task and a couple of anonymised examples. I will tell you whether it suits AI, plain automation or neither, and what a build would involve. You can see the platforms I work with on my tools I use page, and if you are weighing up AI for content rather than operations, my article on using AI for marketing without hurting SEO covers that side.