AI · Automation
I set up AI automation for small businesses, for the tasks that someone in your company repeats twenty times a week, the same way each time. That might mean copying enquiries from emails into a spreadsheet, preparing quotes from your price list, describing the next thirty products or putting together a report from the same four sources. The model takes over work that needs no judgement but still takes hours, and the decisions stay with you.
Scope of work
These are the cases that come up most often in small businesses. Not every company can automate all of them. It works when the data sits in one place and the rule fits into one sentence.
Many messages ask the same questions again and again, about availability, start dates or the price of the standard scope. The model drafts a reply from your price list and terms, and you approve or correct it with one click. Nothing goes out without your approval, and that is how I always set up the first version of an automation.
Some shops have several hundred products, half of them with a description copied from the manufacturer or none at all. The model writes versions from the product attributes and in the way you talk about your products, and you review them in batches. In a shop, this is often the first choice, because the answer is already in the shop’s data and there is a lot of work, all of it repetitive.
A report can cover last month’s sales, requests broken down by type or website traffic set against orders. I pull the data from your systems, and the model turns it into an overview with a short summary in English. The figures come from the database rather than from the model, which is the most important distinction in any report.
Messages from the website and the inbox land in one place, and then someone sorts them by hand. The model assigns a category and an urgency level, adds a two-sentence summary and sends the matter where it belongs. When a request is unclear, it leaves it without a label instead of guessing.
This covers category descriptions, blog posts and versions of a text for other locations or services. The source material comes from you, the model helps to arrange and shorten it, and every text is checked before publication anyway. Deciding what is worth writing about stays with a person.
In all five cases, the data sits in one place, the rule fits into one sentence and someone approves the result. If one of these three conditions is missing, I say so at the start. It can usually be put in place, but that has to happen before the rollout. Texts written for visibility in search engines are separate work, billed in my SEO plans.
Process
The pilot shows whether the process leaves you with less work after the rollout than its maintenance adds. It has four steps, usually spread over a few weeks, and you can stop after the third.
I help you choose the task that repeats most often and needs the fewest decisions, which is not always the one that sounds most interesting in a presentation. This choice affects the result more than anything that comes later. It is usually the thing you describe as “the same every day”.
I count how many cases come in each week, how long each one takes and how many need corrections along the way. Without these three numbers, all that is left after a month is the impression that things are “probably better”, which is not enough reason to spend money on expanding. The measurement takes less time than it seems.
For an agreed period, the automation handles the same cases as you do today, and you compare the results with me. If the model needs corrections, I refine the rules at this stage. That is the job of the pilot, before the solution covers the whole company.
I put the numbers from the second step next to the results of the pilot. Expanding to the next process makes sense when the first one saves more than its maintenance and the model’s usage fees cost. If it does not, I tell you so directly instead of looking for arguments for a bigger job.
Technology
Most of my clients work with WordPress, often with a WooCommerce shop. The automation goes in as an extra component that reads and writes data where it already sits, and everything else stays as it is.
Products, orders, form enquiries and posts are already in the WordPress database. The automation reads them there and stores its results there, so there is no second copy of the data for someone to watch and keep in sync.
I put the logic in a plugin rather than in the theme files. That way it keeps working after a theme change, and if there is a problem, you switch it off with one click in the admin panel without waiting for my reply.
Access to the model runs on your own account and API key, with a monthly spending limit. You see the cost of usage and can cut it off at any time without talking to me. The same goes for ending the collaboration.
Tasks go through a queue with retries. When the model provider does not respond or the monthly limit runs out, the request waits and leaves a trace in the log instead of quietly disappearing. With someone else’s API, this is the most important part of the code.
The mechanics are the same as in the work I describe under WordPress plugins and API integrations, with a model on the other side of the connection. For a shop that has yet to be built or needs a rebuild, see E-commerce website development. If your website is not under regular maintenance today, start with WordPress maintenance, because an automation attached to a neglected website stops working at the first larger update.
Limits
Automation should take routine work off your team and leave the decisions that build customer relationships to people. In the four areas below, the model only prepares material, and you keep full control over what reaches the customer.
How big a discount is, when to depart from the price list and what terms a regular customer gets are commercial decisions, and they stay with you. The model can prepare an option for your approval, but it does not send it on its own or negotiate on your behalf.
A complaint is by definition an unusual situation that needs judgement and sometimes a concession. The automation can describe it, link it to the right case and gather the material, but a person replies, even when the answer is uncomfortable.
This covers your terms and conditions, the privacy policy, warranty information and descriptions that make claims about a product’s properties. I do not publish such content without a check by someone who takes responsibility for it.
If a model is answering, the customer should know. I do not build automations that pose as an employee, and I do not sign mass-produced messages with someone’s name. Sooner or later this comes out, and it costs more than it brought in.
Starting the pilot
In automation, most of the time goes on working out how the process really runs, and far less on the code. That is why the step-by-step description and the starting figures come first on this list.
If none of this is written down, that is normal and does not hold up the start. Organising the process is half of the project, and I can lead it. You do need someone who confirms that the written version matches how the company really works.
Other services
An automation and an assistant solve different problems with the same technology. Both take data from your systems, pass it through a model and work within limits set before the start.
An assistant on the website and on WhatsApp that answers only from your company’s knowledge. For unusual requests it hands the conversation, with its history, to a person.
Custom plugins and connections between WordPress and other systems such as CRM, stock management or applicant tracking, with data import and synchronisation. Quote after a short brief.
The full picture of my AI services, with the limits of a rollout and how I work with models myself, is on the AI for businesses page. If your problem has more to do with questions that customers keep asking than with recurring work inside the company, the right place is AI chatbots and assistants.
FAQ
Questions that come up in the first conversation about automation: the cost, the model’s mistakes and whether the website will have to change.
I prepare the quote after an analysis, because the cost depends mainly on how much has to be organised before the start and how many systems the solution has to connect to. A simple automation that reads data from one place takes a few days of work, and a process connected to several systems with a task queue takes a few weeks. On top of that, there is a fee for using the model, billed by usage. It depends directly on the number of requests, so I estimate it only once it is clear how many there will be each month. The spending limit is set in advance, on your account.
I do not sell automation with that argument, and I do not count savings in jobs. What I can promise is that a recurring task stops taking hours and the person who did it moves on to matters that need decisions. In the pilot, I measure the time and the number of corrections, so after a month you have numbers instead of impressions. If you are looking to cut staff costs, I will tell you directly that this offer is not designed for that.
That is why the first version of an automation never publishes anything on its own. Content goes into a queue for approval. You review it in batches and reject what does not fit, and the rejected cases come back to me as material for improving the rules. Publishing without a click is only an option for processes where a mistake costs little, such as internal reports. For product descriptions that customers see, approval stays in place permanently.
In most cases, yes, because the data the automation works with is already in the WordPress database. I check two things beforehand, the PHP version and whether the hosting allows background tasks, because the task queue depends on both. On cheap shared hosting, the website sometimes has to move to a more powerful server. I tell you that before the quote rather than halfway through the work.
The tasks wait in the queue and are processed once the service is back. Nothing disappears quietly, and nothing is sent twice after a retry. I watch for this in the same way as with any integration with someone else’s API. For processes where a delay matters, I set up a notification, so you learn about the outage without looking at the log. Every provider has outages, and I plan every project on the assumption that one will happen.
It finds the tasks in a business that repeat in the same form, connects a language model to the systems where the data sits and sets limits, so that a person approves the results. OneNet is a one-person business, and I do this work myself. I choose the process with you, run the pilot, build the connection to your systems and review the results, so you always deal with the person who does the work.
Yes. I work remotely with companies in the UK and the EU, among them a chain of tattoo and piercing studios in the UK and education organisations in Belgium. You deal with me directly, in English. I am available by email, WhatsApp, phone and video call, Monday to Friday, 9:00–17:00 Central European Time (8:00–16:00 UK time).
