Building an Agentic AI Assistant for Small Business

I didn’t start this with a grand AI strategy. I watched a YouTube video, bought a used PC on eBay, and started wondering whether this could help a small business owner stop dropping so many small but important things.

My first proper introduction to AI agents was not through a business seminar or some carefully planned technology project. It was a YouTube video.

I watched one video about OpenClaw and AI agents, then did what most of us do when something catches our attention: I watched a few more. A lot of the examples were clever, but many of them were also very coding-heavy. Developer workflows, software projects, technical experiments, all impressive in their own way.

But I kept looking at it from a different angle.

I was not really asking, “Can this write code for me?” I was asking, “Could this help me run my business better?”

Could it save time? Could it help me stay organised? Could it give more structure to the bits of the business that normally depend on me remembering, noticing, following up, and joining the dots at the right moment?

So, after watching a few videos and convincing myself this was definitely a sensible business decision and not just another curiosity rabbit hole, I bought a used PC on eBay and started experimenting.

That became the machine I would run OpenClaw on, and from there I began building, breaking things, fixing things, changing my mind, and trying to work out what agentic AI might actually mean for a solopreneur.

I should say this clearly: I am a long way from having it perfectly sorted. This is not one of those stories where I installed an AI assistant and now the business runs itself while I drink coffee and look thoughtfully out of a window.

It does not.

But I can see some real advantages, especially if you treat it less like a magic robot employee and more like a set of useful systems organised around the way a small business already works.

AI agents are a relatively new phenomenen. They started to gain attention around February / March of 2026. The agent that sparked that early interest was named OpenClaw: a project built by one man, Peter Steinberger, an Austrian programmer and entrepreneur who created the viral, personal AI agent in late 2025. Since then, many of the large tech players have announced their own versions of agents that are either released (Claude Cowork) or are in the works.

The biggest difference between an AI agent and a more familiar AI chat is that you typically don’t ask an agent a question; you give it a task to complete. It then uses tools like ChatGPT, Claude or Gemini to complete the task.

For a small business owner, that task might be: go through the emails I’ve received today, find any that are urgent or require a response and draft a reply for me. Notify me when you’re done.

Or, every morning at 7am, send me a reminder of my upcoming appointments for the day.

The other big difference is memory. When you’re interacting with ChatGPT, it has a limited memory, so it will start to discard old conversations to make space for new. With an AI agent, when set up correctly, it will retain most of what you’ve said previously. The result over time is that your agent gets smarter, knows more about you, remembers more about you. You spend less time going over old ground.

A lot of the language around AI agents makes it sound as if the aim is to create an autonomous employee that goes away and handles everything for you. I understand why that sounds attractive. If you run a small business, the idea of something quietly dealing with the admin, follow-ups, reminders, notes, research and awkward half-finished jobs is very appealing.

But I do not think that is the right starting point. At least, it has not been for me.

The more useful question has been much plainer: where could this take some pressure off?

Not replace judgement. Not make decisions I should be making. Not send things to clients or prospects without me checking them. Just take some pressure off the parts of the business where there is too much context, too many small moving pieces, or too much relying on my memory.

That is where it started to become interesting, because when you run a small business you are not usually short of things to do. You are short of clean handovers between all the different versions of yourself.

There is the version of you who has a good idea while walking between jobs. The version who reads an email and thinks, “I need to deal with that properly.” The version who has a useful conversation with a prospect, spots a content angle, makes a decision, or promises to follow something up.

Then there is the later version of you who has to remember all of that while also doing the actual work.

That is where things get messy.

One of the early lessons was that an assistant is only useful if it can keep hold of context in a way you can trust.

A normal chat window is useful, but it is also limited. You ask a question, get an answer, paste something in, maybe get a decent draft back. That can be helpful, but it still feels like a conversation floating in the air.

If something important gets decided, where does it live? If a preference changes, how does the assistant know next time? If you work something out today, how do you stop yourself having to explain it again next week?

That was one of the first places where OpenClaw started to feel different from simply chatting with AI. The useful bit was not “the AI remembers everything”, because I would not trust that as a business system. The useful bit was being able to create written memory, documents, notes, routines, and operating instructions that the assistant can refer back to.

That sounds boring compared with the usual AI hype, but boring is often where the value is.

A small business needs continuity. It needs decisions written down, notes that do not disappear, and a way of keeping preferences, context and repeatable processes somewhere more reliable than the owner’s head. And yes, I am very aware that I am the owner in that sentence.

One of the most practical parts of the setup has been speech.

I use Wispr Flow, which lets me speak a thought or instruction and turn it into usable text. That might not sound as exciting as an “autonomous agent workflow”, but for me it has been one of the most useful pieces of the puzzle.

Typing creates friction. Not always, but often enough.

You think of something while you are away from the desk, between jobs, just after a call, or doing something else entirely. You know it is worth capturing, but the thought of stopping, opening the right app, typing it neatly, and putting it in the correct place is enough to make you leave it until later.

And later is where good ideas often go to die.

Being able to speak a rough thought into a familiar channel changes that. At the moment, Discord is one of the channels I use, but the bigger point is not Discord itself. It could be WhatsApp, Telegram, Teams, or whatever makes sense for the way someone already works.

The important thing is that the assistant does not have to live behind one perfect interface that I only use when I am sitting at the desk in a calm, organised mood. That is not how most small businesses operate.

If I can speak an instruction, a reminder, a rough content idea, or a note about something I need to come back to, it lowers the distance between having the thought and getting it into the system. That distance matters more than people realise.

A spoken note can become a task, a reminder, the start of a draft, a research brief, an instruction to check an email thread, or a follow-up action after a meeting. It is not glamorous, but it is useful, and useful beats glamorous most days of the week.

I have also experimented with specialist agents, and this is one of those areas where it is easy to get carried away.

You can invent a whole pretend company full of AI job titles and feel very clever for about half an hour, while not actually making the business any better. The useful version is much more restrained.

For me, the coordinator assistant is the front door. I gave my assistant a name and a title: Julie is my Executive Assistant. The role is about continuity, organisation, memory, reminders, documentation, and helping decide what needs attention.

Then there are specialist roles where they genuinely make sense: a content agent (Candy) for drafting and shaping posts, a research agent (Scout – I’d gone away from names at this point) for finding useful ideas and signals, an SEO/search-support agent (Charlie) for search intent and content opportunities, and a marketing-direction agent (Kenny) for positioning, offers and campaign thinking.

The point is not to have lots of agents for the sake of it. The point is to stop pretending one assistant should be equally good at everything.

A content task needs a different brief from an SEO task. Research needs different judgement from calendar awareness. Marketing direction is not the same as tidying up a document. Clear roles help, but only if the work comes back into the same system.

If those specialist agents all disappear into separate conversations, you have not built a business assistant. You have built more places to lose things. The output needs to be saved, the decision needs to be visible, and the next step needs to be clear.

Otherwise it is just more tabs open, but with better branding.

Email is one of the most obvious places an assistant can help, although it is also one of the places where I would be careful.

The inbox is not just an inbox. It is leads, clients, suppliers, admin, invoices, appointments, support requests, meeting details, and those messages that need a thoughtful reply so they sit there longer than they should.

An assistant can help by summarising a thread, pulling out the action, drafting a possible reply, or reminding me what the context is. That is useful, especially when the alternative is staring at an email thinking, “I need to deal with this properly,” and then moving on because the next thing is shouting louder.

But I do not want an AI freely sending emails on my behalf just because it sounds confident. A draft is one thing. A message going to a client, prospect, or supplier is another. I took baby steps with this particular implementation; I gave the AI agent its own email address and put a forward on my business email account so the AI agent gets a copy of every email sent to me. When the agent drafts a reply it’s saved in a text document where I can simply copy and paste into my reply. 

The same applies to calendar awareness. I do not need a fantasy diary manager that rearranges my life while I make a cup of tea, although I admit that would be convenient, and slightly worrying. I need something more ordinary: what is coming up, what do I need to prepare, what did I promise before that meeting, and is there an email or task connected to it?

That sort of awareness can prevent small mistakes, not because the assistant is magic, but because a lot of business mistakes are caused by useful information living in too many different places.

Content seemed like an obvious use case at first, and it is. But it has also shown me one of the biggest limitations of AI: it can produce something that is technically fine and still not feel like you.

That matters.

A blog post can have the right points, the right structure, and the right keywords, but if it sounds like generic AI content, it will not build much trust. It might be tidy, but tidy is not the same as useful or human.

This article is a good example. Earlier versions had plenty of correct information in them, but they did not quite feel like my voice. They were too neat, too explanatory, and too much like the AI was standing at the front of the room with a laser pointer saying, “Here are the lessons learned.”

That is not really how I write.

I tend to come at things through a story, a comparison, a problem, or something that made me stop and think. I would rather admit I am still figuring something out than pretend I have a polished framework for everything. To that point I am currently testing a process of having the content agent run it’s first draft against the ‘tells’ listed in the Wikipedia article – Signs of AI writing, and rewriting to strengthen the final draft.

That is probably one of the more important lessons for business owners using AI for content. Do not mistake a complete draft for a finished draft.

The first version might be a useful starting point, but the voice still needs judgement. It needs your examples, your doubts, your phrasing, and sometimes your refusal to let it sound like every other article on the internet.

Research has been another useful lesson, because agents do not automatically know what “useful” means.

At first, it is tempting to ask a research agent to find content opportunities and expect gold to come back. Sometimes you get something useful. Sometimes you get a neat pile of things that are technically relevant but not actually helpful: vendor pages, bland summaries, and ideas that sound fine until you imagine a real small business owner reading them and thinking, “So what?”

The research became more useful when the brief became clearer. Instead of just looking for information, it needed to look for real business-owner problems: missed enquiries, follow-up gaps, reviews, booking systems, CRM, websites that are not pulling their weight, and practical AI use cases that normal people can understand without needing a glossary.

That was a useful lesson in itself. Agentic AI is not like buying a finished employee in a box. It is more like training a system around how your business thinks.

You review the output, notice what is weak, improve the brief, keep what works, and change what does not. Very glamorous, I know, but that is how useful systems are built.

This might be one of the most important distinctions I have learned: automation does not always mean autonomy.

Some of the most useful parts of this setup are boring. Scheduled checks, backups, moving useful files where they need to be, producing regular research notes, reminders, calendar-aware prompts, and little routines that reduce the need to remember everything manually.

Not every job needs the cleverest AI model in the room. If the task is basically, “Run this routine and only complain if it fails,” then a smaller local model (qwen in my case) or a simple scheduled process may be enough.

For more important work — strategy, writing, judgement, complex problem-solving — it makes sense to use a stronger model such as ChatGPT or Claude. The point is not to use the biggest tool for every tiny job. The point is to match the tool to the risk and importance of the task.

For a small business, that balance matters. AI should not become another expensive toy, but it also should not become a race to use the cheapest option for everything, especially when the output represents the business.

Low-risk routine? Keep it local if possible. Important client-facing writing or strategic thinking? Use the stronger tool, then review it like a grown-up.

The more an assistant can do, the more important the boundaries become.

It is tempting to focus on capability: can it read this, write that, check this, summarise that, connect here, trigger there? But the better question is often simpler and more important: should it be allowed to do that without asking?

For me, the answer is usually no when it comes to anything external. I do not want an agent publishing posts, sending emails, contacting people, or making business decisions without review.

That is not because I do not trust the technology at all. It is because trust is built in layers.

Start with notes, drafts, reminders and internal routines. Then maybe add more automation where the risk is low and the rules are clear. Passwords, private keys, payment details, private client information and sensitive personal material need proper care. That is not really an AI issue; it is just basic business hygiene.

The assistant should make the business more reliable, not more reckless.

There is also the unglamorous side of building this properly: backups, documentation, naming conventions, recovery notes, security checks, clear handoff folders, knowing where drafts live, and knowing what happens if a process fails.

None of that makes a dramatic YouTube thumbnail.

But once an AI assistant becomes part of how the business works, those boring bits matter. If it has memory, files, tools, routines and responsibilities, then it needs to be treated like part of the operating system of the business, not a toy you poke at when you are curious.

That is one of the things I have had to learn while building this. The clever demo is the easy part. Making it reliable enough to be useful is where the real work starts.

If you run a small business, I would not start by asking, “How do I build an AI agent?” I would start with something more practical.

Where do you lose time? Where are you relying on memory? Where do good ideas disappear? Where do emails sit because they need thought before you reply? Where do you repeat the same explanations, checks, notes, or admin jobs?

I would also ask where friction stops you capturing something useful. Where would a spoken instruction be easier than sitting down to type properly? Where would a draft, summary, reminder, research note, or documented decision make the next step easier?

That is where I think the useful version of agentic AI starts.

Not with a robot employee. Not with pretending the business can run itself. Not with handing over judgement to a machine. It starts with building small systems around the work, so fewer things depend entirely on the owner being in exactly the right place, at exactly the right time, with exactly the right thing in their head.

For a solopreneur, that is not a small benefit.

I started this because a YouTube video made me curious enough to buy a used PC on eBay and see what would happen. That is not a very grand origin story, but it is an honest one.

Since then, I have found that the useful version of an AI assistant is less about replacing people and more about supporting the way a small business actually runs. Capturing thoughts, keeping continuity, helping with drafts, surfacing context, supporting research, creating routines, keeping decisions visible, and making sure the same useful idea does not have to be rediscovered three times.

I am still experimenting, and I would be suspicious of anyone who claimed to have all of this perfectly solved. But I can see the shape of something valuable.

For me, agentic AI is not about building a fantasy autonomous company. It is about adding structure around the work, so the business can respond faster, follow up better, and stop relying quite so heavily on everything being carried around in one person’s head.

And if you run a small business, especially on your own, that is probably the part worth paying attention to.

After reading this – if you’re the hands-on, I like to build and experiment type – and you’d like to know more about where I started, which AI models I used, the hardware I built on, then get in touch via the contact page. I’m happy to explain more.

If you’d like to take smaller steps and just introduce a couple of AI helpers into your business, to maybe deal with enquiries through your website, or manage inbound calls after hours then that’s exactly what the Set & Scale Marketing Engine is built for and you can learn more on the link below.

👉 Explore the Set & Scale Marketing Engine