A few weeks ago, the owner of an Osaka building-materials wholesaler forwarded me a sales email with one line on top: "Is this real, or is it another ChatGPT?"
The email promised an "autonomous AI agent" that would "run your back office while you sleep." Twenty-two employees, a sales team that lives in Outlook, quotes built by hand in Excel. He was tempted. He was also suspicious, and he was right to be both.
"AI agent" is the buzzword of 2026. Every software vendor has one, every conference talks about them, and most small business owners I meet in Kansai have no clear idea what the term actually means. So let's fix that. This is my honest take on AI agents for small business: what they are, where they genuinely help a 5–50 person company, what they cost, where they go wrong, and how to start without betting the company on it.
What is an AI agent, really? (And how is it different from ChatGPT?)
Here's the simplest way I can put it.
- A chatbot answers questions. You ask, it replies, the conversation ends. It follows a script or searches a knowledge base. (We covered these in our article on AI chatbots for customer service.)
- ChatGPT-style assistants are much smarter conversation partners. They draft, summarise, translate and brainstorm. But *you* do all the copying, pasting, clicking and sending. The AI thinks; you act. (See 5 practical ways to use ChatGPT at work.)
- An AI agent is given a goal, a set of tools, and permission to use them. It decides the steps, takes actions in your systems (read an email, look up a customer, fill in a spreadsheet, create a draft), checks the result, and keeps going until the task is done or it gets stuck.
The key word is *action*. A chatbot talks. An agent does things.
That's both the promise and the danger. An assistant that writes a bad paragraph wastes thirty seconds. An agent that sends the wrong quote to the wrong client costs you a customer.
A useful mental model: an AI agent is like a very fast, very literal new hire on their first week. Capable, tireless, occasionally confidently wrong, and absolutely not someone you'd give the company bank card to on day one.
5 realistic AI agent use cases for a 5–50 person company
Forget the "fully autonomous company" demos. These are the use cases where I've seen agents earn their keep in small businesses, roughly in order of how safe they are to start with.
1. Inbox triage and reply drafting
The agent reads incoming email to a shared address (info@, sales@), classifies each message (quote request, complaint, invoice, spam, "needs the boss"), pulls the relevant customer history, and prepares a draft reply. A human reviews and clicks send.
This is my favourite starting point because the agent never sends anything on its own. For a team getting 80–150 emails a day, sorting and first drafts alone typically free up an hour or more daily, and nothing slips through the cracks on busy Mondays.
2. Quote preparation
A customer emails: "50 units of X, 20 of Y, delivered to Sakai by the 15th." The agent extracts the items, checks your price list and stock sheet, applies your discount rules, and builds a draft quote in your usual template. A salesperson checks it and sends it.
For the Osaka wholesaler, quotes were taking 20–40 minutes each because the information lived in three different spreadsheets. A draft that appears in two minutes, with the tricky lines flagged, changes the rhythm of the whole sales team.
3. Booking and scheduling
Clinics, salons, repair services and consultants spend a surprising amount of time on back-and-forth scheduling. An agent can read a request, check the calendar, propose slots, and pencil in a booking, escalating anything unusual (cancellations, complaints, urgent cases) to a person. If your customers contact you on LINE, this connects naturally with LINE automation.
4. Research and preparation
Before a sales meeting: "Summarise what this company does, recent news, and anything from our past emails with them." Before a bid: "Collect the requirements from these five PDFs into a checklist." Agents that can browse and read documents are good at this kind of legwork. The output is a briefing for a human, so the risk is low, as long as someone checks the facts before relying on them.
5. Back-office data entry across tools
This is the classic: an order arrives by email or PDF, and someone retypes it into the sales spreadsheet, the accounting software, and the shipping system. An agent can read the document, enter the data in each tool, and flag anything that doesn't match. It's essentially what we did for a Kobe logistics company in the project that saved 40 hours a month, and it's where the biggest time savings usually hide.
Notice the pattern: in every case, the agent does the tedious 80% and a human owns the final 20%. That's not a limitation to apologise for. It's the design.
What do AI agents cost for a small business?
There are three layers of cost, and vendors tend to only mention the first.
1. The AI itself. Either a subscription to an off-the-shelf product with agent features, or usage-based fees if the agent runs on an AI model through an API. Pricing changes often and differs widely between providers, so check current pricing pages directly rather than trusting any number in a blog post (including this one). For a single, well-scoped workflow, the raw AI usage is usually not the largest line.
2. Setup and integration. Connecting the agent to your email, spreadsheets, accounting or booking system, writing the rules ("orders over ¥500,000 always go to the manager"), and testing with real examples. This is where most of the effort goes.
3. Ongoing oversight. Someone needs to review the agent's work, handle the exceptions, and update the rules when your prices or processes change. Budget a few hours a week at first, falling as trust builds.
For reference, our AI projects at SolidTech typically start from around ¥30,000–50,000 per month for a single workflow (we confirm exactly what's included in a free consultation). The right question isn't "how much does it cost?" but "how many hours a month does it give back, and what's an hour worth to us?" If an agent saves 30 hours a month at ¥2,500 an hour, that's ¥75,000 of time you can put somewhere more useful.
Where AI agents go wrong
I'd rather you hear this from me before you sign anything.
They make mistakes, confidently
Large language models still get things wrong: a misread quantity, a wrong date format, an invented detail that sounds plausible. An agent multiplies this, because one wrong step early on can carry through every step after it. The fix isn't hoping for a better model. It's designing checks: validation rules, comparisons against your real data, and human approval for anything that leaves the building.
Nobody is watching
The most dangerous phrase in an AI project is "it runs by itself now." Agents need an owner: a named person who looks at what it did yesterday, handles flagged cases, and notices when something drifts. No owner, no agent.
Data privacy is an afterthought
An agent that reads your inbox sees customer names, addresses, sometimes health or financial details. Before connecting anything, ask: where is this data processed? Is it used to train the provider's models? Can we limit what the agent can access? In Japan, handling customer data falls under the Act on the Protection of Personal Information; the Personal Information Protection Commission publishes official guidance. Business plans from the major AI providers generally offer options so your data isn't used for training, but read the terms for the specific product you choose.
Also give the agent the *minimum* access it needs. If it only drafts replies, it doesn't need permission to send them. If it only reads the price list, it doesn't need write access to accounting.
Starting too big
"Let's have an agent handle all of customer service" is how projects die. Too many edge cases, too many systems, too much change for the team at once. Six months later, nobody trusts it and everyone has gone back to doing it by hand. I've written before about why DX projects fail in small companies, and agents are no exception.
How to start with AI agents: a step-by-step plan
Here's the approach we use with clients. It's deliberately boring.
- Pick one task. Repetitive, rule-based, and time-consuming. Ideally something your team complains about. Write down how long it takes today.
- Map the steps on paper. Where does the information come from, which tools does it touch, what decisions are made, and what counts as an exception? If you can't describe it, an agent can't do it.
- Start in "draft mode". The agent prepares, a human approves. Nothing gets sent, booked or entered without a click from a person.
- Test with real history. Run the agent on last month's emails or orders and compare its output with what your team actually did. This shows you the error types before any customer sees them.
- Run in parallel for 2–4 weeks. Use it on live work, still with human approval, and log every correction. Corrections become new rules.
- Loosen the leash gradually. Once a category of task is consistently right (say, standard reorders from repeat customers), let the agent complete those automatically, while everything else still goes to a person.
- Measure and decide. Hours saved, error rate, team feedback. Then, and only then, choose the next task.
Most single-workflow projects reach step 5 within 4–8 weeks. That's fast enough to see results this quarter, and slow enough that nobody gets hurt.
Frequently asked questions
Are AI agents worth it for a company with fewer than 10 employees?
Often, yes, as long as there's a task that eats several hours a week. Small teams feel the benefit most, because there's no one else to hand the tedious work to. Start with inbox triage or quote drafting; they're low-risk and quick to set up.
Can I just use the agent features in ChatGPT or other off-the-shelf tools?
For personal research and one-off tasks, absolutely, and it's a good way to learn. For a repeatable business process that touches your customer data and internal systems, you'll usually want something configured around your own tools, rules and approval steps.
Will an AI agent replace my staff?
In the projects I've worked on, no. It replaces the retyping, the sorting and the chasing. The people who did that work move to tasks that need judgement and relationships, which is usually where a small company is short-handed anyway.
What's the difference between AI agents and RPA?
Traditional RPA follows fixed click-by-click scripts and breaks when a screen or format changes. AI agents can read unstructured content (emails, PDFs, messy requests) and adapt to variations. In practice, the best setups often combine both: fixed automation for the stable parts, AI for the messy parts.
Where to go from here
If an email like the one my Osaka client received has landed in your inbox, here's my advice: don't buy an "autonomous agent". Pick one painful task, keep a human in the loop, and measure.
That's exactly how we approach AI solutions and business automation at SolidTech: one workflow at a time, built around the tools you already use. If you'd like an honest opinion on whether an AI agent makes sense for your business, and which task to start with, get in touch for a free consultation. We'll tell you if the answer is "not yet."
