Ten months in: what I've actually learned about AI automation for business

Why I started with our own company first

If I was going to talk to customers about AI, I had better have my own house in order. So I built an AI receptionist for our own shop, VoIP PBX Express, our voice arm.

Ten months later there is a whole roster. Alex handles first-line support, Arthur answers the phone, Apollonia manages projects, Athena keeps the books, and Andrew Igloo, our AI network administrator, watches customer networks. Behind them sit an invoicing assistant, a website builder, a marketing agent and a Google Ads agent. I built the agents in Claude Code, and I have no coding background.

Replacing people was never the goal. I wanted to cut costs, get rid of the jobs nobody here enjoys or is good at, and let a small team punch above its weight.

Where does AI automation for business fit?

There are three places AI fits inside a small business, and most of what gets sold as AI for small business blurs them together. Separate them and the starting point gets a lot clearer.

1. Workflow agents that run a tight job and go quiet

Something triggers them, they run a defined workflow, and they go quiet until the next trigger. They keep no memory between runs.

Arthur, the AI receptionist built by Arthur Igloo, our AI arm, is one. A call comes in, he picks up, transfers or takes a message, and sends a summary. He does not remember you next time, and he does not need to.

Alex is our first-line support tech, trained on about seventeen hundred past tickets, our user guide and our procedures. When a ticket lands he drafts a reply and posts a summary in our team chat: what the customer wants, the draft, what is still missing, and his reasoning. I can ignore it, approve it, approve it with notes, or say "I've got this," and he stands down. Once I approve, Alex can go back and forth with the customer about three times to gather what he needs. Then he books an appointment and hands our engineer a summary with every detail already in the format we use.

Nothing reaches a customer that a human has not had the chance to stop. One refinement came up on a recent training call with a long-time customer of our voice arm: if nobody here has acted on Alex's draft within a few hours, he sends it himself. I'm adding that. The draft still lands in our team chat first.

Apollonia, our project manager, coordinates site surveys, shipping, install scheduling and follow-ups from the moment a deal closes. Andrew is a close cousin who never goes quiet; more on him below.

2. Desktop AI working alongside you

This one sits on your computer and works with you. You give it selective access to your browser, email and spreadsheets, you watch it work, and you can stop it any time. I use Claude Cowork, from Anthropic, with a working folder for each project.

Jane used to spend two full days every month billing our phone-line customers, one usage report at a time. In Cowork she logs into the portal once (it never holds her password), hands over the keyboard, and the job takes about fifteen minutes. She still does the final QuickBooks invoice, so a human eye is on the number a customer sees.

3. Personal assistants with memory

This is the newest and least mature of the three. The assistant lives alongside one person, runs continuously, and over time learns how that person works. Jane's reads her email each morning and picks out the ones that matter, and it keeps her calendar and to-do list. The software behind it is very new and its data protection is still maturing, so we set it up ourselves as part of the IT we manage.

Why do many narrow agents beat one general agent?

Because the more you ask one AI to do, the more it makes things up. I fell into this trap early by trying to make one agent do everything. Arthur was answering calls well, so why not have him handle tickets and scheduling too? One agent juggling five jobs with five sets of rules starts mixing them up.

The fix is a tight job for each agent. Arthur answers calls, Alex reviews tickets, Athena keeps the books, and the invoicing folder does invoicing. Many narrow agents beat one general agent almost every time, so if you are shopping for an AI agent for business, ask first how narrow its job is.

Is AI safe to use with business data?

Match where the AI runs to how sensitive the data is. People are nervous about AI for security reasons, and the concern is fair: a public chatbot on a personal account may keep what you typed to train future models unless you opt out. Know the data policy before anyone pastes in a customer list.

A business account that does not train on your data, or Cowork on your own desktop, gives you far more control than a public chat window. Treat AI access as one more part of your network security.

The other side of security: what I am seeing on phone systems

Phone systems are being hit for international toll fraud. We used to see one customer security event every two or three years; this year it has been eight or nine. One Sunday I woke at 2am to charges streaming out of our account: overseas calls from a customer account that had international calling turned on. I shut it down after about twenty minutes. International calling is now off for everyone, and a custom access list (a rule that says which addresses are allowed to talk to the phone system) has held since.

AI has lowered the bar for the people doing this. A person cannot listen to a network all day, but an AI can, and that is Andrew's job. He sits on the customer's network around the clock, and when something looks odd he alerts a second AI at The Igloo, which pages a human. The numbers are in our post on the network threats facing small businesses.

The bottleneck is people, not the technology

The technology moves faster than people do. On nearly every customer call one person gets it and wants to move, and the rest hang back. Getting the rest of them on board is a training job.

When a team is nervous, I start here: AI is a computer program that speaks very good English and wants to please you. It is not sentient, and it is sometimes frustratingly stupid. Right now it is nothing to fear. The training write-up I link below walks through how the chat window, the desktop tool and the agent-building tool differ.

Ten months of training people taught me one thing that write-up does not cover: people trust the tool once it has disagreed with them. On the training call I mentioned earlier, the two IT staff liked that it pushed back, "you did it this way, but it should have been done another way," instead of agreeing with everything.

Hand someone a tool and walk away and they may never open it. Sit with them and work one of their own real problems, and they will use it. We did that on the same call, turning one morning of their phone system's call records into a report they can now rerun every month; the whole session is written up here. Once your team has watched AI do the desk-level work, you stop paying a consultant to do it.

Where should a small business start?

Start with the desktop tool. Install Cowork, give it a working folder, and pick one painful, repetitive weekly task. Just one. Spend an afternoon teaching it, and the first time it works you will get the bug. Autonomous agents come later; their failures are subtler.

One business we work with used this approach to plan a Google Workspace to Microsoft 365 migration with its own IT person instead of hiring a consultant.

What this means in SW Washington

The businesses we are built for have ten to seventy-five people, and every one of them has a Jane: the person who holds the office together and spends two days a month on something a computer should do. The help should come from the local IT team that already does your network and firewall, because that is who can sit in your office and set it up. That is the whole idea behind Wallaby.

The other place is the phone. A dental office in Olympia or a warehouse in Centralia loses calls over lunch and after five, and an AI receptionist that lives inside the phone system fixes that in about a week with the phones you already have.

Ten months in, the thing I am surest of is that the two-day jobs are where to start, and the free on-site assessment is where we find yours. What we set up is on our AI services page.