Six Months In: What I've Actually Learned About AI In Small And Medium-Sized Businesses
About six months ago, I started really digging into AI for our own small business. Not as a hypothetical, not as a "what should we offer customers" exercise, but as a "let's actually use this and see what happens" experiment.
We're VoIP PBX Express. We're a phone systems shop. We move fast. And I figured if I was going to talk to customers about AI, I'd better have my own house in order first.
Here's the thing nobody loves to hear: AI is coming whether we like it or not. Customers are going to expect it. Competitors are going to use it. Employees are going to bring it in through the back door whether you've sanctioned it or not. The only real question is whether you have a plan or you're getting dragged along.
A lot of people hear "AI" and immediately think "replacing people." And honestly, to some degree, that can be true — there are tasks AI is going to absorb that humans used to do. But that's only one piece of the picture, and in my experience it's not even the most interesting one. The bigger story is AI taking on the jobs nobody likes doing — the repetitive admin work, the data entry, the manual reconciliation — and freeing your key employees to focus on the work that actually moves the business forward. Done right, AI doesn't shrink your team. It makes the team you have significantly more productive.
This article isn't a sales pitch. It's a starting point — a framework for beginning to think about your AI plan for 2026 and 2027. Because if you're not thinking about it now, you'll be thinking about it in twelve months when you're behind.
Here's the short version of what I've learned: there are three places AI fits inside an SMB right now, they each solve a different problem, and most of the conversations I have with business owners blur all three together into one big confusing thing. Once you separate them out, picking a starting point gets a lot easier.
I'll walk through all three with real examples from our own shop, then take a swing at where I think the next 12 months go.
Why I Started Doing This
The motivation was pretty simple. Three goals:
Cut costs where I could
Get rid of the jobs nobody enjoys doing — and the jobs we honestly aren't that proficient at
Be a force multiplier for our team
That second one matters more than people realize. Every business has tasks the team is doing because somebody has to do them, not because anybody's good at them. AI is great at the stuff humans don't want to do, and surprisingly often, it's better at the stuff humans aren't naturally good at either. That includes things like consistent data entry, careful spreadsheet reconciliation, sifting through hundreds of records, summarizing long documents, and following the same procedure exactly the same way every time. Most humans are bad at repetition. AI doesn't get bored.
And the force multiplier point matters. Every business hits seasons where the work outpaces the team, and you can't always hire your way out of it. If AI lets the team you already have punch above its weight, that's the ballgame.
I had no idea where to start either. I just knew I needed to start. So I picked the obvious one for our business and built an AI receptionist for ourselves first.
The Three Places AI Fits In An SMB
Here's how I'd break it down now after six months of poking at it. Each of these is a different shape of AI, and each one solves a different problem.
Type 1
Agentic Workflow Processes
These are AI agents that live inside a sequence of events. Something external triggers them, they run through a defined workflow, and they go quiet until the next trigger. They don't have memory across runs in any meaningful way. They know what they're supposed to do, they have access to a specific set of tools, and they execute.
Arthur Igloo is one of these. A call comes in, Arthur picks up, runs through his workflow (greet, understand, transfer or take a message, send the notification), and that's it. He doesn't remember you on the next call. Each call is self-contained.
Alex Igloo is the next one we built. Alex is our level-one support technician. He monitors our ticket queue. When a ticket comes in, he reviews it against the last 1,500 customer service tickets we've handled, applies our standard operating procedures, and drafts a response. Before he sends anything, he pings me on Slack with "here's the ticket, here's what I think, do you approve?" I type approved, and he runs with it. If I say no, he stands down.
Alex is specifically trained on the unique-enhanced hybrid we run, which is Cisco Call Manager Express on the front end paired with Sangoma on the back. That combination is unusual enough that a generic AI wouldn't be able to figure it out. Alex knows the specific quirks: when do we need a remote session, when don't we, when does a ticket need an appointment vs. just a config change.
The newest member of the team is Apollonia. She's VoIP PBX Express's AI project manager — she shepherds every new system from sale through install and into Year-1 service so nothing falls between the cracks. From the moment a deal closes she's coordinating site surveys, shipping, install scheduling, calendar invites, and post-install follow-ups, keeping customers informed at every step and the VPE team focused on the technical work that matters.
One thing worth flagging here: when most people first think about agentic workflows, they jump to "this replaces an employee." That's almost never how it actually plays out, and it's not a useful way to plan. The better mental model is to take a specific employee, break their job down into the tasks they do every week, and look at each task individually. Some are creative, judgment-heavy, relationship-driven — those stay with the human. Others are repetitive, rule-driven, and follow the same pattern every time. Those are candidates for a workflow process. You're not replacing the employee. You're peeling off the parts of their job that drag them down, and pointing those tasks at AI that can do them faster and cheaper. Your employee gets to focus on the work that actually needs them.
Type 2
Desktop AI Working Alongside You
Different shape. This one sits on your computer and works with you on whatever you're doing. You give it selective access to your browser, your email, your spreadsheets, your CRM. You watch it work. You can stop it whenever. It's not autonomous, it's collaborative.
The product I use here is Claude's Cowork. Anthropic makes it. Twenty bucks a month at the entry tier. Installs on your desktop and you give it working folders for different projects.
The example I keep coming back to: Jane used to spend two full days every month on SIP Express invoicing. Hundred and eleven customers. For each one, she'd request a usage report from a portal, wait for the email, download it, open it, copy the numbers (inbound minutes, outbound minutes, DIDs, E911, CNAM, non-recurring charges), paste them into a master spreadsheet, then build the invoice in QuickBooks. Two days, every month, no creative thought required, just clicks.
We rebuilt it in Cowork. Jane logs into the portal once, then hands the keyboard to the agent. It downloads 111 invoice requests. It goes to her email, finds the 111 attachments, pulls them down, renames them with company name and IP address, saves them all to a Google Drive folder for that month. Then it opens each one, extracts the values, and updates the spreadsheet. The whole thing now takes about 15 minutes. She still does the QuickBooks invoicing herself because we want a human eye on the final number, but the painful part is gone.
Other things I've used Cowork for: cleaning up "Spam Likely" caller ID flags by submitting forms and emails to Verizon, AT&T, and T-Mobile across hundreds of customer numbers. QuickBooks reconciliation that I'd been dreading. A construction law project where I needed to write a letter to a contractor referencing specific Washington State code, with quotes from his own recorded conversations as evidence. I gave Cowork the recordings, made it a legal expert on Washington State construction law, and it produced a really sharp letter.
That last example points to something worth flagging on its own. Cowork can become an expert in pretty much any domain you point it at. I've spun one up as a legal researcher. I've used another as a web designer when I wanted to rework our GoDaddy site without hiring a developer. I could just as easily build one for Google Ads, or HubSpot reporting, or marketing copy, or sales follow-ups, or whatever the job needs. Different working folder, different reference materials, different skills enabled — same underlying tool, completely different role.
The thing about Cowork is it's nuanced. You learn how to use it. Working folders matter. Skills matter (you can turn skills on and off so it doesn't get cluttered). You have to think about what access you're giving it. But once you get the rhythm, it's the most useful tool I've added to my workflow since Excel.
Type 3
AI personal Assistants With Memory
This one's the newest and the least mature. Personal assistants are AI agents that live alongside a specific person, learn how that person works over time, and develop a kind of memory about their preferences, projects, and people.
The framework we're using here is OpenClaw. Open source, brand new, exploded in popularity, and frankly not totally secure yet. So we're treating it carefully and deploying it at the IT layer, not letting individual employees install it ad hoc.
The example: Jane needed something to keep track of everything coming at her. Customer emails, my to-dos, her HubSpot tasks, calendar appointments, personal stuff. The problem with using a workflow agent like Alex for this is that workflow agents don't have memory across runs. The problem with using Cowork for this is that Cowork's memory is tied to a working folder, so if she switches projects, it forgets.
OpenClaw is different. It runs continuously, it has memory across days, and it learns. Every morning it reads her email and surfaces the important ones (and learns over time which ones she cares about). It books and cancels appointments. It pulls her HubSpot tasks for the day. It checks Outlook for the next two days of meetings and lays them out. It maintains a running to-do list she can edit.
It's the closest thing to having a real assistant who knows your job. The catch right now is the security model is still maturing, so this is the one I'd be most careful with. Don't let people install it themselves. Plan it at the IT level.
An Easy Mistake To Make
Here's the trap I fell into early: trying to make one agent do everything.
I started loading more functions onto Arthur. He could handle support tickets too, right? He could do the receptionist work AND the scheduling AND the data lookups? Why have multiple agents when one could do it all?
Because the more you ask an AI to do, the more it hallucinates. That's just the math. If your agent's juggling five different domains with five different sets of rules, it starts confusing them. It mixes operating procedures from one with terminology from another. Things get weird in subtle ways.
The fix is to give each agent a tight job. Arthur answers calls and routes them. Alex reviews tickets. The Cowork instance for invoicing handles invoicing. The Cowork instance for legal research handles legal research. Each one is in its own working folder with its own skills enabled. When you do this, accuracy goes way up and the weirdness goes way down.
The rule: Don't ask one agent to be everything. Multiple narrow agents beat one general one almost every time, especially when you're starting out.
The Security Thing Nobody Wants To Talk About
One of the customers I was just on a call with brought this up and it's worth saying out loud: a lot of people are nervous about AI for security reasons, and they should be.
If you ask ChatGPT a question on a personal account, it can retain what you said and use it to improve future models depending on your settings. So if your marketing person pastes in a sensitive customer list to summarize, you should know exactly what your account's data policy says before you do that. For most public chatbots, you have to deliberately opt out.
The way around it isn't "don't use AI." It's pick the right deployment for the right data. Vendor-backed AIs (like the John Deere AI for parts lookup that this customer's team uses internally) keep data in-house. Enterprise OpenAI accounts have data isolation by default. Cowork running locally on your desktop with your own subscription gives you a lot more control than copy-pasting into a public chatbot.
For the stuff that actually matters, choose the deployment carefully. For marketing copy that doesn't contain anything sensitive, the public tools are fine.
Where This Goes In The Next 12 Months
I'm going to be honest, predictions in this space age fast. What I'm sure about is the three categories above will all mature significantly, and most SMBs will be using all three by this time next year whether they realize it or not. Here's where I'd put my chips:
The Agentic Workflow Space Will Commoditize.
AI receptionists, AI support agents, AI sales qualifiers, AI scheduling agents — these are going to be everywhere. The vendors that win are the ones that integrate properly with the systems businesses already use. The ones that lose are the ones that ask you to forward calls or replatform your phone system. We've already seen this play out in vertical SaaS, and the same forces apply here.
Cowork-Style Tools Become The New Excel.
I really mean this. The first time I sat down with Cowork, I had the same feeling I had the first time I used Excel in the 90s. There's a learning curve, but once you're up the curve, you can't imagine going back. Within 12 months, I think every productive knowledge worker has one of these on their desktop the way they have a browser. The companies that get good at deploying them well will pull ahead of the ones that don't.
Personal Assistants Get Their Security Story Figured Out.
OpenClaw and similar frameworks are going to mature. The security model is the only thing holding them back from being adopted by every operations manager in America. Once the enterprise-grade version exists, it'll be a tidal wave. People are tired of drowning in their inbox.
The Cost Curve Keeps Falling Fast.
The Arthur of today uses top-tier voice models, top-tier reasoning models, top-tier transcription. It costs us real money to run. By next year, the same quality will cost a fraction. AI receptionist services that price like a premium product today will be accessible to businesses that couldn't justify them before, and a lot more SMBs will pile in once that happens.
The Biggest Uncertainty: How Fast Can Humans Adopt It?
The technology moves faster than people do. Every customer call I have, the same pattern shows up: somebody in the company gets it and wants to move, and the rest of the team is hesitant or confused. The bottleneck isn't the AI. It's training, change management, and getting your team comfortable with new ways of working.
That's actually the part I think we can help with the most. Hand you a tool and walk away, you might never use it. Sit with you, show you how to set up your first working folder, train your team in a workshop, demo it on actual problems from your business — that's how adoption happens.
Where To Start If You're Just Getting Going
If you're a business owner reading this and thinking "okay, where do I actually begin," here's what I'd do.
Start with Cowork. Twenty bucks a month, install on your desktop, give it a working folder, and pick one painful repetitive task you do every week. Just one. Spend an afternoon teaching it to do that task. The first time it works, you'll get the bug.
Don't start with the autonomous agents. Those are great, but they're harder to set up right and the failure modes are subtler. Cowork shows you what it's doing in real time, which is what you want when you're learning.
Don't try to do too much. Pick small, repeatable, painful tasks. Build confidence. Then expand.
What We're Doing About It
Here's the part I'd be remiss not to mention. We're spinning Arthur Igloo off into a dedicated AI consulting firm focused specifically on grass-roots AI adoption inside small and medium-sized businesses.
The model is the same one that's worked for our phone system business for over a decade. A high degree of technical expertise. A reasonable cost. Implemented quickly and efficiently. No enterprise-scale price tags. No 18-month deployment timelines. No army of consultants billing $400 an hour while you wait for a kickoff meeting.
What we'll do is what we already do for ourselves. Help you pick the right tool for each job. Build the agentic workflows that fit your business. Set up Cowork for your team and teach them to actually use it. Plan the personal assistant rollout for the people who'll get the most leverage from it. Keep an eye on what's coming next and make sure you're not getting blindsided.
The Bottom Line
Six months ago I had no idea what any of this was. Today, our business runs differently, our team gets more done, and we cut multi-day projects down to minutes. That's not a vendor pitch. That's just what is happening.
AI in your business isn't a question of if. It's a question of whether you start now with a plan or whether you wait and play catch-up. The businesses that start now will have a year of practical experience by the time the businesses that wait are still figuring out what to call their first project.
Use this article as a starting point for thinking about your 2026 and 2027 plan. The three categories give you a way to organize the decision. Pick one to start with. Build from there. And if you want help thinking through where AI fits in your specific business, give me a call: (360) 515-2700. Happy to compare notes.