What We Learned Running AI in a Real Business for 9 Months

We've been building and running AI tools in our dental practice for nine months now. Not experimenting. Not piloting. Actually using them, every day, with real patients and real staff.

Some of it went brilliantly. Some of it was harder than we expected. And a few things surprised us completely. This is the honest version — no marketing spin, no "AI transformed our business overnight" nonsense. Just what actually happened.

What worked immediately

Three things delivered value from almost day one.

The AI phone receptionist. This was the fastest return on investment we've seen from any tool we've built. It answers every call, 24/7, knows our practice inside out, and books directly into our schedule. Patients loved it immediately — no hold music, no voicemail, no "we'll call you back." Staff loved it too, because it meant they could actually focus on the patients in front of them instead of constantly picking up the phone. We wrote about the build process in an earlier post, but the short version is: it just worked, and it kept working.

Email automation. We set up AI-powered inbox management early on, and it's been the quietest success story. It classifies incoming emails, labels them, unsubscribes from junk, and routes anything important to the right person. We genuinely forget it's running most of the time — which is exactly the point. The best automation is the kind you stop thinking about.

What took longer than expected

Not everything was plug-and-play.

Staff adoption was slower than we hoped. Even the tools that worked perfectly took two to three weeks before staff actually used them consistently. There's a gap between "this tool is available" and "this tool is part of my routine," and we underestimated how wide that gap is. People don't resist change because they're difficult — they resist it because the old way is automatic and the new way requires thinking. It takes time for the new way to become automatic too.

Getting the AI's tone right was fiddly. Our first version of the phone receptionist was too chatty — it rambled and patients got impatient. We dialled it back, and it became too robotic. Then we made it warmer, and it started sounding like a customer service script. It took several rounds of iteration, listening to real call recordings, and tweaking prompts before it felt natural. Nobody tells you how much time you'll spend on personality tuning, but it matters more than almost any technical feature.

Integrating with existing systems was painful. Practice management software is not built for integration. APIs are poorly documented, inconsistently maintained, and sometimes just broken. We spent more time reverse-engineering how our existing systems worked than we did building the AI tools themselves. If you're planning to connect AI to legacy business software, budget twice as much time as you think you'll need. Then add a bit more.

What surprised us

After-hours usage was much higher than we expected. We knew people called outside business hours, but we didn't realise how many. Turns out a huge number of patients want to book appointments at 9pm after the kids are in bed, or at 6am before work. Before the AI receptionist, those calls went to voicemail and most of them never called back — they just booked with someone who answered. We were losing patients we didn't even know we were losing.

Our AI memory system became unexpectedly essential. We built a tool to capture notes, decisions, and context into a searchable AI memory — mostly as an experiment to help us keep track of everything we were building. Within a couple of months, it became one of the most valuable tools in the entire stack. Being able to ask "what did we decide about the referral workflow?" and get an instant, accurate answer changed how we work. We didn't plan for it to be important. It just became important.

Staff went from sceptical to enthusiastic faster than we expected. The same people who were cautious in week one were coming to us by month two asking "can the AI do this too?" Once they saw one tool work reliably, they started seeing opportunities everywhere. Scepticism isn't the enemy of adoption — it's the precursor. People who ask hard questions early become your best advocates later.

What we'd do differently

Start with fewer tools. We were excited and built too many things in parallel early on. Some of them were half-baked because our attention was split. If we started again, we'd pick one tool, nail it completely, make sure it was rock-solid and fully adopted, and only then move to the next. Depth beats breadth every time.

Involve staff from day one. We made the classic mistake of building tools and then presenting them to the team. "Here, use this." That's backwards. The people who use a tool every day know things the builder doesn't — edge cases, workflow quirks, things that seem minor but completely change whether something gets used. We should have had staff in the room from the first conversation, not after deployment.

Set clearer success metrics upfront. For the first few tools, we didn't define what "working" actually meant. Is 80% accuracy good enough? How many calls per day justify the phone system? At what point does time saved outweigh time spent maintaining? We were sometimes guessing whether something was successful, which made it hard to prioritise improvements. Now we define metrics before we build, and it makes every decision easier.

By the numbers

We're not going to pretend these are precise — some are estimates, and your results would be different. But for context:

• Tools built: 16 (ranging from simple scripts to full applications)
• SaaS subscriptions eliminated: roughly $700–$1,700/month, depending on how you count
• Staff hours saved: approximately 20–25 hours per week across the practice
• Tools abandoned or rebuilt from scratch: 2–3 (not everything works first time, and that's fine)

If you're curious about the financial side, we broke down the real costs in our post on what AI can actually do for a small business. The short version: the savings are real, but they take time to materialise.

The biggest lesson

AI is easy. Change management is hard.

That's it. That's the whole lesson. The technology is the simple part — getting an AI to answer phones, sort emails, or transcribe notes is a solved problem. The hard part is getting people to trust it, use it, and change habits they've had for years.

Every "AI project" is really a people project with a technology component. If you focus only on the tech and ignore the humans, you'll build something clever that nobody uses. We've seen it happen to other businesses, and we nearly did it ourselves a couple of times.

The businesses that succeed with AI aren't the ones with the best technology. They're the ones that take the time to bring their team along. That means explaining why, not just how. It means listening to concerns instead of dismissing them. It means being patient when adoption is slow and celebrating when it clicks.

If you're thinking about bringing AI into your business, read our post on why AI projects fail before you start. Most of the failure modes are avoidable — but only if you know about them upfront.

Nine months in, we're more convinced than ever that AI is genuinely transformative for small businesses. But "transformative" doesn't mean "instant." It means doing the work, learning from mistakes, and getting a little better every week.

If you'd like to talk about what AI could do for your business — honestly, without the hype — get in touch. We're happy to share what we've learned.