AI Tools We Tried and Abandoned (And Why)

We talk a lot about our wins — the AI tools that saved us hours, cut costs, and genuinely made our dental practice run better. But if we only told you about the things that worked, we'd be lying by omission. Plenty of our ideas crashed and burned.

This is the post about those ones. The tools we built, tested, and quietly shelved. Not because AI is bad, but because we aimed them at the wrong problems, or underestimated the human side of things, or simply got carried away with what felt possible.

Here are four AI projects that didn't make it — and what each one taught us.

1. The all-knowing practice chatbot

What we built: A chatbot that could answer any question a patient might have about our practice. Opening hours, fees, treatment information, insurance policies, parking, post-operative care instructions — everything. We fed it every document, every FAQ, every page of our website, and pointed it at the front desk like a digital receptionist who never takes a lunch break.

Why it seemed like a good idea: Our reception team fields the same questions dozens of times a day. "Do you do payment plans?" "How much is a crown?" "Can I claim through my health fund?" If a chatbot could handle even half of those, it would free up significant time.

How it failed: It hallucinated. Confidently. A patient asked about a procedure we don't offer, and instead of saying "I'm not sure about that," the chatbot made up a detailed answer. It quoted fees we don't charge. It described a booking process that doesn't exist. The broader the scope, the more opportunities it had to invent things. And the problem with a chatbot that's right 85% of the time is that you can't tell which 15% is made up.

What we learned: Narrow AI that does one thing well beats broad AI that does everything badly. A chatbot that only handles three tasks — booking, cancellations, and opening hours — can be made nearly bulletproof. A chatbot that tries to answer every possible question will eventually tell someone they can claim a procedure on Medicare when they can't. That's not a minor inconvenience. That's a trust problem.

What we did instead: We scrapped the all-knowing approach and built focused tools for specific tasks. Our after-hours phone answering system handles a tight scope and does it reliably. No improvisation, no hallucination, no making things up when it doesn't know the answer.

2. AI voice calls for appointment reminders

What we built: An AI that would phone patients the day before their appointment to remind them. It used text-to-speech that sounded reasonably natural, could confirm or reschedule, and logged the outcome back to our practice management system. Technically, it worked quite well.

Why it seemed like a good idea: No-shows cost dental practices a fortune. A typical missed appointment is 30–60 minutes of dead chair time. Reminder calls reduce no-shows significantly, but making those calls manually is tedious and time-consuming. An AI caller seemed like the perfect solution — tireless, consistent, and scalable.

How it failed: Patients hated it. Not "mildly disliked" — genuinely hated it. People don't want to receive robot phone calls. Some hung up immediately. Others felt it was impersonal and told us so. A few said it felt like a scam call. We were solving a real problem, but the solution made patients feel like they were dealing with a call centre, not their local dentist. When your reminder system actively annoys the people you're trying to retain, it's doing the opposite of its job.

What we learned: Sometimes the low-tech solution is better. Not everything needs to be clever. The best tool is the one people actually respond well to — and for appointment reminders, that turned out to be a plain SMS. No AI. No voice synthesis. Just a text message that says "Hi Sarah, just a reminder you've got an appointment tomorrow at 2pm. Reply Y to confirm or call us to reschedule."

What we did instead: Automated SMS reminders. Simple, cheap, and patients actually appreciate them. Confirmation rates went up. Complaints went to zero. We spent two weeks building a sophisticated AI voice system when a $0.05 text message did the job better.

3. AI-generated social media posts

What we built: An automated pipeline that generated social media content for the practice. It would pull topics from a list, generate a post with AI, create a caption, suggest hashtags, and schedule it for publishing. Hands-free social media management. The dream.

Why it seemed like a good idea: Nobody at the practice had time to manage social media consistently. We'd post in bursts — three posts in a week, then nothing for a month. An automated system would keep the feed active without anyone having to think about it.

How it failed: The content was technically correct and completely soulless. Every post read like it was written by a committee. "Did you know that regular dental check-ups are important for maintaining oral health? Book your next appointment today!" It was the kind of content you scroll past without registering it exists. Generic, safe, and utterly forgettable.

Social media works when it has personality. The posts that actually get engagement are the ones where someone shares a genuine moment, a behind-the-scenes photo, or an honest opinion. AI can't fake that. It can produce content that looks like a social media post, but it can't produce content that feels like it came from a real person.

What we learned: AI should assist creative work, not replace it. There's a meaningful difference between "AI writes the post" and "AI helps a human write the post." The first produces wallpaper. The second saves time while keeping the human element that makes content worth reading.

What we did instead: We still use AI in our content workflow, but as a drafting tool, not an autopilot. A human picks the topic, decides the angle, reviews the draft, and adds the personality before it goes out. It's faster than writing from scratch but doesn't sacrifice the authenticity that makes social media actually work.

4. Real-time clinical decision support

What we built: A prototype that listened to clinical notes and patient symptoms, then suggested possible diagnoses and treatment options. Think of it as an AI second opinion that could flag things a clinician might have missed.

Why it seemed like a good idea: Clinical decision support systems exist in hospitals and large health networks. They catch drug interactions, flag unusual lab results, and prompt clinicians to consider differential diagnoses. Bringing something like that to a small dental practice seemed like a natural extension of AI into clinical care.

How it failed: Three ways, all of them serious. First, the liability implications were enormous. If an AI suggests a diagnosis and the clinician follows it and it's wrong, the legal and ethical territory is a minefield. Second, the complexity was beyond what we could responsibly build. Clinical decision-making draws on years of training, tactile examination, patient history, and professional judgement that can't be reduced to pattern matching. Third — and most importantly — the clinicians didn't trust it, and they were right not to. A suggestion from an AI system that's sometimes wrong is worse than no suggestion at all, because it adds noise to an already complex decision.

What we learned: Know where AI's boundaries are. Admin automation? Brilliant. Sorting emails, answering phones, transcribing notes, managing schedules — AI handles these reliably because the stakes are low and the tasks are well-defined. Clinical decision-making? Not ready. Maybe not for a long time. The consequences of getting it wrong are too high, and the problem is too complex for current AI to handle responsibly.

What we did instead: We shelved it permanently and redirected that energy into admin automation where AI can genuinely help clinicians — tools that save them time on paperwork so they can focus on the clinical work they're trained for.

The meta-lesson: fail fast, fail cheap

Here's the thing about all four of these failures: each one took roughly one to two weeks to try. Not months. Not six-figure budgets. We had an idea, we built a quick version, we tested it in the real world, and when it didn't work, we stopped. The total cost of all four failures combined was probably less than one month of a SaaS subscription we'd never use.

The cost of not trying would have been worse. We'd still be wondering whether a practice chatbot could handle everything. We'd still be debating whether AI voice calls were worth pursuing. We'd still be guessing instead of knowing.

If you want more detail on what we learned from the tools that did work, we've written about nine months of running AI in a real practice. And if any of these failures sound familiar, our piece on why AI projects fail digs into the structural reasons these things go wrong.

The honest pitch

We're not going to pretend we have a perfect track record. Nobody does, and anyone who claims to is either lying or hasn't tried enough things. What we will say is this: we've made these mistakes so you don't have to. We know where AI works brilliantly for small businesses, and we know where it falls flat. We'll tell you both.

If you're thinking about AI for your business and you want a conversation with someone who'll be honest about what's realistic — including the things that probably won't work — get in touch. We'd rather save you from a bad idea than sell you one.