Building AI Systems That Actually Stick

I've been building software systems for over 20 years. Before I focused exclusively on AI automation for field service businesses, I led engineering teams at companies ranging from an outdoor fitness brand with 100,000 subscribers to an ecommerce platform serving 50,000 merchants. I've seen a lot of technology rollouts — good ones and bad ones.

The bad ones fail in a predictable way.

Why Most AI Projects Fail

Here's the pattern: a business owner reads about AI, gets excited, buys a tool or hires someone to build something, and launches it. Three months later it's not being used. Either the team reverted to their old process, the tool broke and nobody knew how to fix it, or it turned out the problem it solved wasn't actually the most important problem.

This isn't a technology failure. It's an implementation failure.

The technology works. n8n, Make, OpenAI, Anthropic — these are mature, reliable platforms. The failure happens upstream, in the design phase, when the person building the system didn't understand the operation well enough to build the right thing.

The Three Failure Modes I See Most Often

1. Building for the ideal process, not the real one.

When you ask a business owner to describe their workflow, they usually describe how it should work, not how it actually works. The ideal process is clean and logical. The real process has workarounds, exceptions, manual patches for system gaps, and idiosyncratic habits that the team has developed over years.

An automation built on the ideal process breaks when it hits reality. The workarounds aren't accounted for. The exceptions cause errors. The team bypasses the system because it doesn't match how they actually work.

The fix: spend time observing and documenting the real process before building anything. This is most of what an operations audit is for.

2. Over-automating too fast.

There's a temptation to automate everything at once. Why not? If you can automate one thing, you can automate ten things.

The problem is adoption. When a team gets hit with ten new automated workflows at once, it's overwhelming. They don't understand the system, they don't trust it, and when something breaks (it always breaks somewhere) they lose confidence in the whole thing and go back to manual.

The right approach is to automate one thing, let it run for a month, let the team get comfortable with it, and then add the next thing. Adoption builds confidence, and confidence builds adoption.

3. Building without ownership.

Every automation needs a human owner — someone who knows how it works, gets notified when something breaks, and has authority to adjust it. Without an owner, the automation becomes a black box that nobody understands or maintains.

For most small businesses, the owner is the owner. That means they need to understand, at a basic level, what the system does and how to tell when it's not working. Building systems that are too complex for the client to own is a consultant's failure, not a client problem.

What Stick Looks Like

A system that sticks has a few characteristics:

It solves the most painful problem first. Adoption is easiest when the first automation makes an obvious, immediate difference. Start with the thing that's costing the most time or money. When the team sees the impact, they become advocates rather than skeptics.

It's simple enough to explain in two sentences. "When someone fills out the contact form, they get a text within a minute and a follow-up email if they don't respond in 24 hours." That's it. The team understands it. They can verify it's working. They know when it's not.

It has a fallback. Every good automation has a human fallback for edge cases. If something doesn't fit the pattern, it doesn't crash silently — it flags for a human. This is important for trust: the team knows the system isn't going to drop something important.

It's maintained. At minimum, someone checks in on it monthly. APIs change. Integrations break. Tools update. An unmonitored automation is an automation that will quietly fail at some point and damage the customer experience.

The Investment Worth Making

The most expensive part of building a good AI system is the upfront thinking — the audit, the process documentation, the design work that happens before a line of automation is written. This is also the part most people want to skip.

Don't skip it. A week of careful design saves months of troubleshooting and replacement. The systems that stick are the ones that were understood before they were built.

I've built production AI systems at scale and for small businesses. The principles are the same: build for the real process, start with the highest-impact problem, and keep it simple enough to maintain. [Learn more about how I work](/about), or [book a discovery call](/contact).