An AI barnacle is a flashy tool bolted onto a process without a clear, measurable purpose. It looks impressive in a slide deck. It doesn't drive revenue, reduce cost, or make anything meaningfully faster.
Every organisation has them. The chatbot on the internal wiki that nobody uses. The AI analytics dashboard that duplicates what a SQL query already does. The "intelligent" classifier that's less accurate than three lines of regex.
They accumulate because saying "we're using AI" feels like progress. It's not. It's mostly theatre.
The pattern
AI barnacles follow a predictable lifecycle:
- Someone sees a demo and gets excited
- A team gets tasked with exploring an AI solution
- They bolt an LLM onto the nearest process
- It works in the demo environment
- It half-works in production
- Nobody can prove it's delivering value, but nobody wants to kill it either
The root cause is starting from the technology instead of the problem. "We need an AI strategy" is almost always the wrong framing. You need an operations strategy that uses AI where it actually helps.
When you start from "we should use AI," you go looking for problems to attach it to. When you start from "this process costs us $2M a year in manual effort," you evaluate every option, and AI might be one of them. Or it might not.
The cost
Barnacles aren't free. Every AI integration adds latency, unpredictability, maintenance overhead, and direct cost.
LLM calls take time. If the AI step doesn't add enough value to justify the wait, you've made the process slower. LLMs are also non-deterministic, so every call you can replace with deterministic logic is one that won't hallucinate, drift, or surprise you. Models change, prompts need tuning, context windows shift. An AI integration is a living dependency, not something you deploy and forget. And API calls add up. A process handling thousands of requests a day with an unnecessary LLM call on each one is burning money for no return.
The biggest wins don't come from the most advanced AI. They come from the highest-value solution to the actual problem. Sometimes that's a sophisticated multi-agent system. Sometimes it's a lookup table and an if-statement.
The test
Before adding AI to any process, answer three questions.
What metric does this move? If you can't name a specific number, whether that's cost reduced, time saved, revenue generated, or error rate lowered, you don't have a use case. You have a demo.
What's the simplest solution that works? If a rule-based system, a database query, or a basic script solves the problem, use that. Reach for AI when simpler tools can't do the job.
What happens when it's wrong? Every AI system will produce incorrect outputs. If the cost of a wrong answer is high and there's no human check in place, you're building a liability, not a feature.
Strip the barnacles
The organisations getting real value from AI aren't the ones with the most integrations. They're the ones who deploy it precisely where it moves a number that matters, and leave it out everywhere else.
Strip the barnacles. Keep what works. Build from the problem, not the technology.
Views expressed here are my own and don't represent any other parties.