Sixteen years learning exactly why automation fails

For sixteen years I worked in telecom and network infrastructure — the systems underneath everything else, where downtime is measured in seconds and nobody notices you until something breaks.

It was a good career, and it taught me how large systems genuinely behave rather than how the architecture diagram says they do. But the frustration that eventually pushed me out was always the same: I could see the automation that would have saved hundreds of hours, and I could not build it.

Not because it was technically hard. Because the integration touched a vendor system whose contract forbade it. Because the security policy was written for a threat model that predated the problem. Because getting approval meant eighteen months and three committees, by which time the tool we had scoped was two generations out of date.

The thing that changed my mind

What changed was watching how quickly the same problems fall away in a smaller business. No procurement cycle. No committee. The person who understands the bottleneck is often the person who can approve the fix. A workflow that would have taken an enterprise a year to even schedule can be running in six weeks.

That gap is the whole business. Small and medium businesses have a structural advantage over the enterprise on exactly this kind of work, and almost nobody tells them so.

How I think about the tools

I am deliberately not loyal to any one model or vendor. Frontier models from the likes of Anthropic and OpenAI are extraordinary at some things and needlessly expensive at others. Open-weight models running on hardware you control are the right answer when data sensitivity or per-call cost matters, and the wrong answer when you need the absolute best reasoning available.

Choosing between them is an engineering decision that should be made per workflow, written down, and explained to you in language you can challenge. Anyone who tells you there is a single right answer is selling something.

What I will always tell you

If a workflow should not be automated, I will say so. If a spreadsheet formula would beat a language model, I will say that too. I would rather lose a project than build something that quietly makes your business worse, because in a business this size my reputation travels faster than any marketing I could write.