Practical writing on custom AI agents, workflow automation and what it takes to run them in production — costs, failure modes and where to start.
These three get sold as competitors and they are not — they solve different problems, and two of them are cheaper than the third. Here is the single question that decides which one your process needs, and what it costs to get the answer wrong.
The framing hides a third option, and the deciding factor is almost never the build price — it is who owns the maintenance in two years. Here is when buying is genuinely right, when building is, and the middle path most companies end up on.
A working glossary for anyone reading an AI automation proposal. Thirty terms defined in plain English, grouped by where you will meet them, with a note on which are load-bearing and which are decoration.
This question is usually asked backwards. Fine-tuning is requested to solve a knowledge problem, which is the one thing it does not reliably do. Here is what each approach actually changes, and the test that tells you which one your case needs.
The word “agent” now covers everything from a support chatbot to a scheduled script, which makes it nearly useless in a sales conversation. Here is a definition narrow enough to be worth something — and an honest account of when you do not need one.
Every quote for a custom AI agent is a different number, and the reason is rarely the model. Here are the five things that actually move the price, why the running cost should fall rather than rise, and what makes an estimate balloon after work starts.
Almost every failed AI agent worked in the demo. The gap between a convincing demo and a system your business can depend on is made of five specific failures — none of which is the model being bad at its job.
Picking the wrong first workflow is the most common way an automation programme dies — not because the automation fails, but because it succeeds at something nobody cared about. Here is how to score the candidates.
Prefer a reader? Subscribe to the RSS feed.