Starting flashy AI projects is easy. The demo goes well, the video circulates, everyone applauds. What is nearly impossible is keeping that project alive for six months. Profitable AI projects are easy to start and cheap to sustain, because the cost of maintaining them was paid before the first line of code. That asymmetry, what it costs to launch against what it costs to stay alive, is the filter we use to decide what to build.
It is not a fringe view. Gartner estimates that over 40% of agentic AI projects, the kind that acts on its own, will be canceled by the end of 2027. Projects do not die at launch. They die in month six.
What kills projects is not the model
We run our own product, a WhatsApp appointment assistant that handles health data, and it has taught us where a system actually dies. It is never the dramatic outage. It is the degradation nobody watches.
Our worst incident tells it well. A badly calibrated alert started firing on a state that was not a real failure. Within days it emitted 5,667 events and exhausted the monthly quota of the alerting system. All monitoring went blind.
Six days later, thirteen real reminders failed for a clinic, and the corresponding alarm fired thirteen times without a single one leaving the machine. The clinic told us. The system did not.
Three rules came out of that, and we now apply them to everything: every alert carries its own cap, none can exhaust the budget of the others, and silence gets watched too, because zero warnings in 24 hours does not mean all is well.
This has had a name for a decade
We did not discover the asymmetry. In 2015 a team at Google published “Hidden Technical Debt in Machine Learning Systems”, the reference work on the hidden technical debt of these systems, with a figure that stuck. A mature system can be 5% machine learning code and 95% everything else, the data, the infrastructure, the monitoring and the glue between the pieces.
The warning came from an earlier paper by the same authors, which called machine learning the credit card of technical debt, the kind that is easy to open and paid back with interest. It was said by one of the largest operators in the world, a year before most of the industry had touched a model.
A decade on, with generative AI, the debt is the same and the interest has gone up. The same input can return a different answer tomorrow, so everything around the model weighs more, not less. The demo shows the 5%. Day 180 is lived in the other 95.
Staying alive is a system, not an intention
The difference between the flashy project and the profitable one is not talent or model choice. It is that the profitable one was designed to be maintained. In practice, that means concrete pieces:
A test battery that runs before every change. With non-deterministic components, the ones that do not always answer the same input the same way, any tweak can break what worked yesterday without anything visibly failing. If quality drops, the change does not ship.
A record that lets you reconstruct every decision. When the complaint arrives, the useful question is not “which line failed” but “what input, context and rules produced that answer”. Without the record, that question has no answer.
Alarms that get tested by forcing the failure. A defense you have never seen go off is not a defense. Every alert gets verified by deliberately breaking what it watches.
A real fire drill, on a schedule. In our product, a weekly batch of tests creates real appointments, sends real messages, and cleans everything up afterwards. It finds what no simulation finds.
The uncomfortable question for your vendor
If you are evaluating an AI project, the question that will tell you the most is not about the model or the demo. What happens on day 180? Who watches the alarms, who runs the tests, who finds out if the system degrades in silence? A flashy project cannot answer that. A profitable one answers with names.
Our answer is what we call observability by default, and it sits at the center of how we do AI agent development. If you would rather start with the general map, it is in the AI agents guide. Maintenance has a price too, and what moves that monthly fee is broken down in the cost guide.