AI systems fail differently
Classic software either falls over or works. An AI system keeps answering while it is already wrong, and that is an entirely different kind of failure. The model drifts away from reality because the data changes. Costs climb because someone lengthened a prompt. A user talks the system past its instructions.
None of those three takes a server offline, so none of them is caught by standard monitoring. That is why we also monitor model behaviour, cost per request, drift and injection attempts. Otherwise you discover the failure when a customer calls.
Taking over what is already built
More and more often something arrives here that someone else made. Sometimes by an agency that has shut down, sometimes assembled with AI in a few weeks. That can work perfectly well and still not be sustainable: no tests, no documentation, no idea what happens at the first peak.
We take such codebases over. First an audit with an honest verdict. Sometimes the answer is that rebuilding is cheaper than repairing. Then hardening, and then support with an SLA.
Being accountable
We are willing to be judged on the result. That is not a new AI promise: it is how we worked with ANWB, ING, KPN and the Dutch Postcode Lottery for fifteen years. Technology changes, responsibility does not.