Building Agents Got Easier. Shipping Quality Didn’t
Writing an agent is easy now. Getting one into production is not, and the gap between those two facts is where most AI programmes currently live. Nearly every team is experimenting. Only a small share of those experiments ever serve real traffic, and the reasons are consistent enough to be worth naming. Mark has been on the wrong side of this repeatedly. He founded Domopult, grew it to over 2 million monthly users, and started applying AI there long before it was fashionable. Later came computer vision deployments across logistics and retail. Different problems, different generations of tooling, and the same three factors decided every time whether the work reached production or quietly died as a pilot:
-
- Time. How fast the team could get to a result they trusted. With AI, this collapsed almost entirely into a data question: the model was only ever as good as what it was trained and grounded on, and the data work was always underestimated.
- Security. Prompt injection, sensitive data leaking into prompts and traces, agents granted tool permissions nobody had reasoned about, and models acting on untrusted input.
- Cost. Which stayed invisible until it became a line item someone in finance had questions about.
The uncomfortable part: each of these is manageable on its own, and teams routinely optimise one at the expense of the other two. The session works through what it actually takes to hold all three at once, what to instrument, what to evaluate continuously, what to measure before the pilot rather than after, and is honest about the trade-offs that remain.
Book Now