Something I keep seeing play out the same way across teams.
The model works great in testing. Accuracy looks good, latency is acceptable, and everyone's excited. Then it hits production, and everything slows to a crawl, not because the model broke but because the organization around it wasn't ready.
The real blockers are rarely technical in the way people expect. It's stuff like the data pipeline that feeds the model, pulling from three systems that format timestamps differently. Or the security team needing four weeks to approve an API endpoint nobody told them about. Or two teams, both thinking the other one owns monitoring for the inference service. Or a compliance review that nobody scoped into the timeline because "it's just an internal tool."
At enterprise scale, the implementation discipline ends up mattering as much as model quality. The teams I've seen actually get AI into production and keep it running treat the rollout more like an infrastructure project than an experiment. They map out the data dependencies, approval chains, ownership boundaries, and security constraints before writing the first line of integration code. Not because they love the process, but because skipping it means discovering it all in week six, when everything is on fire.
The pattern that works: treat AI deployment as organizational engineering, not just model engineering. The model is maybe 30% of the problem. The other 70% is making it work inside a real environment with real systems, real people, and real constraints.
What's been the biggest non-technical blocker you've hit trying to get something AI-related into production?
Top comments (0)