Most teams I meet have an AI demo that worked. A notebook that ran the once. A PowerPoint that played well in front of the executive team. A vendor pitch that implied "production-ready" without specifying for whom, under what load, and at what unit economics.
That demo is not your system. The gap between them is mostly four things, and the model is not one of them.
The four gaps
Data. The demo ran on a snapshot. Your system needs an owned, observed, permissioned, replayable pipeline. If your data lineage diagram does not exist, your AI is one schema-drift away from a production incident your team cannot debug.
Governance. The demo had no review trail. Your system has compliance, risk, legal, and - in regulated industries - an auditor who reads PDFs. Build the artifact trail into the inference path, not bolted on top.
Evaluations. The demo passed a vibe check. Your system needs a golden set, a rubric, a regression suite, and a cost-aware CI gate that blocks merges when the eval drops below a defensible threshold.
Cost and latency. The demo was free at one request per minute. Your system will see ten requests per second on day one, and the per-tenant cost model has to survive that. Latency is part of the product surface - not an SRE concern you defer.
Why the model is not the problem
The model gets all the attention because it is the visible piece. But when an enterprise AI initiative stalls - and most of them do - the failure mode is almost never "we picked the wrong base model." It is "we shipped without the four foundations above and the team cannot defend the system to anyone outside the room."
You can swap the model. You cannot swap a missing audit trail in a regulated deployment six months after the fact.
What I would do this week
If you have a demo that is being asked to become a system, run four short exercises before you write a single line of production code:
- Diagram the data path end to end. Mark every place a permission is checked, every place data leaves a tenant boundary, and every place a retention policy applies. The places without marks are the work.
- Define an evaluation rubric per task you care about. Pick five rubrics. Generate twenty examples per rubric. Score them by hand once. That is your golden set. Everything else builds on it.
- Write a one-page risk register. For each risk: blast radius, probability, mitigation owner, mitigation status. If "owner" is "we'll figure it out," the risk is live.
- Cost-model a realistic peak day. Per request, per tenant, per incident. Include the cost of being wrong.
The model selection happens after these. By the time you pick it, the rest of the system is robust enough to swap models without panic.
The lesson
Demos are cheap. Systems are not. The work between them is most of the engineering. If your team has not started that work, you do not have an AI initiative - you have an unfinished prototype with executive attention.