Every dev shop added "generative AI" to its homepage in the last two years. Very few of them have shipped an AI feature that survives real production traffic. Here's what's worth asking before you pick one.
Start with the problem, not the model
A vendor who leads with "we build on GPT-4, Claude, and Gemini" before asking what problem you're solving is telling you something. The model is a detail; the right vendor starts by scoping the actual workflow — a copilot, an agent, an automation — and picks the model to fit it, sometimes using more than one.
Questions to ask about data and privacy
- Where does our data go once it hits the model provider, and is it used for training?
- How is sensitive or customer data isolated from prompts and logs?
- What happens if we need to switch model providers later — is our system locked to one vendor's API?
Questions to ask about production readiness
- How do you evaluate output quality before and after launch, not just during a demo?
- What's the plan for hallucination or incorrect outputs reaching a real user?
- Who owns monitoring and cost control once the feature is live and usage grows?
Red flags to watch for
A polished demo with no discussion of evaluation, guardrails, or cost at scale. A team that can't explain how the system behaves when the model gets something wrong. Pricing based purely on hours with no accountability for whether the feature actually works once real users touch it.
What working with Oplyx looks like
We scope the workflow first, choose models per use case rather than defaulting to one vendor, and build in evaluation and guardrails from the start — because a generative AI feature that only works in the demo isn't a shipped feature. If you're evaluating vendors right now, that's the bar to hold everyone to, including us.