Why most "AI companies" are not AI engineering companies
Since 2023, most software vendors have rebranded with "AI" somewhere in their pitch. Few of them are AI engineering companies. The honest test is what their day looks like: a real AI engineering company spends most of its time on data pipelines, evaluation harnesses, retrieval systems, and production observability — not on hand-rolled prompts.
When a vendor cannot tell you, in detail, how they evaluate model output, monitor regressions in production, or handle drift, they are not an AI engineering partner. They are a generalist software shop with an OpenAI API key.
The five questions that separate serious vendors from the rest
1. Show me the AI systems you currently operate in production. (Not demos. Live systems with real users.)
2. How do you evaluate model output, and how often? (Look for golden datasets, automated evals, regression tests, and human review queues.)
3. What is your MLOps stack? (You should hear about feature stores, model registries, monitoring, retraining schedules.)
4. How do you handle a production incident in an AI system? (You should hear about rollbacks, fallback policies, and observability — not just "we retrain".)
5. Will you work inside our cloud account if our data sensitivity requires it? (The answer should be yes, with no friction.)
Team composition that actually works
A working AI delivery team usually contains an AI/ML engineer, a backend engineer, a frontend engineer, a designer, and a part-time product manager. For larger projects add an MLOps engineer and a data engineer. The AI/ML engineer should be senior — five years or more — because most AI work is debugging data and evaluation, not training models.
Avoid vendors that promise to staff AI work with junior engineers and a senior "architect" who shows up on calls. AI work fails in the details. The people building it must understand the details.
Red flags to walk away from
Walks away signals: vendor pitches a model architecture before asking about your data; vendor promises a single-digit-week timeline for a custom model; vendor will not commit to working inside your cloud account; vendor cannot point to a single live production AI system; vendor offers a fixed-price quote before any discovery; vendor refuses to sign an NDA before the first technical conversation.
How Zeven fits this framework
Zeven runs AI engagements as senior teams of two to five engineers, working in two-week sprints with continuous demos. We share live production references on request, work entirely inside the client's cloud account when sensitivity requires it, and stay engaged for MLOps and ongoing operation rather than disappearing after launch. We sign NDAs before the first technical call.