AI Team Culture Over Models: How Full-Stack Crews Ship

In March 2025 I staffed the same RAG boilerplate with two crews for the same client. Crew A had the larger model budget. Crew B had three habits: a shared failure log, Friday demos on real transcripts, and one person who could stop a release with a single failing example. I gave both crews the same document set, the same eval set of 200 fixed questions, and the same four second response budget.

Crew A changed models twice in five weeks. Each swap cost nine days of re-tuning. Their answer quality moved from 71% to 73% correct on the eval set. Crew B kept the cheaper model, met daily for fifteen minutes around the failure log, and shipped to ten dispatchers in week six with 78% correct and a median reply time of 3.1 seconds. I watched the difference live on a shared screen. Model choice added two points. Culture added seven.

Screenshot: Grafana panel 2025-03-28 shows Crew B median latency 3.1s vs Crew A 5.8s, same model, same store. Source table: failure_log_2025_03.csv (214 rows).

Why I track culture first

I interview teams now with one question: show me last week's failure log. Teams that open a table with owner names, transcript links, and source pages handle the next model change in a day. Teams that open private notebooks need a week. The log predicts delivery speed more than the model card does.

Three seats that carry the habit

I hire for ownership, not for titles. Three seats carry the crew when they share the log.

• AI Engineer owns retrieval and release. She connects the model to the document store via RAG, sets cost and latency budgets, keeps the 200 question eval set green, and signs the release that touches customers.
• Prompt Engineer owns instruction reliability. He writes system prompts and few-shot examples, versions them in git, and reruns the regression file of past failures before every release. He reads raw transcripts, not summaries.
• Responsible AI lead owns guardrails. She approves outbound messages, strips phone numbers and IDs before logging, sets 30 day retention, and maps each stored field to its legal basis. She stops a release the day before launch if a single source is stale.

One habit ties them: anyone stops the release. No defense beyond one failing example. The team thanks the person who stops it. That rule teaches the crew that the product belongs to all three.

How we kept the cheap model

Crew B routed simple reads to a small model at $0.02 per hundred calls and saved the large model for damage claim letters where a wrong clause costs money. The AI engineer cached repeat tariff questions, trimmed the document set from 1,400 to 620 pages, and hit the four second budget without touching the model. The prompt engineer rewrote confirmations to name street, dock, and time window after he read 40 transcripts where drivers phoned dispatch despite a bot confirmation. Callbacks fell 49% in two weeks. The responsible lead caught a stale tariff PDF dated March before launch and forced a one day delay that saved forty wrong quotes.

Your next week

Open one table this Monday. Every wrong answer gets transcript, source pages, and an owner. Friday, each owner shows one failure and one fix in fifteen minutes. Keep the model you have. Measure the log for a month. You will see where culture beats the next model.