The early-2024 AI team org chart was usually one of two shapes. Shape one: a centralised AI team that owned model selection, prompts, and the platform, with product teams as customers. Shape two: AI capability embedded inside each product team, with no centre. By mid-2026 the teams that ship reliably have converged on a third shape, and the teams that have not converged are mostly the ones still struggling to get into production.

The third shape, in the form it shows up across Australian and global product orgs we work with:

A small AI platform team that owns the inference stack, the evaluation framework, the prompt and config registry, the guardrails, and the cost and observability backbone. This team is typically four to ten engineers and is not customer-facing.

Embedded AI engineers inside each product team. These engineers ship the AI behaviour for their product. They use the platform team’s stack. They write the prompts, design the agent loops, and own the evals for their product’s AI features.

A separate evaluation and quality function that runs across products. This is the team that designs the offline eval sets, the online eval pipelines, the regression tests against model updates, and the red-team programmes. The teams that ship treat eval as a craft and staff it separately from product.

Three observations from the orgs that have settled into this shape:

The platform team’s product is the inference stack and the eval framework, not the AI features. This is the hardest mental shift for a centralised AI team that started in 2023 and was used to owning model output. Letting product teams own the prompts is a discipline.

The product teams’ AI engineers are increasingly senior. The “AI engineer” role of 2024 was junior-coded in the market. By mid-2026 the role is mid-to-senior product engineering with deep familiarity with the inference stack, the eval framework, and the product’s AI failure modes. Hiring for this role at junior level is no longer working.

The eval team is the one most undervalued in 2024 and 2025, and the one with the highest impact-per-headcount in 2026. The orgs that staffed eval seriously are the ones whose production AI is calibrated, observable, and improvable. The orgs that did not are the ones still firefighting in eval-shaped problems.

For Australian companies still in the “should we have a centre” debate in mid-2026: the answer the production-ready cohort has converged on is “yes, but a small one, owning the platform and the eval, not the features.” The features belong with the product teams. The platform belongs in the centre. The eval belongs as a separate craft.

This pattern is not new in technology org design — the platform-plus-product-plus-quality shape is the same shape that mature web and mobile orgs landed on a decade ago. AI is not exotic on the org chart. It just took the market two years to remember the lessons.