---
id: "lib-2026-10-15-enterprise-ai-has-barely-started"
type: "library"
title: "Enterprise AI has barely started"
title_original: "AI在大企业中的应用才刚刚开始，人才供不应求？｜对话微软CoreAI副总裁Marco Casalaina"
lang: "ZH"
org: "课代表立正"
source_type: "Talk"
date: "2026-10-15"
week: "2026-W42"
url: "https://nosov.ai/library/lib-2026-10-15-enterprise-ai-has-barely-started/"
original_url: "https://www.youtube.com/watch?v=KpQwWNkyEsQ"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "Interview with Microsoft CoreAI VP Marco Casalaina on a Chinese AI channel; concrete practice on context capture, rubric evaluation and agent identity, but a vendor executive's account with customer examples not independently checked."
summary_basis: "transcript"
original_minutes: "43"
tags: ["business transformation", "data dictionary", "evals", "agent identity"]
frames: ["verification-capacity"]
---

# Enterprise AI has barely started

Original title: AI在大企业中的应用才刚刚开始，人才供不应求？｜对话微软CoreAI副总裁Marco Casalaina

Tier B (Useful, limited): Interview with Microsoft CoreAI VP Marco Casalaina on a Chinese AI channel; concrete practice on context capture, rubric evaluation and agent identity, but a vendor executive's account with customer examples not independently checked.

Summary based on the full transcript.

## In short

A 43-minute conversation (15 October 2026) on 课代表立正, a Chinese channel run by Sun Yuzheng of Superlinear Academy, with Marco Casalaina, VP Products for Core AI at Microsoft. The talk is in English with Chinese captions and was published ahead of AI Engineer Shanghai. His three practical points: models need business context that has to be collected from people, agents should get their own identity with least privilege, and evaluation is moving to weighted rubrics that encode business rules.

## Key numbers

- six months: time a BMW team spent touring departments to build a data dictionary before engineers could query test-car data
- 5,000: sensors on BMW test cars, logging a terabyte or more of data per drive
- 11,500: models listed in the Microsoft Foundry catalogue
- three months: period over which agent evaluation shifted from generic metrics to rubric-based evaluation, per Casalaina

## What it reports

- Consumer agents act as the user and inherit all the user's permissions; Microsoft's "Autopilots" get their own Teams account, mailbox, files and security context and report to a human manager.
- BMW's Mobile Data Recorder: test cars carry about 5,000 sensors and log a terabyte or more per drive. Until about a year ago only a few specialist "wizards" could query the data and engineers queued.
- A BMW team spent about six months gathering a data dictionary across departments (what each sensor means, thresholds, local terms). After that any engineer could ask for, say, all hard-braking events in wet weather last week.
- Evaluation: generic metrics such as task completion can pass while the answer is wrong (an apartment-pricing agent with badly off three-bedroom prices). Rubrics state business rules in natural language, each weighted, with an error threshold agreed before production. Users must say why an output is wrong, and the reason becomes a new rubric dimension.
- Microsoft has run rubric evaluation on M365 Copilot for over a year; he names "eval maintainer" as a new job category.
- Advice: start with the rubric (banks already have one in compliance documentation), run an "abnormally long" pilot, and give every engineer a baseline in context provision and rubric writing.

## How they got it

Interview. Examples come from Microsoft's own products and customer work; nothing is shown in data.

## Limits

- Vendor executive; customer examples (BMW) are Microsoft's account and were not checked.
- The six-month BMW figure is an informal estimate in conversation.
- The episode carries a promotional code for a conference.

## Why it matters for companies

The BMW case puts a time cost on capturing tacit knowledge: half a year of talking to colleagues before an agent could answer basic questions. Rubric evaluation gives companies a way to reuse documents they already hold, such as policies and compliance rules, as success criteria for agents. An agent with its own account and a named manager is also an access-governance and accountability question for HR and security, not only for IT.

## Key passages in translation

The conversation is in English; the channel's title is in Chinese.

> AI在大企业中的应用才刚刚开始，人才供不应求？

AI adoption in large enterprises has only just begun: is talent in short supply? (episode title)

> They spent like six months going all around BMW and gathering a data dictionary.

(original English)

> Meanwhile, there's an entirely new job category that has been created over the past couple of years, which is eval maintainer.

(original English)

Original: https://www.youtube.com/watch?v=KpQwWNkyEsQ