Uber gives 30,000 people a general AI agent. Microsoft says narrow agents are safer. Both are right
Uber began rolling out ChatGPT Work company-wide as Microsoft reported 2,000 responsible AI reviews and argued that narrow scope is a safety feature. The difference is the layer being scaled.
Listen to this article · 5 min
AI-generated narration of the full article.
Two stories published on October 22 seem to give opposite advice.
Uber announced the company-wide rollout of ChatGPT Work, OpenAI’s agent for multi-step knowledge work inside ChatGPT Enterprise, to more than 30,000 employees, after a pilot with about 1,000. The agent connects to Google Drive, Gmail, Google Calendar and Slack, prepares briefs, builds presentations, analyzes data and completes workflows across those apps. GPT-6.1 Sol is the default model. Uber backs the rollout with an internal library of AI skills, approved integrations and employee training.
The same day, Microsoft’s internal IT team published how it governs AI across the company. Its responsible AI reviewers have completed more than 2,000 internal-tool reviews, with thousands more in the pipeline. And its most quoted lesson points the other way: “Sometimes scaling AI means intentionally limiting what an AI system can do.” The vice president behind more than 100 production agents in Microsoft’s cloud supply chain put it plainly: a planning agent “shouldn’t be able to answer a question about the weather.”
One company is handing a general agent to everyone. The other is narrowing agents to one job each. They are describing two different layers.
Two layers of enterprise AI
How wide an AI system's scope is, and what governs it
One job, one processAny task, any app the user can reach
- Microsoft's supply chain agentsScoped to a business purpose, red-teamed for out-of-scope behavior
- Microsoft's legal discovery agentCurated document set, reproducible reasoning
- Uber's skills libraryShared, approved procedures on a general agent
- ChatGPT Work for 30,000 employeesGeneral agent, user's own permissions and approved connectors
The general assistant layer works for one person, with that person’s access, on whatever that person is doing. Its scope is wide by design. It is governed the way employee tools are: identity, approved connectors, data loss controls, training, and a clear statement that the person remains responsible for the result. Uber’s choices, a central skills library, a list of approved integrations and training, are exactly those controls.
The business agent layer does a defined job inside a process: plans supply, reviews documents for litigation, answers customers. Its scope should be narrow, because its outputs feed decisions other people rely on. Microsoft’s legal team required that every AI conclusion come with “a reproducible, evidence-backed explanation” and restricted the agent to a curated document set. That is how a production system is governed.
The mistake is to govern one layer with the other’s rules. Reviewing every employee prompt like a production system is impossible at 30,000 people. Letting a general agent quietly become the system that runs a business process skips the review that process needs.
What Uber has not said yet
Uber’s announcement is about access. It does not report adoption, task completion or business outcomes, and the companies say they will share those later. That is normal at launch, and it means the hard part has just started. Access is a license count; value depends on how many people use the agent for real work and whether that work is any better. We saw this in the European data, where agent intensity varied nearly fourfold between countries with the same access.
The skills library is the piece to watch. It is where a general agent starts to acquire business procedures, and where it starts drifting toward the second layer. A skill that drafts a meeting brief is a personal productivity tool. A skill that prepares a supplier payment is a business process, and it should be reviewed like one, ideally as a controlled document with an owner.
What Microsoft’s 2,000 reviews teach
Two thousand reviews with thousands waiting is the real number in Microsoft’s story. It shows that review capacity, not policy, becomes the constraint. Microsoft’s answers are reusable: standardized assessments, patterns learned from previous reviews to decide which cases need deeper attention, and requirements built in at design time so teams arrive at review with scope, data limits and human checkpoints already defined. It echoes the risk-tiered approach Microsoft’s IT described earlier this month.
A review count is not proof of safety, and Microsoft’s account is an internal case study, not an audit. But the operating model travels: if your governance team reviews every AI use case from scratch, the queue will decide your adoption speed.
A rule of thumb
Govern the general assistant like a powerful employee tool, at scale and by default. Govern business agents like production systems, one by one, with narrow scope as the default. And watch the line between them: when a personal skill starts running a shared process, it has changed layers, and its review should change with it.