Saturday synthesis: The week AI moved deeper into the workflow

Unilever is using five coordinated AI agents to read, interpret, enter, and check nonstandard customer orders, sending missing information and exceptions to people. Across 11 pilot markets, the company says the first stage of order processing fell from 15 to 20 minutes to about 90 seconds, with 95% of orders handled touchlessly. The important lesson is not the number of agents. It is that Unilever started with customer friction, designed a clear exception path, and measured a completed business process. Source: Microsoft customer story

Four patterns from the week

1. The unit of value is becoming the completed workflow

Unilever measured order-cycle time and exception handling. Legora reported reductions in contract-review time. Anthropic’s drug-development work reported lab-tested molecular results and also documented limitations. These examples are more useful than general productivity claims because they connect model behavior to a defined job and an observable outcome. Unilever, Legora, Anthropic life sciences

2. Human review is becoming an architecture decision

The better implementations did not simply add a human somewhere near the end. They specified what uncertainty triggers review, who owns the exception, what context reaches that person, and which actions require confirmation. That pattern appeared in Unilever’s order process, Meta’s credentialed agent, and CAIS’s advisor workflow. Meta Muse, CAIS

3. Shared context is the new operational layer

Useful agents need access to the same facts that shape human judgment: customer history, entitlements, pricing, installed assets, documents, policies, and live operational signals. Google’s connected-app layer, ServiceNow’s OpenAI partnership, Salesforce’s work inside Claude and Slack, and xAI’s shared bots all point toward assistants acting over governed enterprise context rather than isolated chat histories. Salesforce, xAI

4. The model market is becoming a delivery market

Lower prices and faster models matter, but the week’s larger enterprise story was distribution through clouds, marketplaces, workflow platforms, partner networks, and existing interfaces. Buyers will increasingly evaluate where a model can run, which data-processing terms apply, how actions are governed, and how easily the workload can move if requirements change.

Executive snapshot

Read more