AI Field Notes by Michael Nemtsev

The pilot phase ended quietly

Agents moved out of the demo box this week and into the tools people already use to write code, screen candidates, move documents, and run operations.

Agents moved out of the demo box this week and into the tools people already use to write code, screen candidates, move documents, and run operations.

The week had plenty of big numbers.

Cognition's valuation doubled. A London lab with no product drew a $5 billion bet. OpenAI's growth math looked strained. Anthropic flirted with a near-trillion-dollar valuation. Azure, Google Cloud, Amazon Bedrock, KKR, Dell, and the data-center names kept showing that AI infrastructure is now a market in its own right.

But the more important shift was smaller and closer to work.

Microsoft let Copilot touch actual Word, Excel, and PowerPoint files. Cursor said the IDE was the wrong abstraction. Visual Studio put a cloud-agent button next to normal development work. Mistral moved remote coding sessions to the cloud. Codex agents began driving desktops while people worked. Amazon Connect Talent made voice interviews scalable. Sales teams started building their own agent coworkers.

That is not a pilot pattern. That is production seepage.

The tool no longer waits in a separate chat window. It edits the document, opens the ticket, drives the desktop, asks the interview question, checks the code, and returns a diff. The user is still there, but their job moves one step up the stack: write the instruction, review the output, catch the failure, own the consequence.

This is why the security stories mattered so much.

A Cursor agent wiping a startup database, LeRobot shipping with a backdoor, Semantic Kernel turning prompt injection into shell access, agent incidents hitting 88% of organizations, and Common Crawl data fights all pointed at the same production flaw: teams gave agents real tools before they built real blast-radius controls.

An agent that can only draft text is annoying when it is wrong. An agent that can delete a database, move money, leak credentials, or create a hiring record is a different class of system.

The physical world joined the same pattern. Humanoids crossed from demos to factory orders. Robotaxis froze when the cloud dispatch layer failed. Spatial AI and warehouse-picking systems turned motion data into an industry. The robotics story is no longer "a robot walked on stage." It is "which repeated task gets recorded, standardized, and priced into a contract?"

For buyers, the mistake is to treat all of this as tool adoption. It is process redesign.

If Copilot can edit the spreadsheet, the approval step belongs inside the spreadsheet workflow, not in a policy PDF. If an agent can open a pull request while the developer sleeps, code review capacity becomes the bottleneck. If a voice bot runs the first interview, the employer needs a record of what it asked, why, and how a human can intervene.

For sellers, the opportunity is not another agent demo. It is the boring proof layer: permission scopes, audit logs, sandbox defaults, rollback paths, evals on real workflows, and pricing that survives a week of actual usage.

The infrastructure and valuation stories are still relevant. OpenAI's contracts, Google's TPUs, AMD's ramp, Huawei's chip split, and KKR's infrastructure fund all shape who can run these systems cheaply. But the buyer pain is already inside the workflow.

That is the quiet end of the pilot phase. The tools are not asking whether they can help. They are starting to act.

The week in one line: AI adoption is no longer a question of access to agents, but of who reviews their work before it becomes real.

Prefer email?

Get the daily brief and weekly deep dives delivered free.

Read on Substack

Want this in your inbox?

The week in AI, once a week.

A weekly long read on what actually shifted in AI and what it means for the work. Free, unsubscribe anytime.