The week in AI for CEOs: three things that matter, and one workflow to run.

Members: no offer this week. You already have everything DeskTheory makes. Straight to the news.

Microsoft is spending $2.5 billion to fix AI's real problem: adoption

On July 2, Microsoft launched Frontier Company, a $2.5 billion unit with 6,000 people whose whole job is to sit inside customer companies and get AI working there. Early partners include the London Stock Exchange Group, Unilever, and Land O'Lakes. Two days earlier, Amazon put $1 billion into a deployment unit of its own. (TechCrunch)

What it means for you: The companies selling AI just built multibillion-dollar consulting arms, because the models alone were not producing results for customers. That is Microsoft telling you the hard part is not the technology; it is getting a company to change how it works. You do not need 6,000 engineers to copy the approach at your size: pick one team and one workflow, put one person in charge of making it stick, and compare the results to what the team did before. That costs almost nothing.

OpenAI's newest models go public today, and AI just got cheaper again

OpenAI opens all three of its new GPT-5.6 models to the public today, July 9, after a US government security review that had limited them to a small group of vetted partners. There are three tiers: Sol, the flagship; Terra, the everyday mid-tier; and Luna, a budget model that runs at about a fifth of Sol's price. (Engadget)

What it means for you: Ignore the model names. The thing worth remembering: today's AI is the worst and the most expensive it will ever be, and it gets better every few weeks. So don't judge AI by an old experience. If something wasn't ready the last time you tried it, try it again.

Claude's agent is mostly doing office work, not coding

On July 7, Anthropic brought Claude Cowork, its AI agent for everyday work, to the web and your phone: start a job at your desk, close the laptop, and pick up the finished work later. The real news is what people use it for. Anthropic analyzed 1.2 million sessions across 600,000-plus organizations and found a third of the work is routine business operations (reports, checklists, spreadsheet clean-up), about 16% is writing and content, and less than 9% is actual coding. (TechCrunch)

What it means for you: If you thought AI agents were just for engineers, Anthropic's own numbers say otherwise: most of the work is the same ops your office runs on, and coding is the smallest slice. The unlock is delegation, not chat. Hand the agent a whole task, let it run while you do something else, and review the result when it's done. Which is exactly what this week's workflow does to your inbox.