Your Word doc is now an agent
Agentic AI inside Word and Excel went from feature flag to default this week, which changes who has it from the early adopters to everyone with a 365 license. Plus four shorter reads on Big Tech earnings, GM and Gemini, the largest seed in European history, and an open model crossing a frontier coding threshold.
AI Ready.
A newsletter by Seeko. Seeko is a leading AI transformation partner for startups.
THE AI BRIEF
Today's signal: agentic AI inside Word and Excel went from feature flag to default this week, which changes who has it from the early adopters to everyone with a 365 license. Plus four shorter reads on Big Tech earnings, GM and Gemini, the largest seed in European history, and an open model crossing a frontier coding threshold.
THE READ
Microsoft made agentic capabilities generally available in Word, Excel, and PowerPoint and put GPT-5.5 Thinking behind them. The "AI inside the office stack" question moves from rollout to adoption.
On April 22, Microsoft announced general availability of agentic capabilities in Word, Excel, and PowerPoint for Microsoft 365 Copilot, Microsoft 365 Premium, and Microsoft 365 Personal and Family customers. The blog post from Office Product Group president Sumit Chauhan describes Copilot taking "multi-step, app-native actions" inside documents, worksheets, and presentations rather than only answering questions in a sidebar. On April 28, Microsoft layered GPT-5.5 Thinking and ChatGPT Images 2.0 into the same Copilot surface, and the agent features are being set as the default in the affected SKUs.
The shift I'd flag for operators is not the model. It's the configuration. Copilot has been able to draft a Word document and write Excel formulas for a year. What changed this week is that the agent mode is on by default for any tenant on the qualifying plans, including Microsoft 365 Personal and Family. The decision to use it is no longer in the hands of a workplace AI lead piloting features. It's in the hands of every employee who opens a document.
For most operators, the practical question this week is governance, not capability. Copilot can now execute multi-step actions inside documents that touch organizational data through Microsoft Graph. The audit trail, the labeling, the data-loss-prevention rules, and the approval flow for what Copilot is allowed to do without a human review are all configurable, and most tenants have not configured them with agent-default behavior in mind. The default settings were written for a sidebar assistant.
A reasonable move this week is to ask whoever owns the Microsoft 365 admin console for two things. One: which Copilot agent capabilities are enabled by default in your tenant, and on which SKUs. Two: what the company's policy is on Copilot acting on documents containing customer data, financial data, or material non-public information. If those answers are not on hand, the tenant is now defaulting to agent behavior the policy hasn't been written for.
Independent of what shows up in the actual rollout calendar, the design choice Microsoft just shipped is what to track. "Generally available" is no longer a feature toggle. It's a posture change. Operators who treat the office stack as a tool will find their employees increasingly treating it as a colleague.
FREE WORKSHOP
I'm running a free 30-minute session on how you can build a real workflow inside ChatGPT Workspaces later today. RSVP here:
ALSO WORTH KNOWING
Big Tech reported earnings simultaneously last night, and every company beat estimates
Alphabet's Google Cloud grew 63%, Azure grew 40%, and AWS hit its fastest growth in 15 quarters at 28%. Meta beat on revenue but raised capex guidance by $10B, sending shares down 7% despite the beat. Combined 2026 AI infrastructure spend across the four has been revised up to roughly $725B. The buildout is not slowing.
GM is rolling Google Gemini out to about 4 million vehicles via OTA update
The April 28 announcement covers model-year 2022 and newer Cadillac, Chevrolet, Buick, and GMC vehicles with Google built-in, replacing Google Assistant. It's the first time a major automaker has standardized on a frontier LLM for its in-car assistant.
Ineffable Intelligence raised $1.1B at a $5.1B post-money, the largest seed in European history
Sequoia and Lightspeed co-led, with Google and NVIDIA participating. The London lab is run by former DeepMind VP David Silver on the thesis that the next generation of AI has to learn from experience, not human-authored text.
Open-weight models crossed 60% on SWE-bench Verified
Ant Group's inclusionAI team released Ling-2.6-Flash, a 104B mixture-of-experts model with 7.4B active parameters and an MIT license, scoring 61.2%. A year ago, that score required a closed frontier model.
The Fed held rates, and three regional presidents dissented against dovish guidance
The April 29 FOMC decision kept the target range unchanged. Higher-for-longer rates raise the cost of capital underwriting the AI infrastructure buildout, which is what the earnings calls last night were testing against.
WATCHING TOMORROW
Apple's Q2 earnings call tonight closes out the Big Tech reporting block. Tomorrow, we’ll mention whether Apple finally puts a number on AI infrastructure spend, or stays the only one of the four without a clean capex line. Also, April's jobs report drops Friday at 8:30am ET, two days after the Fed held with three dissents. Labor data is the next variable in the rate path underwriting the buildout.Back tomorrow,
Haroon