The takeawayTest one useful workflow. Make permissions explicit. Measure the work after human review.
AI tools & LLMsNext move: Test
GPT-6 Astra raises the bar for computer use.
What changed
OpenAI reports 72.6% on OSWorld 2.0 for GPT-6 Astra versus 65.7% for GPT-5.6 Sol, with roughly 47% less time per task in latency simulations.
Why it matters
Our read: retest work that spans research, spreadsheets, and browser actions. Judge the finished deliverable and correction time, rather than assuming a benchmark gain transfers to your workflow.
Meta Muse puts delegation inside everyday messaging.
What changed
Meta introduced Muse on September 8: a personal agent powered by Muse Spark, operating in a dedicated cloud VM through the Muse app or WhatsApp. US rollout includes iOS, Android, and web.
Why it matters
Our read: delegation is becoming accessible beyond technical teams. A familiar messaging interface may lower the learning barrier, while connected accounts make permission choices more consequential.
The UK NCSC’s September 7 guidance warns that unapproved AI can expose company data and extend attackers’ reach through an agent’s permissions. It calls for secure alternatives and open communication with staff.
Why it matters
Our read: an AI policy needs an adoption path. When approved tools cannot meet a real need, employees have a reason to work around them. Learn which jobs are driving that behavior.
In the New York Fed’s August regional survey, 61% of service firms and 51% of manufacturers used AI, versus 40% and 26% in 2025. Findings were published September 1.
Why it matters
Among adopting firms, the median worker-use share was 17% in services and 7% in manufacturing. Our read: access is only the start; help more people apply AI to a useful recurring task.
Proposed pilot: two colleagues compare the current workflow with one approved AI tool on five similar, low-risk tasks. Keep inputs and acceptance criteria consistent.
Owner
Team lead assigns a pilot owner and reviewer.
Success measure
Compare total time including corrections; require equal or better reviewed quality. Record interventions and tool cost.
Review
September 18: bring the five outputs and decide whether to expand, revise, or stop.
Stop rule
Pause for unapproved data access, an unexpected external action, or repeated quality failures.