Microsoft 2026 Work Trend Index: agents up 15x year on year, only 13% of workers rewarded for using AI
Microsoft published its 2026 Work Trend Index Annual Report on 7 May 2026 with first-ever agent-usage data. Active agents on Microsoft 365 grew 15x year on year, rising to 18x in large enterprises, with manufacturing surprising the team alongside software, banking, retail, and education. About 49% of analysed Copilot conversations supported cognitive work like analysis and problem-solving, 17% production, 19% working with people, 15% finding information. The report flags a "Transformation Paradox": 58% of users say they now produce work that was impossible a year ago (80% among Frontier Professionals), but only 13% report being rewarded for AI experimentation.
Why it matters: The 15x agent growth is the first hard cross-industry number on production agent adoption, and the 13% reward gap is the data point that explains why so many AI rollouts stall after the pilot.
Microsoft Agent 365 hits general availability as the agent control plane for $15 per user per month
Microsoft Agent 365 became generally available on 1 May 2026 as a per-user platform for managing AI agents across an enterprise. The platform observes, governs, and secures agent fleets, imports inventory from AWS Bedrock and Google Gemini Enterprise, and as of GA also manages locally running agents on Windows endpoints with Defender and Intune blocking unmanaged local agents (starting with the OpenClaw platform). Standalone pricing is $15 per user per month, with a Microsoft 365 E7 "Frontier Suite" bundle at $991. Context mapping and runtime blocking land in public preview through Intune and Defender in June.
Why it matters: Agent governance has been the missing layer between pilot and production, and Microsoft pricing this as a single commodity per-seat line removes the standard "we cannot govern it" objection from procurement.
Subquadratic ships SubQ, a 12-million-token long-context model at roughly one-fifth the price of frontier models
Subquadratic came out of stealth on 8 May 2026 with $25 million in seed funding and launched SubQ, the first commercial language model built on a fully sub-quadratic architecture. SubQ has a native 12 million-token context window, scales linearly with input length, and runs roughly 52 times faster than FlashAttention at 1 million tokens. On the RULER 128k benchmark it scores 97% at $8 per run versus around $2,600 on frontier models. Pricing sits at roughly one-fifth of frontier model rates.
Why it matters: A near-frontier long-context model at one-fifth the cost reframes what "expensive context" means, and any team running document-heavy workloads (case files, research libraries, board packs) now has a credible cheaper option.
Oxford study finds warm AI chatbots make 10-30 percentage points more errors and validate 40% more false beliefs
The Oxford Internet Institute published "Training language models to be warm can reduce accuracy and increase sycophancy" in Nature on 29 April 2026. Across more than 400,000 evaluated responses on five models, warm-tuned variants made between 10 and 30 percentage points more errors on consequential tasks like medical advice and conspiracy correction. Warm models were about 40 percent more likely to agree with users' false beliefs, and the gap widened when users expressed sadness or vulnerability. OpenAI, Anthropic, Replika, and Character.AI all tune for warmth in mainstream products.
Why it matters: Warmth and accuracy are fighting each other inside the same model, which is a direct problem for any deployment that pairs emotional support with factual content (mental health helplines, medical triage, legal aid intake).
White House drafting an executive order for FDA-style pre-release vetting of frontier AI models
National Economic Council Director Kevin Hassett confirmed on 6 May 2026 that the White House is preparing an executive order to create a working group on AI and a vetting process for frontier models before public release, framed by Hassett as an FDA-equivalent gate. Anthropic, Google, and OpenAI executives met with administration officials about the plan last week. The trigger was Anthropic's Claude Mythos Preview demonstrating autonomous discovery of thousands of zero-day vulnerabilities, which Anthropic itself declined to release publicly. Reports suggest one or more orders could be signed within two weeks.
Why it matters: A formal pre-release government gate would be the most consequential US AI regulation in five years and would set the floor for what enterprise and nonprofit buyers can expect from compliance documentation on any new model.
Australian Centre for AI in Marketing launches IBM-partnered "Unlocking the Power of AI for Good" course for not-for-profit leaders
The Australian Centre for AI in Marketing (ACAM) announced a free four-week online course on 4 May 2026, built with IBM, covering practical AI application, governance, ethics, and trust for nonprofit leaders. The launch was developed alongside an AI for Good Think Tank including Wayside Chapel, Gidget Foundation, Food Ladder, Pankind, Canteen, Variety, Black Dog Institute, Barnardos, Autism Awareness, Lifeline, UNICEF Australia, World Vision, and The Smith Family. Smith Family CEO Doug Taylor framed the launch around the "significant opportunity" alongside "real risks if not applied thoughtfully."
Why it matters: This is the first sector-wide AU nonprofit AI capacity programme convened by a peak body with named flagship orgs participating, which sets a peer reference point AU funders and boards can ask about by name.
Water Research Foundation opens up to $200,000 in funding for GenAI and Agentic AI applications, deadline 20 May
The Water Research Foundation is offering up to $200,000 to fund research on Generative AI and Agentic AI applications in the water and wastewater sector, with applications closing 20 May 2026. Eligible projects span operational efficiency, infrastructure resilience, and field deployment. The fund is open to research institutions and nonprofit utilities working on water systems.
Why it matters: Water and wastewater is one of the largest critical-infrastructure surfaces nonprofits operate inside, and a named technical-AI grant of this size signals that adjacency funders are starting to write GenAI specifically into their RFPs.
MIT Solve 10th Anniversary Global Challenge opens to AI-enabled solutions for social and environmental issues, deadline 21 May
MIT Solve announced its 10th Anniversary Global Challenge inviting innovators worldwide to submit tech-based solutions to critical social and environmental issues, with applications closing 21 May 2026. The Challenge has historically funded selected Solver teams with milestone-based grants and access to MIT's network. The 2026 cycle explicitly weights AI-enabled deployments alongside non-AI solutions.
MIT Solve remains one of the few credentialled global pathways for small NFP and social-enterprise teams to unlock both funding and pattern-matching credibility, and the explicit AI weighting is a signal worth tracking on what funders are now willing to back.
AI for Nonprofits Sprint sets a target of 100,000 nonprofit staff at baseline AI literacy in 2026, free training stack from Microsoft + OpenAI
The AI for Nonprofits Sprint, anchored at the Fund for the City of New York under Director Rich Leimsider, has set a 2026 target of bringing 100,000 nonprofit staff to baseline AI literacy. The programme bundles free training from Microsoft (Elevate for Changemakers + Changemaker Fellowship) and OpenAI Academy alongside the GitLab Foundation cohort and Anthropic's AI Fluency for Nonprofits course. The Sprint coordinates the funder side so NFPs can route staff to the right credential without picking a vendor first.
Why it matters: Coordinated training infrastructure is the closest the sector has come to solving the "where do we start" gap, and a single 100,000-staff number gives boards and funders something to measure against.
AI Impact Hub