November 12, 2026
DePaul moved the administrative work, not the relationship | Education and Higher Education
DePaul University’s financial-aid counselors used to spend about half of a 30-minute appointment assembling a student’s situation from four separate data sources. A staff-facing AI agent now produces the overview and action plan in about one minute, roughly doubling the time available for the student conversation. EDUCAUSE reports more than 100 hours freed and a 90% average reduction in processing time since the November 2025 launch, while counselors retain responsibility for advice and student support. Source
Executive snapshot
DePaul moved the administrative work, not the relationship. Its agent compresses a multi-system financial-aid review from roughly 15 minutes to one, giving counselors more time with students.
Google created a lower-cost decision lane for agents. Gemini Decision 1 evaluates bounded choices across multimodal inputs, returning confidence scores without generating prose.
Salesforce published encouraging sales-agent results with an important caveat. The survey includes only 24 active customers and relies on self-reported outcomes.
Microsoft showed why model provenance belongs in agent security. Its researchers inserted a hidden trigger by editing fewer than ten dormant neurons, without retraining the model.
Top developments
The generative model gets a decision lane
Google introduced the Gemini Decisions API, powered by Gemini Decision 1. It can evaluate as many as 1,000 yes/no, multiple-choice, or rubric-scoring questions against a shared text, image, audio, video, or PDF input. Google reports 150–250 millisecond end-to-end latency, a price of $0.10 per million input tokens, and no output-token charge. Every result includes a confidence score, which makes the API useful for routing, qualification, policy checks, context grading, and escalation. Google’s benchmark and cost claims are provider-reported, and Google Cloud support is still listed as coming soon. Google
A clean evaluation can still hide a loaded model
Microsoft Research demonstrated “Lazarus,” a technique that changes fewer than ten dormant neurons so an agent behaves normally until a hidden trigger appears. The modification requires no model retraining and can cause the agent to emit commands using whatever permissions it has. In one reported experiment, seven edited neurons triggered a payload on at least 98.8% of held-out trigger trajectories and never on benign ones. Microsoft says it has not observed threat actors using the technique. The operational lesson is concrete: source models from declared publishers, verify hashes, compare weights with the declared base, isolate tool execution, and keep high-impact actions behind approval. Microsoft Research and Microsoft Security
AlphaGenome moves from tissue averages toward individual cell states
Google DeepMind released AlphaGenome 2 for non-commercial research and placed commercial access in private preview. The model was trained on more than 107 million individual cell profiles across 68 tissues and predicts how DNA variants affect specific cell types and states. DeepMind says it identifies nearly twice as many variants that alter gene expression as the prior model. This is a research release, not a clinical product: Google explicitly says it has not been validated or approved for clinical or diagnostic use. Google DeepMind
AI for Sales and field execution
Six hours back, with a small denominator
Salesforce surveyed 24 active users of Sales Cloud’s AI sales agents. Respondents reported an average of six hours saved per seller each week; 63% reported more accurate and complete CRM data; and two-thirds of account executives said deals moved 10% to 20% faster. Prospecting, account research, and pipeline administration were each deployed by 71% of respondents.
These are useful directional signals, not broad market evidence. The sample is small, customers were already active users, outcomes were self-reported, and Salesforce did not publish independently verified CRM-quality or conversion data. A practical field test should measure accepted updates, corrections, missing context, lead coverage, follow-up latency, stage movement, and customer response separately. Salesforce
This is where an end-to-end system matters: YouEx.ai Salesforce Package can carry an inbound conversation into lead research, scoring, routing, follow-up, and scheduling, so the team can measure whether better context actually changes the next action instead of stopping at time saved. And it has MCP included out of the box. YouEx.ai
The field-service pattern is evidence, recommendation, boundary
KT’s AIO-Net agents analyze network anomalies and produce incident reports for human operators. KT says issue detection fell from more than ten minutes to two or three, while investigation and reporting now average five minutes, about half the prior time. The design is as important as the speed: the system may inspect equipment and recommend a response, but it can use only allow-listed, read-only commands. That boundary is enforced in the application, not left to a prompt. The deployment currently covers 70 units of critical equipment and is planned to expand to 280. Microsoft and KT
Education and Higher Education
DePaul starts with the counselor’s first fifteen minutes
DePaul’s team interviewed every financial-aid counselor, documented more than 60 actions, and spent four weeks in discovery before an eight-week build and test cycle. The agent combines institutional knowledge, federal and state guidance, and student-specific data. It is staff-facing by design, reducing privacy risk and preserving the counselor’s role in a stressful, high-context interaction.
The published evidence is stronger on process than on student outcomes. DePaul reports more than 100 hours freed and 90% average processing-time improvement, but it has not yet published accuracy, repeat-visit reduction, resolution, equity, or satisfaction results. Those are the next measures that will show whether the faster preparation improved the student experience. EDUCAUSE case study
Ohio State joins the data foundation to the academic mission
Ohio State and Microsoft announced a five-year partnership spanning clinical care, research, education, cloud infrastructure, and workforce development. Work already underway includes ambient clinical documentation and smart-hospital infrastructure. Planned elements include moving roughly 70% to 80% of storage and compute to Azure, building a unified Fabric and OneLake data foundation, exploring AI-supported tumor-board workflows, and expanding Microsoft training and certification for students and employees. Most outcome claims are still prospective, so milestones, adoption, reliability, and clinical or educational impact will matter more than the partnership’s breadth. Microsoft and Ohio State
AI adoption becomes curriculum design
Stanford committed $10 million to help departments rethink undergraduate majors for an AI-rich environment, including introductory courses, capstones, and what students must demonstrate independently. The important signal is institutional: AI fluency is moving beyond optional workshops into the structure of degrees and assessment. The initiative is funded and organized, but it has not yet produced redesigned programs or student-outcome evidence. Stanford
What to watch
Decision-layer adoption: whether teams use Gemini Decision 1 for bounded routing and grading, and what calibration looks like on real sales, service, and education data.
Sales-agent measurement: whether Salesforce publishes the full distribution behind its averages, plus correction rates, stage conversion, and retained revenue outcomes.
DePaul’s student outcomes: repeat visits, resolution time, counselor corrections, privacy incidents, and student satisfaction.
Ohio State’s execution: named deployment milestones, adoption, measured workload reduction, and evidence that the shared data foundation improves care, research, or learning.
Model supply-chain controls: whether weight diffing, signed provenance, agent isolation, and post-update checks become standard procurement requirements.