Applied AI Weekly Brief
Test accepted work and independent learning before scaling AI services
· Edition 1 · Source review recorded November 6, 2026
Download this Applied AI brief (PDF) ↓This week's judgment
For SMB clients, the next agent pilot should prove that a defined task gets done at an acceptable cost. For conversational tutoring, pair review of learner questions with independent checks of understanding. Both services need a clear test of whether the help actually worked.
2026: The State of Enterprise AI
What changed
Menlo surveyed nearly 400 U.S. enterprise decision-makers. Median AI spending reached 2.2 times the previous year's level against a planned 1.8 times. The report also found limited evidence that enterprises are replacing core CRM or HR systems with AI.
Why it matters
For an SMB, improve a specific workflow around its existing systems. Before quoting a recurring service, estimate cost per accepted task, including agent retries, external tools, and staff review. Set a spending cap during the pilot.
Evidence limits
Enterprise survey results are not an SMB benchmark. Market totals are modeled annualized estimates, not audited revenue. Menlo has investments in companies discussed; spending growth does not demonstrate customer return.
Read in full?
Yes, selectively. Read the production-agent and methodology sections for workflow selection and budgeting; treat market sizing as context.
Original source: Tim Tully, Joff Redfern, Deedy Das and Derek Xiao / Menlo Ventures ↗Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
What changed
After an August-October review of internal data and staff interviews, METR judged that Anthropic's progress on well-specified tasks was likely running at roughly twice its earlier rate, with high uncertainty. Staff described the acceleration as primarily AI assistance, with fully autonomous contributions playing a smaller role.
Why it matters
Test assisted workflows before promising an autonomous service. In a client pilot, compare completed, accepted work with the previous process and record human review time. Specify which decisions still require an experienced practitioner.
Evidence limits
This is one frontier lab, not evidence of a twofold SMB productivity gain. Attribution is uncertain, and Anthropic redacted much of the most informative evidence, limiting independent scrutiny of METR's judgment.
Read in full?
Yes for delivery leads. Read the summary, empirical findings, and limitations to distinguish measured progress from subjective interpretation.
Original source: METR ↗What Students Ask an AI Tutor May Signal How They’re Learning
What changed
Kaplan analyzed 6,060 learner-course records in tax and accountancy. Its heaviest TutorBot users scored lower in five of six subjects. Around the pass boundary, stronger performers asked more conceptual questions; weaker performers relied more on broad explanations and direct answers.
Why it matters
For an Academy tutor, review recurring question patterns alongside assessment results. Repeated requests for a complete explanation could trigger a diagnostic question or human support. Validate any flag with a reviewer before acting on it.
Evidence limits
Records are learner-course combinations, not 6,060 distinct people. This observational provider account cannot establish that AI caused lower scores; struggling learners may seek more help. It does not validate a predictive alert system.
Read in full?
Yes. A short, directly relevant account of how conversation review could improve support for professional learners.
Original source: James Mansbridge / Kaplan ↗Student-Initiated Questioning in LLM-based Physics Tutoring: Metacognitive Regulation and Dialogic Uptake
What changed
In a study of 18 Korean undergraduates, researchers analyzed 112 spontaneous questions during thermodynamics tutoring. Questions that checked ongoing understanding were associated with richer dialogue. The analysis concerns selected student-initiated exchanges, not engagement across every interaction.
Why it matters
Let learners interrupt the tutor to question assumptions, check a step, or challenge an answer. Include occasional invitations to ask questions, then test understanding on a fresh problem without assistance.
Evidence limits
This small, text-only study had no comparison condition or demonstrated long-term learning gain. Only 64.5% of generated problems matched expert judgments of appropriate difficulty, reinforcing the need to review teaching content.
Read in full?
Yes for tutor designers. Read the dialogue examples, methods, and limitations for concrete interaction patterns and content-quality checks.
Original source: Eunhye Kim and Hunkoog Jho / Journal of Science Education and Technology ↗Next product test
For one Academy lesson, have a reviewer classify requests for targeted help versus complete solutions. Compare those patterns with performance on a fresh, unaided problem. Use the result to refine prompts before introducing automatic learner-risk alerts.
Publication dates were verified on publisher pages; Menlo's date is listed in its reports index. Implications and proposed tests are Audeamus analysis. Publicly indexed Scribd searches yielded no additional qualifying source; subscription-only material was not reviewed.