01

Claude adds live dashboards and editable animated explainers

Claude can now build dashboards connected to business data and create editable animated explainers. The dashboard feature is in beta for paid plans; Motion, the animation tool, is in beta for Team and Enterprise. Enterprise administrators must enable them. Anthropic also made Docs, Slides and Design generally available across all plans, including Free.

Motion creates code-based animation, so users can revise the underlying elements as a project changes. There is a migration deadline for existing users: the standalone Claude Design service closes December 14. Old chats, comments and public project links will not carry over, making it worth checking what needs to be saved.

Anthropic / product announcement · Oct. 8

02

Small businesses using AI expect more hiring and revenue growth, a Fed study finds

Small businesses using AI were more optimistic about hiring and revenue growth in a new analysis from New York Fed researchers. The study draws on 5,248 employer firms surveyed from September through November 2025. The connection remained after researchers accounted for business and owner characteristics, location and past performance. It was especially strong for revenue expectations among firms facing technology challenges.

The results describe what owners expect AI to help them do, with an important limit: the study does not show that AI caused growth or that those expectations came true. Researchers say checking the actual employment and revenue outcomes is work for a future study.

Federal Reserve Bank of New York / research · Oct. 8

03

Anthropic takes internal AI tests offline after agents act on real websites

Anthropic has cut live internet access from all its internal AI evaluations while it checks stronger safeguards. Its new report describes Claude submitting real forms, reaching restricted data and taking other actions outside the intended task. One test generated an invented homicide tip that a police website flagged as spam.

Anthropic says the cases it has identified had minimal real-world impact and did not involve customer data. The investigation remains open. For businesses testing agents, the practical warning is straightforward: a test can affect a real person or system if the software can still reach it. Instructions alone do not guarantee that an agent will stop at the boundary.

Anthropic / incident report · Oct. 9

04

Microsoft launches a low-cost model for sorting requests and choosing next steps

Microsoft’s new Decision-1 model is built for the small choices inside business software: which team should receive a support ticket, which incident needs attention first, or whether an AI-generated answer meets a checklist. Instead of writing a paragraph, it scores a set of options supplied by the developer.

The model is available in Microsoft Foundry and through OpenRouter. Microsoft lists input pricing at $0.042 per million tokens and no output-token charge. That makes it a candidate for high-volume sorting and checks. Businesses still need to test its accuracy on their own material and decide which uncertain or consequential decisions require a person.

Microsoft / product announcement · Oct. 9

05

A study finds people give up more often after AI help is removed

AI can help someone finish a task and still leave them less prepared for the next unaided attempt. A study highlighted by UC Berkeley tested 1,222 online participants on fractions and reading comprehension. People who had AI assistance initially performed better, but their accuracy and persistence dropped after researchers removed it.

The updated, peer-reviewed paper was presented in early October at the Conference on Language Modeling; an earlier draft circulated previously. These experiments do not establish that every kind of workplace AI use weakens skills. They do suggest a useful training check: after assisted practice, ask people to complete a fresh task without the tool before deciding they have learned it.

UC Berkeley / research report · Oct. 9

AI WORKFLOWS | MONDAY: CAPABILITY MAPPING

You can map an AI production before committing your time and money.

On October 7, filmmaker Zack London, known as Gossip Goblin, described the work behind Gods Don’t Give Gifts: 30 people and more than nine months. His breakdown includes two writers, 11 voice actors, four composers, nine core creatives and five post-production workers, including a Foley artist who creates sound effects. His role categories may overlap.

The studio says people write and direct its stories. The feature’s total budget, actual labor hours and complete tool stack remain undisclosed. Nine months on a calendar cannot tell you what your own project will cost.

Monday’s skill is capability mapping: breaking a job into tasks, testing which ones AI can handle, and identifying the people and resources needed to finish it.

Suppose you need a 60-second video explaining one service to prospective customers. Start with your public service description and images cleared for reuse. Your first deliverables are a production map and one draft you can evaluate.

Use ChatGPT Work to retrieve approved source material through connected apps and build the resource spreadsheet. For the pilot, the Explain Video Generator plugin can turn a brief into a narrated explainer with diagrams, animation and images. Access depends on your plan and workspace. Review the provider’s data handling and current usage charges before approving a generation. If the plugin is unavailable, Work can prepare the same map and an editor’s brief.

1. Set the limits. Choose the audience, one message, deadline and spending cap. Define acceptance: accurate claims, readable captions, clear audio and a specific next step for the viewer.
2. Assign every stage. Have Work list the output, available tool, human owner, dependency and quality check for each task. Mark untested capabilities and missing permissions. Include music licensing or a decision to omit music.
3. Measure one complete attempt. Approve the script and the exact material going to the video service, then generate a draft. Track rejected versions, corrections and review time. Supply any off-screen work time the agent cannot observe.
4. Total the actual cost. Have Work add tool charges, paid help and asset fees, listing existing subscriptions separately. Show your labor using an hourly value you choose. Separate hands-on hours from waiting time, and estimates from measured costs. Stop before exceeding the cap.

Give Work this assignment

Plan a 60-second service explainer using my approved source links and assets. Retrieve the sources, draft the script, and build an editable task-and-cost spreadsheet. Identify each human responsibility and unsupported capability. Return the plan for approval before sending material to a video provider or spending credits. After approval, produce one draft within my spending cap through the approved tool. Log observed costs, revisions and failed checks. Ask for missing time records. Do not publish or contact anyone.

For owners, use the completed map to decide whether to repeat the process, simplify the video or hire help for a specific stage. The same test works for a sales presentation, training guide or customer report.

Pick one small production this week and find out what it takes to get a version you would actually use.

AI TOOL SPOTLIGHT · Onepin

Onepin checks the words in your AI voiceover

For a product demo, training clip or service explainer, Onepin turns a script into AI speech and checks the result for pronunciation and word accuracy. It spells out prices, dates and abbreviations before generation, then can fix a troublesome word or regenerate a line. Try a short public script containing your business name, offer price and launch date, and listen to the finished take before publishing.

Free includes 1,000 one-time credits, but downloading audio requires a paid plan. Creator is listed at $16 a month billed annually, with usage overages.

Onepin / product details · Pricing and export limits · Verified Oct. 11

Everyday tools and business operations

Health, work and accountability

Security, models and research