Get tomorrow's brief in your inbox
Today: OpenAI's Decisions API is generally available for fast classification and routing, with reported speed gains that deserve testing in your own workflows. Canva launched Craft for image generation, while 1Password introduced agent logins without exposing passwords. Microsoft's account of the ConnectWise breach shows why agent access needs tight controls.
Small decisions are becoming a practical AI workload
OpenAI's newly generally available Decisions API handles bounded questions rather than open-ended writing. A developer supplies context and a question with permitted answers, and the API returns a choice, probability, or score. For a small business, that could mean routing incoming support requests, categorizing transaction descriptions, or deciding which requests need human review. OpenAI says the service accepts text, images, and native JSON, with up to 400 questions per request.
The attraction is responsiveness. OpenAI claims speeds up to 10 times faster than GPT-6 Luna through its Responses API. Separately, language-learning company Speak told TechCrunch that the API reduced time between actions from 1.4 seconds to 0.13 seconds. Those are vendor and customer reports, not a promise that your workflow will see the same improvement. A faster decision also is not necessarily a more accurate one.
Sources: OpenAI announcement, TechCrunch reporting
Canva Craft: try it for text-heavy marketing graphics. Canva's first image-generation model is available to all users through Dream Lab, with output up to 2K resolution and 30 aspect ratios. Canva says it supports legible text and precise edits, making menus, promotions, and social graphics sensible test cases; check every price, date, and product claim before publishing, and confirm your account's generation allowances rather than assuming unlimited free use. Business impact: If its editing claims hold up on your designs, you could spend less time fixing layouts and rebuilding assets. Source
1Password for Agents: delegate logins without handing over passwords. The new platform verifies an agent's identity, asks the user to approve a login, and fills credentials without exposing them to the agent. If you already use agents for account-based tasks, check supported integrations and current pricing, then pilot a low-risk account; password protection does not remove the need to limit what an authenticated agent can do. Business impact: It offers a more controlled alternative to copying credentials into assistant conversations. Source
HubSpot Agent Builder: start with a recommendation, not an irreversible action. HubSpot says its users can describe an automation in a no-code builder and connect it to CRM data, with controls over which actions run independently. Try a repeatable job such as reviewing lost deals and recommending follow-up material, but verify account eligibility and charges before committing. Business impact: Your sales team can test whether an agent reduces recurring analysis work while keeping customer-facing decisions with a person. Source
Copy Vetano's test-and-approve workflow, not its entire software stack. In Zapier's customer story, three-person hiring startup Vetano used Claude and Next-Gen Zaps to revise welcome-email workflows that founder Chris Fairley estimated would otherwise take two or three hours of editing. Before deployment, a test email goes to Fairley for approval, and the previous workflow is retained but switched off. Your move: Apply that pattern to one onboarding automation, with separate tests for customer types, duplicates, and test accounts; the opportunity is faster maintenance without silently sending the wrong message. Source
Check AI claims before they trigger another action. n8n recommends layered hallucination detection, including checking whether individual claims are supported by retrieved documents and escalating questionable outputs. Its important distinction: agreement between repeated model answers is not proof of accuracy, and faithful use of a source does not prove that source is correct. Your move: Add source checks and human approval before an AI answer changes a record or reaches a customer; the payoff is less downstream cleanup from confident mistakes. Source
Measure context-gathering time before buying another assistant. A Dropbox survey of 219 professionals at The AI Conference 2026 found that respondents spent an average of 5.8 hours weekly retrieving context for AI work. That conference sample is not a benchmark for every small business, but it suggests a useful audit: track time spent finding documents, supplying background, and saving usable results. Business impact: Organizing approved reference material may address your bottleneck more directly than adding another model subscription. Source
Treat ERP results as modernization evidence, not isolated AI ROI. Small Business Trends reports that Volumetric Building Companies cut purchasing, manufacturing, and inventory cycle times by 50% within three months after implementing SAP Cloud ERP with IBM. The account does not isolate AI's contribution, so use it to frame a process-modernization discussion rather than forecast equivalent savings. Your move: If disconnected operational systems are slowing you down, request a phased proposal with integration, training, and disruption costs included. Source
Self-hosting exchanges API dependence for operating responsibility. n8n explains that keeping prompts inside your network requires both the model and the workflow handling requests to run on infrastructure you control. Its recommended approach is to identify the smallest model that reliably handles the task, then assess deployment requirements. Your move: Compare hardware, maintenance, and operating effort against cloud usage, and use only approved nonsensitive examples when testing external providers; a hybrid arrangement may fit better than moving everything locally. Source
Existing Muse artifact links have changed. Meta says sharing is back on after a fix, and previously shared artifacts now have new URLs; old links direct visitors to request replacements. Researchers described discoverable shared conversation pages containing sensitive business information, while Meta says the issue concerned artifact URLs rather than chat sharing. Your move: Review sensitive shared material and its audience before distributing replacement links, then update legitimate collaborators to avoid broken handoffs. Source
ConnectWise customers should check exposure with their IT provider. Microsoft researchers say China state-aligned attackers stole application credentials from ConnectWise's cloud backup service and used AI agents to compromise around 2,250 organizations. ConnectWise said it paused access and was investigating while working to restore service. Your move: Ask whether your organization uses the affected service, review relevant Microsoft-environment sign-in activity, and follow current vendor guidance on credential remediation and recovery; using ConnectWise alone does not establish compromise. Source
Agent Skills is moving toward vendor-neutral governance. Simon Willison reports Anthropic's announcement that it is donating the Agent Skills standard to the Agentic AI Foundation under the Linux Foundation. For businesses documenting repeatable agent procedures, this is a reason to favor reusable instructions over deeply proprietary configurations, not proof that every agent supports interchangeable skills. Your move: Ask vendors what they support and test portability before making it a purchasing assumption. Source
Create a source-bound drafting prompt. Try: "Using only these approved documents, draft an answer to this customer question. Identify the supporting passage for each factual claim. Flag anything that needs confirmation. Do not send the answer." Review the result yourself, since a request for citations is not an accuracy guarantee. Related guidance
Audit one automation's API version today. Record where the version is specified, who owns the integration, and how you will test a provider change before production. n8n explains that version pinning separates your workflow from incompatible updates; checking this now can reduce avoidable maintenance surprises later. Related guidance
The practical pattern is narrower AI jobs paired with stronger controls: faster classification, editable creative output, and approval-based account access. This week, prioritize one measurable pilot and one access review rather than a broad autonomous-agent rollout. Judge success by accepted work, errors, operating cost, and review time, not output volume alone.