Get tomorrow's brief in your inbox
Today: Zapier's workflow benchmark found that even its top-scoring AI model completed only about half of complex business tasks correctly, a warning against unattended automation. Google is opening Search and Gemini integrations to U.S. developers through a preview waitlist, creating a potential new customer channel for businesses with apps.
AI agents still need supervision, and the most expensive model is not necessarily the best choice.
Zapier's AutomationBench tests whether agents actually finish business workflows, rather than simply produce convincing answers. Its leaderboard covers 657 tasks across 47 simulated apps, including CRM updates, email routing, spreadsheets, and support cases. Gemini 4 Argon leads at 51.3%, followed by Claude Sonnet 5.5 at 44.7%. These are results from Zapier's particular test environment, not predictions of your business's failure rate, but they make a strong case against trusting an agent's "done" message.
The dangerous failures are quiet ones. In one example, an agent must distinguish similarly named customers, use the current spreadsheet record, and find an email containing the notification rules. Zapier says agents often report success even when the underlying records show otherwise. For your business, verify the changed records and recipients, not just the generated summary. Start with draft-only or approval-gated workflows before allowing customer-facing actions.
Price also fails as a shortcut for quality. Zapier reports that GPT-6 Sol achieves roughly 80% of GPT-6 Astra's score for about one-sixth of its cost per task, while Claude Sonnet 5.5 outperforms the more expensive Opus 5.5. Its cheapest examples, GPT-6 Luna and DeepSeek V4 Flash, cost $0.04 per task but complete fewer than a quarter of full workflows. Test models on representative work and compare cost per correctly completed task, including review and corrections. Cheap classification can be useful; cheap unattended operations can become expensive.
Google Connected Apps: a potential distribution channel, not an automatic traffic boost. Google's platform lets developers submit apps that users can access through Search, Gemini, or both, with monitoring for traffic, latency, and errors. The preview is available through a waitlist for U.S. developers; it uses a cloud-hosted Model Context Protocol server or Android MCP through AppFunctions APIs. If your business already operates an app, identify one useful customer action and investigate the preview before committing development resources; the opportunity is reaching customers inside an AI interface, but additional sales are not established. Sources: TechCrunch, Small Business Trends.
Slackbot Shareable Skills: reuse workflows across company boundaries. Small Business Trends reports that Slackbot skills can now be shared externally through links, customized by recipients, and controlled through link revocation and incoming-request management. If you collaborate with agencies, clients, or partner firms, try sharing a non-sensitive weekly-summary workflow after checking your workspace's access and inspecting the skill for confidential instructions. The business opportunity is less duplicated setup, provided someone owns updates and reviews what leaves the organization. Source
ServiceNow and ChatGPT: separate today's plugin from the roadmap. OpenAI says the ServiceNow MCP plugin is available in its marketplace today, while enhanced capabilities are expected by year-end; security early access is planned for January 2027, voice for the first quarter, and Codex in AI Workflow Factory during 2027. Existing ServiceNow customers should ask their administrator which actions are currently supported and test permissions before enabling record changes, rather than purchasing around preview demonstrations. The potential benefit is fewer tool switches during IT work, but planned features should not enter today's savings forecast. Source
Start conversational analytics with one trusted dataset. Small Business Trends describes analytics assistants that answer ordinary-language questions about sales, customers, and costs, with permissions and activity records as important selection criteria. For a shop or service business, pilot a read-only sales dataset and ask questions whose answers you already know before exploring unfamiliar trends. This can reduce report preparation, but an explanation of a sales decline still needs checking before you change staffing, purchasing, or promotions. Source
Give every automation an owner before adding another one. n8n's internal automation leader says the company has more than 12,000 internal workflows and argues that a central team cannot indefinitely build and maintain everything. Apply that lesson at small-business scale: record each workflow's owner, purpose, dependencies, failure alert, and manual fallback, then have the people closest to the process help maintain it. This is a recommended operating practice, not a reported n8n checklist; its value is preventing your time-saving projects from becoming an invisible maintenance backlog. Source
Automate contract triage against an approved playbook, not an open-ended judgment call. In an OpenAI customer story, Legora's general counsel estimates that NDA review fell from about 20% of her legal team's capacity to under 1% after introducing a workflow that flags terms outside historically accepted parameters. Those are customer-reported results, not a guaranteed outcome for another firm; ask your legal adviser whether a narrow, repeatable contract category would suit a similar first-pass process. The opportunity is faster triage while keeping exceptions and final decisions with qualified reviewers. Source
Measure AI-search visibility weekly instead of reacting to individual answers. HubSpot recommends daily collection where supported, weekly interpretation, monthly conversion reviews, and quarterly pipeline context; its AEO tool tracks prompts across ChatGPT, Perplexity, and Gemini daily. If customers research your category through AI search, keep a stable set of buyer questions and compare weekly patterns before rewriting content. The benefit is avoiding work driven by fluctuating answers rather than meaningful changes in visibility or customer acquisition. Source
Claude marketing plugins require a paid plan. HubSpot's guide says plugins are available on Claude Pro, Max, Team, and Enterprise, with discovery under the Customize menu. Plugins bundle capabilities such as reusable instructions and connectors; the HubSpot connector exposes CRM records available to the connected user. If you already pay for an eligible plan, test one recurring campaign brief or reporting task with limited permissions, and confirm current subscription and connector costs before expanding access. The useful purchase is a repeatable workflow, not another source of generic marketing copy. Source
Perplexity offers a free starting point for source-based research. Zapier's guide says Standard Search is available on the free plan, while Pro Search is the default on paid plans; it also describes Deep Research and a Computer agent for multistep work. Start with Standard Search for a low-risk research task and inspect the cited pages yourself; verify current plan eligibility and limits before buying for advanced features. The opportunity is faster initial research, not removing fact-checking or spreadsheet review. Source
Run a "prove it worked" automation test. Choose one low-risk workflow and write down its expected record changes, recipients, and prohibited actions. Run it on test data, inspect every outcome, and log corrections alongside execution cost. This turns Zapier's outcome-based testing approach into a practical buying and rollout check.
Try a tightly scoped analytics prompt: "Using only this approved sales table, total revenue by service for last week. State the date range, identify missing or duplicate records, and show the calculation behind each total. Do not infer why sales changed." Compare the answer with your existing report before asking broader questions. This follows the plain-language analytics article's recommendation to check answers against known numbers.
AI tools are moving closer to business records and real actions, making permissions, process ownership, and outcome checks more important than polished demonstrations. Watch Google's preview access and ServiceNow's delivery milestones, but prioritize one supervised workflow you can measure over a broad promise of autonomous operations.