|
| Agentic Risk and Reward Most of the business leaders I'm talking to are seeing benefits from AI, and the numbers back it up. 78% of small and mid-sized businesses (SMBs) say AI has improved their productivity, according to a report by Intuit.
And, most of those businesses got there on a similar path. First, individuals prompt on their own, and AI is helpful but isolated. But after continued success and comfort, AI gets securely connected to a limited set of company data, like an inbox, files, or the CRM. Finally, an agent is deployed to use that information and take over a process. Each step on the path makes AI more useful because it has access to more information and can do more with that information.
But as AI moves along this path, the risk moves with it. The question is: can corporate governance keep pace? In a recent survey of SMBs by Pax8, more than a third of the respondents found unapproved AI tools, workflows, or agents, and fewer than half formally review what their people are building. | Step one on the path is prompting. Deloitte surveyed 25,000 workers across 22 industries. They found that 63% have used generative AI at work, and nearly a third of those do it without their employer knowing. Most of that use is ordinary and time-saving: drafting emails, summarizing documents. On their own, those tasks carry little risk. But if this AI use is not governed, the risk is in what employers can't see: which tools people pick, and what company information they paste into them. Key Takeaway: As an owner or manager, most of the fix is communication: train early, ask your team what they're using, and talk about the tools from your experience. If you're using AI your company hasn't approved, speak up now so your experience can help shape your company's policies. |
Step two is connecting AI to your data. One example of this is Claude's plugin for Small Business, which links to QuickBooks, PayPal, HubSpot, and many others. Two safeguards are built in: your existing permissions carry over, so an employee can't see more through Claude than they can today, and nothing sends, posts, or pays until you approve it. The Practical Angle: A connector gets the same access as the login it uses. So start with one system and configure read-only access. Then widen the scope once you've confirmed the security and the results. See Plugin Details → Step three is handing the process to agents, and this is where the payoff grows. McKinsey's 2026 State of AI survey found that 40% of large organizations and 22% of SMBs are scaling agents in at least one function. As well, 32% of the respondents skipped buying at least one software product because agentic coding tools let them build it in-house. My Take: The survey shows that large organizations have led agentic AI adoption. But because the models and tools are more powerful, and less expensive, SMBs are now well positioned to take advantage of this third step in the AI journey. Read McKinsey's State of AI → | Quick Hits.Foundations Controlling AI Agents Agentic AI is a combination of an LLM, connected data, and tools. The model itself only proposes an action. A separate layer of software (the harness) decides whether the action runs, needs your approval, or gets blocked. It also decides what the model can and can't read. | .Video Connect Your AI to Your Business Data This segment shows how to connect AI to business data to analyze pricing, invoices, and contracts, and how to use agents to automate tasks in payroll and reordering. It also highlights the key AI risks (hallucinations, bias, privacy) that businesses must govern. Watch from 4:20 to 27:10. | .Deep Dive Malicious AI Skills and MCP Connectors Skills and MCP connectors share a vulnerability: both load instructions straight into the model. Messages hidden in a skill's instructions or a connector's description can maliciously alter an agent's behavior. Proper governance includes reviewing and managing access for every skill and connector used.
New to skills or MCP?
Start with the Agent Skills overview or the MCP overview. |
|
| Industry DevelopmentsAdvancements in Open-Weight AI Mistral's ML4 and Reflection's Beam join a growing field of open-weight models, ones anyone can download and run on their own servers. The big win is control: sensitive data stays in-house. If the cost and performance claims hold up, more businesses will find they don't need the frontier models. | AI Significantly Improves Weather Forecasts Microsoft Research built an open-source AI layer that corrects existing weather models' errors 2 to 6 weeks out, the hardest range to predict. It correctly forecast two major U.S. winter storms 19 to 26 days ahead. For utilities, farms, logistics, and families, that's weeks of extra warning. |
|
| |
|
