14:35 Lena: I want to stay on that comparison piece for a second. If I’m comparing two Excel sheets, say, two different versions of a budget, can Claude actually tell me, "Hey, the marketing spend in row 42 went up by 15%"?
4:19 Miles: Exactly. It’s not just saying "the text is different." It’s performing the calculation. It uses the `Get Range` tool to pull data from both workbooks, does the math in its internal environment, and then it can even use the `Update Range` tool to create a "Variance Report" in a third spreadsheet. It can even add a chart—there’s a tool called `Add Chart` in the Excel MCP—to visualize that variance automatically.
15:12 Lena: Okay, that’s impressive. But what if the files are in different places? One is in my OneDrive and one is on a SharePoint site I just got access to?
15:21 Miles: That’s where the power of the "Unified Bridge" comes in. As long as you’ve authenticated through Composio or the M365 connector, Claude doesn't care if the file is in OneDrive, a SharePoint library, or even an email attachment in Outlook. It uses the Microsoft Graph API to bridge those silos. You just say, "Compare the budget in this email with the one in the Finance site," and it handles the "where" so you can focus on the "what."
15:45 Lena: And the "Output" part of that—the listener asked about sharing with a specific user. We talked about permissions, but can Claude actually *send* the notification?
15:55 Miles: It can do better than just a notification. It can actually draft the email or the Teams message. There are specific MCPs for Gmail and Slack, and of course, the native M365 connector handles Outlook and Teams. So the workflow looks like this: Compare the files, generate the summary, and then use the `draft_email` or `send_message` tool to deliver it directly to the person who needs it.
16:18 Lena: So I’m not even leaving the terminal to send the "Hey, here’s the comparison" email.
16:23 Miles: Not at all. You can even have Claude include the sharing link it just generated in the body of the message. It’s a completely closed loop. But here is where we have to talk about guardrails again. Because if you give an AI the power to send emails and change permissions, you need to make sure it’s not "hallucinating" a new boss.
16:41 Lena: Right! I don't want it sharing the "Executive Salaries" sheet with the entire company because it got a little too enthusiastic.
4:19 Miles: Exactly. This is why we use "Deterministic Guardrails." We move away from just telling Claude "Be careful" in a prompt—because prompts can fail under pressure—and we move toward "Hooks."
16:59 Lena: Hooks? Like Captain Hook?
17:02 Miles: More like a safety catch. A Hook is a little piece of code that runs *before* or *after* Claude uses a tool. For example, you can set up a "Pre-Tool Hook" that triggers whenever Claude tries to use the `Add Workbook Permission` tool. The hook can check: "Is this user on the approved list?" or "Is this file marked as 'Highly Confidential'?" If the answer is no, the hook returns an exit code that literally blocks the command.
17:27 Lena: So even if Claude *thinks* it should share the file, the system says "Nope, not allowed."
4:19 Miles: Exactly. It’s a "Hard No." It’s not a suggestion; it’s an enforcement. You can even set up hooks to automatically run a linting check or a security scan whenever a file is edited. There’s a great plugin called "Hookify" that makes this easy. You can just type `/hookify Block sharing of any file in the /Confidential directory` and it creates the rule for you. No complex JSON editing required.
17:57 Lena: That makes me feel a lot better about giving it access to SharePoint. It’s like having a security guard who never sleeps and doesn't get distracted by a "pretty please" from the AI.
18:06 Miles: That’s a great way to put it. And because today is 2026, these security patterns are much more mature. We’ve learned from the early mistakes—like the time a developer asked Claude to document their API config and it accidentally hardcoded the secret key into a public markdown file. That cost $30,000 in fraudulent charges. Hooks are the answer to those kinds of "human-in-the-loop" failures.
18:28 Lena: $30,000... yeah, I definitely want the hooks. But let’s talk about the "SharePoint Access" part specifically. I noticed in the materials that there are different ways to connect—like "Claude Code" versus "Claude Cowork." What’s the difference?
18:43 Miles: That is a really important distinction for an enterprise user. "Claude Code" is that terminal-based agent we’ve been talking about. It’s great for engineering tasks, bulk file operations, and using MCPs. "Claude Cowork" is more of a general-purpose autonomous agent that runs on your desktop. It has local file access and can even use a browser.
19:03 Lena: So Cowork is more like an AI assistant that lives on my Mac or PC?
4:19 Miles: Exactly. It uses OneDrive Sync to see your files. This is a big point for IT teams: for Cowork to work with SharePoint, those libraries *must* be synced to your local machine via the OneDrive client. If it’s not synced, Cowork can't "see" it. Claude Code, on the other hand, uses the API to talk to the cloud directly.
19:25 Lena: Okay, so if I’m doing high-volume automation, I want Claude Code with MCP. If I want a "digital buddy" to help me with my daily tasks, I’m using Cowork with synced folders.
19:36 Miles: You nailed it. And Cowork has its own set of challenges—like the fact that its activity is currently excluded from Anthropic’s standard Audit Logs. So if you’re in a regulated industry like healthcare or finance, you really need to lean on things like OpenTelemetry to capture what the agent is actually doing.
19:53 Lena: "OpenTelemetry"? That sounds like something I’d need a rocket scientist for.
19:57 Miles: It sounds more complex than it is! It’s basically just a standardized way to stream logs. You can point Claude to an "OTel Collector," and it will spit out every tool call, every prompt, and every file it touches. You can then pipe that into a dashboard like Grafana. It’s how you prove to an auditor that the AI didn't do anything it wasn't supposed to.
20:17 Lena: So, transparency is the final piece of the puzzle. We have the bridge, the data architecture, the workflows, the guardrails, and now the audit trail.
20:27 Miles: It’s a complete governance framework. And it’s what turns a "cool AI tool" into a "production-ready enterprise system."