ClawPilot: A New Way to Work with AI

Chris McNulty
1 hour ago
4 min read

Personal, powerful, and enterprise-ready

Last October, in our annual AI predictions, I argued that 2026 would bring “agent boss” platforms — tools that let business users coordinate multiple AI assistants without endless app-switching. Halfway through the year, that prediction is playing out, but in a more interesting way than I expected.

Three trends are converging on something I’ve started calling ClawPilot — a personal AI assistant with the capability of an open agent and the trust profile of an enterprise copilot.

ClawPilot (n.) — the convergence of open-agent versatility (the “claws” that do real work) with enterprise copilot orchestration (the “pilot” that provides guidance, security, and trust).

From “Agent Boss” to “ClawPilot”

My original prediction focused on the messy, manual process of toggling between ChatGPT, GitHub Copilot, and other point tools. I expected a major vendor — likely Microsoft — to deliver a control surface over all of them by 2026.

The pieces are arriving. But they’re consolidating differently than I forecast. Instead of one boss managing many bots, what’s emerging is one cohesive assistant that absorbs many of those bots’ capabilities. That’s the shift worth paying attention to.

Three forces are driving it:

OpenClaw and personal agents. Highly capable agents that take real action across email, calendar, web, and messaging. They’re extensible — you can teach them new skills — and deeply personalized. In their raw open-source form, they’re also unmanaged, which makes them a serious risk inside a company.
Enterprise copilots. Microsoft 365 Copilot has matured from a chat assistant into a platform that respects identity, permissions, and audit. It’s safe for business use, but it’s historically been more about helping you draft than acting on your behalf.
Long-form, persistent workflows. New patterns like Microsoft’s Copilot Cowork — where an agent runs tasks in the background and checks in at decision points — treat AI as a continuous collaborator instead of a turn-by-turn responder.

Each of these alone is interesting. The convergence is what matters.

What OpenClaw proved

OpenClaw landed earlier this year and became the fastest-growing open-source AI project I’ve ever seen — more than 100,000 GitHub stars in its first week. The reason isn’t mysterious. People want an AI that does the thing, not one that explains how to do the thing. Early users have it triaging email, negotiating with vendors, pulling data from the web, and running personal errands across half a dozen apps.

The capabilities are real. The security profile, in its raw form, is awful. An open agent on your laptop is effectively untrusted code with persistent credentials. A prompt injection hidden in a single email can turn it into a data-exfiltration tool. Microsoft’s own security team reportedly described it internally as a virus with a friendly face.

That’s the gap leaders need to think about clearly. The capabilities aren’t going back in the box, but the raw form isn’t enterprise-deployable.

What enterprise copilots solved

Meanwhile, Microsoft 365 Copilot — and specifically Cowork — has been quietly closing the other half of the gap. Cowork doesn’t just generate content; it executes multi-step work across Outlook, Teams, SharePoint, and the rest of the stack. Triage a calendar. Pull together a client brief from a dozen sources. Schedule the prep call. Surface checkpoints for human approval before anything goes out.

The important part isn’t the demo — it’s the substrate. Every action runs under your identity, against data you already have permission to see, with audit trails behind it. Multi-model orchestration (including Anthropic’s Claude models alongside Microsoft’s own) means the system picks the right engine for each task. Users don’t have to care which model is doing what.

What enterprise copilots haven’t fully delivered yet is the always-on, deeply personalized feel of an open agent. They’re still mostly invoked, not ambient.

The convergence

ClawPilot is what you get when you stop treating these as separate categories. Take the capability surface of OpenClaw — the cross-tool reach, the extensibility, the willingness to act — and put it inside the trust boundary of an enterprise copilot. Add Cowork-style background execution with approval checkpoints, and you have an assistant that can genuinely run a knowledge worker’s day without exposing the business.

Concretely, that means an assistant that:

Wakes up before you do, sweeps overnight email and Teams traffic, and assembles a morning brief.
Spots that your afternoon is back-to-back, moves a non-urgent check-in, and protects an hour of focus time.
Drafts the deck for your 2pm using your team’s SharePoint files and past notes, then pauses for your review before anything ships.
Operates entirely under your corporate identity, with every action logged and every external-facing step gated by a human checkpoint.

Microsoft’s internal “Project Lobster” is the early proof point — an OpenClaw-style assistant built by Omar Shahine’s team that scaled from 100 to over 3,000 daily internal users in days. People who get a taste of this don’t go back.

Why it matters

The productivity case writes itself. Every knowledge worker effectively gains a small team. The more important point for leaders is the governance case. If you don’t give your people a sanctioned ClawPilot, they will build or download one. Shadow AI is already a problem; raw OpenClaw deployments will make it acute.

Leaders who get this right in the back half of 2026 will do three things:

Update AI strategy to assume agentic, not chat-only, usage.
Engage security and IT early — agent identity, least-privilege access, and approval gates are now first-class design problems.
Train people to manage AI, not just prompt it — defining goals, reviewing output, providing feedback. Delegation is a skill.

ClawPilot isn’t a product yet. It’s a pattern. But the pattern is converging fast enough that I’d treat the second half of this year as the planning window, not the wait-and-see window. The age of the personal AI agent has started — the question is whether yours arrives sanctioned, or smuggled in. P.S. For the record, nothing beats melted butter for lobster. As for lobster rolls, Maine-style (cold on hot toasted buns) is my personal favorite, but those can be fighting words in Connecticut.