Source: raw/OpenClaw_Was_Too_Dangerous_To_Install._7_Months_Later_Microsoft_Is_Bringing_It_To_Your_Company..md — Creator: Nate B Jones, AI News & Strategy Daily · URL: https://www.youtube.com/watch?v=avN5R3OhxYs · Platform: YouTube (auto-captions render OpenClaw as “open Claude”; fetched 2026-10-21). One comparison with OpenAI’s Dots is from his follow-up video, raw/Should_You_Pay_100_A_Month_For_OpenAI_s_Dots_When_Meta_s_Muse_Has_A_Free_Version.md (https://www.youtube.com/watch?v=lnB4Zckx_34).
Jones’s argument: Microsoft’s Autopilot puts an OpenClaw-style persistent agent inside Excel, Outlook and Teams, and it will reach more workers than anything OpenAI or Anthropic ships this year because of distribution, not intelligence. “I’ve played with Autopilot, and it’s frankly, it’s not super intelligent … The value comes from the data it can use and how well you know how to use AI.” Most of the video is five habits that work with the Copilot people already have, which makes it useful for anyone whose employer allows only Microsoft tools. Product facts and seat counts are as Jones relays them.
Key Takeaways
- What Autopilot is. A Microsoft agent with an identity “your IT department can audit” that holds “an assignment while you’re off doing something else” inside Microsoft 365 apps. It is in “expanded private preview”, and Microsoft describes Scout, its earlier name, “as literally powered by” OpenClaw’s open-source technology.
- Distribution, as Jones relays it. Microsoft reported 450 million+ paid Microsoft 365 commercial seats in January and 30 million+ paid Microsoft 365 Copilot seats in July; OpenAI reported 9 million+ paying business users in February.
- Habit 1: say what good looks like. “Asking Copilot to summarize everything leaves the entire standard … largely unstated.” His Copilot hit rate is “50 60% of the time”; it goes up when you show a good previous briefing and explain what is useful about it.
- Habit 2: supply the sources and their priority. “Tell it which source takes priority when documents disagree. Most people do not do this.” Ask for linked evidence and open the links: does the source say what the summary says, and is it this week’s update or last month’s plan?
- Habit 3: make the assignment recur. Specify what to check, how often, “what counts as a material change”, when to bring you in, and whether it may contact anyone or “only prepare a draft”. Until Autopilot reaches you, scheduled Copilot prompts are the earlier route. Check whether the next run noticed what changed.
- Habit 4: match reasoning to difficulty. Microsoft’s auto router picks models on accuracy, speed and cost, but a router that sees only “What is this?” cannot tell a syntax typo from a problem you are three hours into. Tell it what you already tried and what a mistake would cost. “Count the time you spend repairing the result”: a cheap model that misses a customer commitment costs more than the bill shows.
- Habit 5: find out why a run failed, then share the fix. Missing source, unclear instruction, or evidence misread? Each needs a different fix. Save the correction, check it helped, and share it. In enterprise after enterprise, “a small group, call it 5%, figure out how to get useful work out of AI.”
Details
The worked example: a customer renewal. Ask Copilot to identify unresolved customer commitments and what must happen before the renewal conversation; then judge whether it separated a promise someone made from a request the customer made. Microsoft’s Work IQ announcement uses a similar case, where the sales record looks healthy but service tickets and correspondence show warning signs. When the answer is weak, “check what information the tool had before deciding that the model can’t do it”. Often the relevant system is not connected or your account cannot read it.
Copilot Notebooks can hold project references and instructions in one place, though you still rebuild context in each prompt.
The counter-argument he takes seriously. Better models may remove steps rather than automate them: a report and a meeting may exist only for human coordination. Ask the agent “what a proposed change would remove, what information it would need, how you would check the result”. He notes that Claude has add-ins for Excel, PowerPoint and Word, so a company could keep its Microsoft documents while employees hand the work to Claude.
Router economics. Jones says Microsoft has not disclosed what its router uses as input, and that an auto router “may burn more tokens figuring all of that out than you would just sending it to a smart model in the first place.” Don’t resubmit the same request hoping the fifth answer is better.
Versus OpenAI Dots (follow-up video). “Autopilot may not run on a frontier model, but Autopilot is going to be trusted because it’s Microsoft”, whereas OpenAI put its strongest model, Astra, behind Dots. Facts about Dots belong to the OpenAI articles linked below.
Meta’s move. Meta announced Meta Enterprise Platform, with Meta Business Agent and developer tools, following its Muse personal agent.
Try It
- Write the standard into your next Copilot (or any assistant) request: audience, length, what must be included, and one example of a good past output.
- List your sources in priority order and say which wins on conflict; ask for a link per claim and open two of them.
- Turn one weekly report into a scheduled prompt with a written “material change” rule and draft-only permission, then compare run 2 with run 1.
- When you escalate to a stronger model, include what you tried and why the cheap answer failed.
- Keep a shared doc of fixes (“missing source”, “instruction unclear”) so a second person benefits from each failure.
Open Questions
- Autopilot’s rollout dates, pricing and which models it runs on are not given.
- The seat counts are Jones’s relay of Microsoft and OpenAI announcements; the primary releases were not checked here.
- The “5%” figure is anecdotal: “an actual number I’ve heard over and over again”.
Related
- Microsoft Agent Governance Toolkit — Microsoft’s governance side of enterprise agents
- OpenClaw Concepts Walkthrough — the open-source agent Autopilot descends from
- Cost and Intelligence Levers — routing and the true cost of a cheap model’s misses
- Recurring Task Automation (Nate B Jones) — habit 3 in more depth
- AI Rollout Resistance — why most employees don’t reach the 5%
- Building an AI Champions Network — a structure for spreading the fixes habit 5 produces
- GPT-6 Astra, Sol and Luna — the model behind OpenAI’s Dots