FOUND FRIDAY · SEASON 5 · ISSUE 12

The assistant got a desk. The signature is still yours.

Grok Bot, OpenAI’s dots and Claude Duet point to the same shift: assistants with cloud computers, connected tools and continuing context. They arrived at different times and with very different access limits; a November 3 dots update brings the comparison back into focus. Meanwhile, SharePoint agent migration and Microsoft’s preview-channel changes put practical decisions on the calendar. The capability is arriving faster than the ownership model.

What I’m seeing this week: The assistant is moving from a window you open to an actor with an identity. It has an inbox, a browser, a cloud desktop and a growing memory of your preferences. That is a useful step, and it is also a new thing to inventory. The question is no longer whether the assistant can act. It is what it can reach, who authorized it, and where the record lives.

The pattern

The assistant now has its own computer, connected apps and memory.
Access is a plan, region and admin decision, not a demo.
The durable control is the approval boundary and the record.

Agentic Platforms / Personal Assistants

Your next assistant has a name, a memory and its own computer

Sources: xAI — Introducing Grok Bot; Grok Bot product page; OpenAI — Introducing dots; Anthropic — Introducing Claude Duet.

Summary: Grok Bot launched in beta on August 11, followed by OpenAI’s dots on September 29 and Claude Duet on October 21. This is an emerging category, not three launches this week. Grok Bot works on a persistent cloud computer, signs into tools and continues while your laptop is closed. Its current page lists paid individual, Teams and Enterprise access, with general availability for Enterprise and Teams. OpenAI’s dots are rolling out on Pro, Business Premium and Enterprise in eligible markets; a November 3 update added GPT-6.1 Terra to their model options. Claude Duet remains a US waitlist research preview for Pro, Max and Ultra subscribers, offering a morning brief and work on a cloud computer.

The operating details differ. Grok Bot’s bots share one computer per user, including its files, browser and logins: a separate bot is not a separate security compartment. OpenAI says dots’ proactive research uses read-only connected-app tools; that does not mean every task you explicitly assign is read-only. Anthropic says Duet asks for approval before sending messages, making payments, signing documents or deleting anything. These are vendor-described controls, not interchangeable guarantees.

Why it matters

The category moved from chat that answers to an assistant that acts. That changes the inventory. Each product can hold an identity, reach into connected tools and accumulate context about how you work. Useful, but it means a business now has to ask which assistant is signed into what, under whose account, and which actions it can take without asking. The access differences matter more than the demos. One is already included on team and enterprise plans, one is rolling out by plan and market, and one is a waitlist preview. A pilot that treats all three as equivalent will misjudge both risk and support load. The approval language is also not a guarantee: each vendor describes its own review flow, and the exact actions covered differ.

What to do now: Pick one assistant and one low-risk workflow. Write down what it may read, what it may change, and which actions require approval. Confirm the plan, region and admin setting that actually grant access before you promise it to a team, and name the person who reviews its work.

SharePoint / Agent Migration

Microsoft is retiring the .agent file, not the agents your team built

Source: Microsoft Learn — Upgrade agents in SharePoint and OneDrive

Summary: Microsoft Learn says agents stored as .agent files in SharePoint or OneDrive keep working across SharePoint, Teams and the Copilot app until the end of March 2027. Starting in mid-November 2026, creating an agent in SharePoint or OneDrive uses Microsoft Copilot Agent Builder and no longer saves a .agent file. The right path depends on the job: Copilot in SharePoint for site questions, a skill for repeatable tasks, Agent Builder for custom question-and-answer or analysis, or republishing a Copilot Studio agent to the Copilot app. Microsoft says there is no automatic upgrade. Existing permissions and sharing do not transfer automatically because the upgraded agent uses agent-level sharing. After March 2027, .agent files stop responding.

Why it matters

This is a migration with a date, not a rename. The file is the easy part. The hard part is that a working agent may be wired into a Teams chat, a page, a pay-as-you-go billing setup and a sharing model that will not come with it. The Teams detail is the one that catches people: an upgraded agent loads automatically in the Copilot app, but an agent already placed in a Teams chat or channel has to be removed and re-added. A site owner who tests only the builder will miss that. So will anyone who assumes the people with access today still have it after the move.

What to do now: Build a list of .agent files with owner, audience, page or Teams placement, knowledge sources and any pay-as-you-go dependency. Have the inventory and owner decisions ready before mid-November, then test one business-critical agent through its upgrade path, sharing and billing as soon as the upgrade experience is available. Put a named owner and a review date on every agent you keep.

Microsoft 365 / Preview Governance

Microsoft’s preview channel is changing. Your pilot group will not move itself.

Sources: Microsoft Learn — Targeted release retirement; Microsoft Learn — release options; Microsoft Copilot Blog — Frontier expansion.

Summary: Microsoft says administrators can no longer add, remove or modify Targeted release enrollment in November 2026; the channel retires in January 2027. At retirement, users follow their configured Standard or Deferred release preferences, defaulting to Standard if those preferences were not configured. Targeted release users are not automatically enrolled in Frontier. Starting in November, Frontier adds eligible previews for new Outlook for Windows, OneDrive, SharePoint and Teams. It remains optional and admin-controlled: enrollment itself needs no Copilot license, but each preview still needs its product’s licensing. Frontier and Deferred release are not currently available in GCC, GCC High or DoD environments.

Why it matters

Preview programs are where governance often goes quiet. A small group opts in for a demo, then a feature appears in a production workflow before anyone wrote a support plan. Consolidating previews makes the decision clearer, not smaller. One admin setting now spans four products, so one mistaken enrollment can widen the test surface well beyond the intended pilot. The licensing split is the practical trap: enrollment itself does not need a Copilot license, but the products do. And Targeted release is not the same as Frontier, which is exactly the assumption an admin is likely to make.

What to do now: Record your existing Targeted release pilot and decide its intended destination. Configure Standard or Deferred preferences where supported, and separately enroll the intended Frontier pilot. Confirm product licenses and cloud eligibility. Check the tenant’s notices for the November change rather than assuming enrollment can still be edited.

OpenAI / Workflow Migration

A migrated GPT can look finished while its integrations are missing

Source: OpenAI Help Center — Custom GPT retirement and migration FAQ

Summary: OpenAI’s migration FAQ says custom GPTs are scheduled to retire on 11 December 2026, with 11 February 2027 for FedRAMP workspaces and affected Enterprise workspaces with a qualifying, approved deferral. New GPT creation ended on 26 October for affected Enterprise workspaces, with FedRAMP excepted. Migration turns instructions into a skill, copies knowledge files and adds connected apps, but custom actions do not transfer. The selected model does not carry over. A migrated plugin starts private and does not automatically install or restore the original audience. Migration access varies by account and workspace, so the target date is not a guarantee that everyone can migrate on the same day.

Why it matters

The visible part of a GPT is the prompt. The part that makes it useful is often the action that writes to a CRM, opens a ticket or pulls a quote. Those are exactly the pieces migration does not carry. A team can test the new plugin, see sensible answers and miss that the workflow no longer completes. The sharing model is the second gap: a private replacement can look healthy to its creator while original users lack access. Check where the GPT was created, not just your own subscription: a public GPT created in an affected Enterprise workspace follows that workspace’s transition even when you access it from a personal account. Its replacement plugin is not automatically yours to use.

What to do now: List business-critical GPTs with owner, users, knowledge files and custom actions. Migrate one, then test the full task end to end, including the action and the sharing. Confirm your workspace’s actual retirement date in the product notice before you set a deadline.

Anthropic / Human Review

The human in the loop can still be following the agent’s lead

Source: Anthropic — How AI agents shape experts’ judgment calls

Summary: Anthropic researchers asked 50 experienced analysts to complete a realistic statistical analysis with Claude Code. Unknown to the analysts, the agent was quietly steered toward one of two reasonable options at four decision points. Analysts chose the agent’s preferred option 86 percent of the time; with no effect, the expected rate was about 50 percent. When analysts had stated a preference first, they switched 60 percent of the time when the agent disagreed, against 6 percent when it agreed. In post-task explanations, 77 percent reused the agent’s arguments but only 15 percent credited the agent. Anthropic says the setup deliberately amplified the agent’s preference and the sample was small and task-specific. It also reports that analysts using the agent finished faster and made fewer serious errors than a small, less experienced comparison group.

Why it matters

Human review is often treated as a control, but a reviewer who reads the agent’s summary before forming a view is not independent. The failure is quiet: the decision feels like the reviewer’s own, and the reasoning sounds familiar because it is. That matters most where the agent frames the options, such as vendor selection, risk triage or research conclusions. It does not mean the agent is usually wrong. It means the review process has to capture the human’s judgment before the model supplies the frame. Otherwise the approval step documents agreement, not independent judgment.

What to do now: For one consequential decision, ask the reviewer to write an initial view and confidence before the agent’s recommendation is shown. Then compare. If the final answer borrows the agent’s language, ask for the reasoning without the generated summary in front of them.

Google / Vault Compliance

Vault can now export the Gemini conversation, not just the document

Source: Google Workspace Updates — Google Vault now supports filtered exports for Gemini in Workspace

Summary: Google says Vault now supports eDiscovery filtered exports for Gemini in Workspace conversations across Gmail, Drive, Docs, Sheets, Slides and Meet. Admins can filter by custodian and date range, for organizational units up to 5,000 users, and export matching histories as JSON through the Vault interface or API. Exports include prompts, responses, generated attachments such as images, links to referenced Workspace files and citations. The feature is available now on Rapid and Scheduled Release for Business Plus, Frontline Standard and Plus, Enterprise Essentials Plus, Enterprise Standard and Plus, Education Fundamentals, Standard and Plus, and with the Vault add-on. Google says retention rules and litigation holds for Gemini in Workspace will follow. API support arrives within 14 days of the interface rollout.

Why it matters

Most AI governance conversations start with a usage dashboard. The harder question is what you can produce when legal, HR or an auditor asks what was said. A dashboard is not a record. This gives administrators a way to retrieve the conversation itself, with the files and citations it touched, which is closer to how email and documents are already handled. It is not complete: retention and holds are still coming, API access trails the interface, and the export only covers conversations where history was turned on. The edition list is also specific, so “we have Google Workspace” is not the same as “our plan can export this.”

What to do now: Ask your compliance owner to run one export against a non-sensitive test conversation. Confirm the edition or Vault add-on, the retention setting and whether conversation history is actually on for the group you expect. Write down who is authorized to run an export before the first real request arrives.

Field Notes from My AI Workstation

I handed over the follow-up, not just the draft

This week I delegated the coordination of a 30 minute meeting. Storm checked my Outlook availability, sent three available options from its own agent-owned inbox, and set a 15 minute check for replies on that exact thread. It did not book anything. No reply had arrived when I wrote this.

That is the part worth reporting. The agent did not just draft a message for me to send. It sent the message, from an identity of its own, and then waited inside a boundary I could describe. The boundary was narrow: one thread, one purpose, a short check interval and no authority to put a meeting on the calendar.

The useful milestone is that the follow-up exists, the agent uses a sending identity of its own, and the coordinator is set up to answer questions within the original scheduling request. It must bring a scope change back to me. Those are operating instructions, not a claim that the system is infallible. The meeting was not booked when this workflow was set up.

The business lesson is that delegation has a middle state. Most agent demos jump from “the assistant suggested something” to “the assistant did everything.” Real work spends time in between: the message is out, the answer has not arrived, and the system has to know what it is waiting for and when to stop. That middle state needs an owner, a scope and an end condition, or it becomes a background process nobody is watching.

For a small business, this is where the operational value sits. Follow-up is the work that slips when everyone is busy. An agent that can send, watch one thread and stop is useful. An agent that can send, watch everything and improvise is a liability. The difference is not the model. It is the instruction and the permission boundary around it.

I authorized the task. Storm carried the sending and the waiting. A person still owns the booking, and the agent still has to report what it did.

What to do now: Before you delegate a follow-up, define the thread, the allowed action, the check interval and the stop condition. Give the agent its own identity if it will send mail. Then verify what it actually sent and what it is still waiting for.

Getting Started with Hermes

Want to build your own practical AI workstation? My beginner-friendly Hermes setup guide walks through the mindset, setup, safety boundaries, and first workflows for a personal AI assistant.

Get the Getting Started with Hermes guide

An assistant with its own computer is not a chatbot upgrade. It is a new identity inside your business, and an identity needs an owner.

Closing Thought

This week’s news is about the same boundary from different angles. An assistant can now hold an identity and act on its own computer. A SharePoint agent has a migration date whether or not its owner is ready. A custom GPT can be retired while its integrations quietly fail to follow. A compliance export can exist while retention rules are still on the roadmap. The tools are becoming more capable and more operational at the same time. That makes the management questions more important, not less. Who authorized this? What can it reach? Where is the record? Those are still the questions that decide whether AI saves work or creates a new kind of cleanup.