Found Friday Season 5 Issue 13: Five failures in five days. The fallback plan is the story.

Found Friday newsletter
FOUND FRIDAY · SEASON 5 · ISSUE 13

Five failures in five days. The fallback plan is the story.

Microsoft 365, the FAA, Workday, Chase and United experienced disruptions inside five days. No common cause has been established. For businesses, the useful connection is the work that stops when a critical service disappears. This issue also looks at AI tools taking on more responsibility, and my own experience handing real work to Storm and a growing roster of Grok bots.

What I’m seeing this week: The outages deserve to be considered together, without assuming they are connected. Email, air travel, HR systems and banking each support work we tend to take for granted until access disappears. Meanwhile, AI tools are taking on more of that work. The practical question runs through both stories: who owns the handoff, who checks the result, and what happens when the usual path is unavailable?

The pattern

  • Five disruptions in five days. No common cause has been established.
  • Check which critical processes depend on the same services.
  • Test whether your fallback still works when the primary service does not.
Microsoft 365 continuity badgeMicrosoft 365 / Business Continuity

Five failures in five days, and no common cause has been established

Sources: FAA — nationwide ground stop questions and answers; Ars Technica on Microsoft’s preliminary post-incident report; Techstrong IT on Workday’s statement; Associated Press on the JPMorgan Chase statement; CNN on the United Airlines outage.

Summary: Between Thursday November 12 and Monday November 16, five separate failures hit systems that a lot of organizations share. Microsoft 365 lost Outlook, Teams and Copilot for about a day. Microsoft’s preliminary post-incident report traces it to a latent defect in a July Windows fix for its network gateway servers; lock contention exhausted CPU and memory, and more than 80 percent of affected gateway servers stopped responding at the worst point. On Friday, the FAA halted all US departures for roughly four hours after a contractor’s router work at a site in Arkansas flooded its legacy telecom network, and the agency says it disabled the same broadcast function on more than 60 similar routers. The FAA explicitly attributes the failure to that legacy system and not to its modernization program. Workday’s HR and payroll platform went into maintenance Friday evening and was back by Saturday morning, after its disaster recovery site connected to live production databases during a maintenance validation. On Monday, JPMorgan Chase said a software bug in its own on-premises hardware disrupted Chase logins, transfers and some branches; customers at other banks saw delays only when transacting with Chase through Zelle, and Bank of America and US Bank said their systems were not directly affected. United separately lost the system that produces takeoff performance data for under an hour that evening.

All five operators said on the record that they found no indication of a cyberattack, and none has linked its failure to another. Timing is what puts them in one story: the Workday outage began inside the FAA ground stop, and the United outage began the same afternoon as the bank disruption. No source has established a connection between either pair, and the Department of Homeland Security said directly that the ground stop was not linked to the Microsoft outage. Treat the calendar as a coincidence of dates and read the responses as the pattern.

Why it matters

The pattern here is architecture, not coordination. Three of the five trace to change the operators described as routine: a fix deployed in July, a router swap at a remote site, a backup validation. What made each one consequential is that it sat at a single point of concentration, one company’s front door, one agency’s telecom network, one HR platform, one bank’s core. Your continuity plan probably assumes failures are independent, brief and local. This week tested the opposite: an email outage that also takes Teams and Copilot with it, a payroll platform that goes down on a Friday evening, a bank problem that surfaces as a failed transfer inside somebody else’s app. The useful question for your team is not whether this was an attack. It is which of your critical processes fail together, and whether your fallback shares a vendor with the thing it is backing up.

What to do now: Pick one critical process and trace every dependency to a named provider, including the fallback. Where two share a vendor, write down what you would actually do instead. Keep contacts, runbooks and the last approved version of key documents somewhere that does not require the same identity provider to open. Give one person the authority to declare a manual workaround. If your payroll or billing cycle crossed an outage, decide now how missed entries get captured and who checks for duplicates before resubmitting. And treat the incident number as phish bait: Microsoft Threat Intelligence reported fake service health notices quoting the real incident ID from this outage, so verify status inside your own admin center rather than through an emailed notice.
Microsoft 365 governance badgeMicrosoft 365 / Copilot Governance

Copilot can now change the system of record, and one click stops the asking

Source: Microsoft Copilot Blog — Turn Copilot insights into action with connector write capabilities

Summary: Microsoft made connector write actions generally available across Copilot Chat and Cowork on November 9. Connected tools can now create, update or delete information in the applications they reach, across hundreds of Microsoft, partner and custom connectors. Office app support is coming later, not available now. By default Copilot asks before it creates, updates or deletes anything, and the confirmation card names the tool and connector along with the details being sent. Per tool, a user chooses to allow once or to always allow, and always allowing removes the confirmation step for that specific tool from then on. Microsoft 365 admins can open each connector in the admin center to review its tools and identify which ones can change or delete data.

Why it matters

Connecting an application is no longer just a search decision. It can also authorize changes to the system of record. The approval screen is the visible part; the saved per-tool preference is what persists. Microsoft starts read tools at always allow and new write tools at needs approval. A standing approval for a delete-capable tool deserves a deliberate policy decision, not a reflexive click. The admin inventory helps identify what is actually available before a pilot expands.

What to do now: Pull the list of MCP-based connectors in your tenant and identify which ones can create, update or delete. Review saved tool approvals rather than assuming everything still prompts. Start with one low-risk connector and one-time approvals, and reserve always allow for read tools. Then write down who is allowed to change that preference.
OpenAI small business badgeOpenAI / Small Business Presence

OpenAI gave small businesses a profile to claim, not a ranking to buy

Source: OpenAI — Introducing Businesses in ChatGPT and the Business Agents policy

Summary: OpenAI’s Businesses in ChatGPT beta, announced November 10, adds a Business Hub where an eligible business can claim its organization, and once verified, correct key details and add self-service product feeds with pricing and availability. Claiming is open to eligible businesses wherever ChatGPT is available; existing enterprise customers and advertisers go through their OpenAI representative, and Shopify merchants can connect from Business Hub or the Shopify admin. Insights are aggregated and appear only above a volume threshold: mentions, website clicks and discovery topics. OpenAI says it does not share users’ conversations or agent chats with businesses. Business Agents, which are a separate step after claiming, are available to eligible ecommerce and online service businesses in the United States. OpenAI’s policy adds a qualifier worth reading twice: business verification alone does not guarantee agent eligibility, and the company may decline access or defer a category.

Why it matters

This is a data quality job with an owner, not a marketing campaign. The profile work is small and worth doing: if an assistant is describing your business to a customer, wrong hours, wrong pricing and a stale product list are now a customer service problem that nobody assigned to anyone. What it does not do is promise placement, recommendation or sales. Claiming and launching a customer facing agent are two different decisions with two different risk profiles, and treating them as one package is how an SMB ends up with an unsupervised bot speaking for the company. The verification gate also runs in both directions. Being eligible to claim is not the same as being approved to deploy an agent.

What to do now: Check whether your business is eligible and whether the information an assistant would surface about you is correct. Name one person who owns verification and the quarterly update. Treat any Business Agent as a bounded pilot with a named reviewer, and keep your existing support channel in place while you test.
Google Workspace automation badgeGoogle / Workspace Skills

A shared skill is a copy, not a subscription, and that is the catch

Sources: Google Workspace Updates — Skills in Google Workspace and Workspace Studio; Google Workspace Admin Help — Turn skills on or off

Summary: Google announced skills for Workspace on November 10: reusable custom instructions that carry team rules, templates and reference files. They can be built in Workspace Studio or edited as Markdown based SKILL.md files, drafted collaboratively in Google Docs, and tested before they are enabled. They run on the web in Workspace Studio, Gmail, Docs, Slides, Drive and Chat, plus the Google Chat mobile app by mention. Sheets and Meet are not supported yet, and prompts and instructions are English only at this point. Rapid Release domains started a gradual rollout on November 9; Scheduled Release domains begin November 30. The feature is on by default for eligible Business and Enterprise editions and can be disabled at the domain, organizational unit or group level. Sharing a skill gives each recipient an independent copy, and later edits by the author do not overwrite the copies people saved. Skills do not sync with Gemini app skills.

Why it matters

Consistency is the promise. Version drift is the bill. Every organization that has ever published a template knows this failure mode: the document gets shared, everyone edits their own copy, and six months later there are four approved processes. Google has reproduced that dynamic inside its AI tooling, deliberately and with real benefits, because a copy means a team can adapt a skill without waiting on the author. It also means the author loses control the moment it is shared. Referenced Drive files have to be shared with recipients separately, so a skill can be technically present and practically broken at the same time. The admin console logs creation, editing, sharing and running, which gives you a way to find the copies that matter, but only if someone is actually looking.

What to do now: Pilot one recurring workflow with one owner for the instructions. Before you share anything, confirm the referenced Drive files are accessible to the people who will run it. Decide how revisions reach the copies already in circulation, either by re-sharing a new version or by retiring the old one. Check the sharing controls and log events in the admin console rather than assuming the defaults match your policy.
Anthropic human review badgeAnthropic / Enterprise Defaults

Claude unified its workspace by default, and December 1 is the date to check

Source: Claude Help Center — Set up and manage the new unified Claude experience for your Enterprise organization

Summary: Anthropic’s admin documentation says the unified Claude experience, which folds chat and Cowork into one conversation, turned on by default on November 16 for Enterprise organizations where no admin had already set it. Every conversation starts as a chat; when a task needs code execution or files, it continues in an isolated cloud sandbox managed by Anthropic and returns the work for review. Owners and Primary Owners can turn it off in Organization settings under Capabilities, and if the setting had already been turned on or off, that choice stays. If chat or cloud code execution is off, nothing turns on. Organizations using zero data retention, the HIPAA ready configuration or customer managed encryption keys are excluded for now. The documentation also adds a forward date: on December 1, custom roles that include chat and code execution get the unified experience even if an admin turned it off for that role, unless the organization level setting was explicitly changed.

Why it matters

A capability switched itself on inside a paid workspace, and the settings that bound it are the ones nobody opens. Anthropic’s article names three separate approval switches: always allow for connector tools, a skip all approvals mode, and an automatically approve mode. Those are three different decisions with three different blast radii, and an organization that reads the first one and assumes it covers the rest has misread its own posture. Network egress from the sandbox is off by default for Enterprise, and connector policies are enforced outside the sandbox on every call, which are both reasonable defaults and neither is a substitute for reading them. The December 1 detail is the sharpest executive point in this issue: a role that was deliberately opted out comes back in, unless someone changed the organization level setting rather than the role.

What to do now: Open Organization settings and confirm the unified experience toggle, the role assignments, network egress and connector policies. Check all three approval settings separately: always allow for connector tools, skip all approvals, and automatically approve. Decide deliberately who should not have agentic execution and put them on a custom role. Then calendar December 1 and verify the custom role behavior in your own tenant rather than trusting this description of the documentation.
Hermes journey badgeField Notes from My AI Workstation
Hermes: My Agentic Journey

The site I asked for, and the bots I actually lean on

Storm built it. I reviewed it.

I gave Storm the instructions and source documents. Storm built the site. I reviewed it, corrected what did not fit, and directed the next pass.

This was real work handed off, not a request for instructions I would then have to follow myself. Storm built the SharePoint AI Resource Center, added catalogs and supporting pages, and used my six policy and framework documents to build a new guided front door. I reviewed the results and gave specific corrections: photographs in the actual page headers, more readable links, and organization-neutral language for the prototype. Storm made those changes, and I reviewed them again.

That is the human-in-the-loop part in practice: I set the direction and review what comes back; the agent does the building and the next pass. The result is a working prototype, with real policy approvals, support destinations and request automation still to settle before an employee launch. Those remaining decisions do not erase the accomplishment. They define what I need to approve next.

Grokbot: 22 bots I find invaluable

I now use 22 Grok bots and find them invaluable. There are more than three worth talking about, so this issue I am highlighting Friday, Cypher and Cart. We may feature different bots in future issues.

Friday is a concierge for my family. She keeps me apprised of birthdays, anniversaries and other important dates, and helps with meal planning and general family organization.

Cypher watches the network. When I installed smart home IP bulbs, Cypher flagged the new IP addresses immediately. That is exactly the job I wanted from it: tell me what appeared on my network, in time for me to look.

Cart handles shopping research when I need it, an especially useful role heading into the holiday season.

These are different kinds of help: keeping important family details in view, flagging a change on my network, and researching a purchase. The value is in a clear job that matters to me, not simply in having more bots.

A command center for my AI workstation

My workstation picked up a second screen this week: a 12.3-inch touchscreen with a 1920-by-720 display, mounted where I can glance at it instead of switching windows. The question I designed around was simple. What are my assistants doing right now, and what needs me?

I split the work the way I did the SharePoint build. Claude worked the design conversation and the interface concepts with me. Storm is doing the integration work behind the screen. I set the direction, sketched the layout I wanted, reviewed each pass and sent back specific corrections. The layout keeps a menu down the left side so I can look at one assistant or all of them at once, next to computer health: CPU, memory, GPU, disk, network and uptime.

I also wanted it to feel like my workspace rather than a vendor dashboard. That means a changeable logo and a configurable sci-fi background, either a still image or a subtle video loop, with the information staying readable on top of it. Decoration that competes with the numbers is a failed design.

These are three captures of the interface from this week, and the corner badge tells the honest part: mock data. The panels are shaped and laid out, but a shaped panel is not a live feed. The ChatGPT column is the one I would point to. It reads no verified integration, and it says nothing is inferred until a source is added in Configuration. That is the rule I want running in the build. A signal that is not connected gets labeled, not filled in with something that looks plausible.

Storm Ops overview dashboard on the touchscreen with agent cards for Storm, Central Command, Claude and ChatGPT, computer health tiles, and an awaiting-review list

First capture: each assistant in its own card with its state and review count, computer health in the middle, and the items waiting on me on the right. The mock data chip marks the sample values.

Storm Ops dashboard with four agent columns for Storm, Central Command, Claude and ChatGPT above a row of CPU, memory, GPU, disk, network and uptime gauges

Second capture: one column per assistant so Storm, Central Command, Claude and ChatGPT can be compared side by side, with the machine health strip running along the bottom of the screen.

Storm Ops dashboard with Storm as the primary assistant showing 4 of 7 tasks today, a critical disk alert, a machine health column and the left navigation menu

Third capture: Storm promoted to a primary assistant panel with a 4 of 7 task count, the alerts and machine health moved into the right column, and the same left menu driving navigation.

Right now, I’m working on optimizing token use to make the most out of all of my bots because they are so helpful.

What to do now: Assign a reviewer before you assign an agent. Write one sentence for each agent covering what it does, what it can change, and who acts on what it finds. If you cannot write that sentence, the agent is not ready to run unattended.
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

A capability that arrives by default is still a decision. Someone has to own it before it does something you have to explain.

Closing Thought

The outages do not need a common cause to deserve a common response: know what your work depends on and what you will do when it is unavailable. The AI stories ask a related question. As assistants move from answering to completing work, someone still needs to set the direction, review the result and decide what happens next. That is not a reason to avoid handing work off. It is how the handoff becomes useful.

Rob Niles illustrated portrait from the Hermes Journey

Hang tight…the weekend is almost here!

Rob Niles

AI Adoption Advisor | Microsoft 365 Strategist

Follow my AI journey

Need a practical AI plan for your team?

I help leadership teams, departments, and small businesses turn AI from scattered experiments into useful, governed workflows.

Talk with Rob about practical AI education

CincyAI: practical AI, local community

CincyAI community members gathered for a local event

Community photo from CincyAI.org.

CincyAI runs practical AI conversations and shares updates for the local business community. If you are in the Cincinnati area, it is a way to connect with others exploring what AI means for their work.

Sign up for CincyAI updates