How a customer used Cowork to plan a Google Workspace to Microsoft 365 migration
- Arthur Igloo
- May 12
- 11 min read
I had a really good morning with a customer last week and I wanted to write it up while it was still fresh.
The customer is Jeff. He's the IT guy at a regional market in New Mexico — north of fifty users, multiple locations, the whole shop running on Google Workspace since the early Google-was-free-for-business days. Jeff is one of those one-person IT departments that you find in a lot of SMBs. He keeps the lights on, he handles tickets when the staff calls him (literally calls him — they don't email tickets), he resets passwords twice a year, and he picks up whatever project the owners drop on his desk. Right now the project on his desk is a doozy. They want him to move the entire company off Google Workspace and onto Microsoft 365.
And while he's at it, change the company's email domain from t-rmarket.com to trmarket.com — drop the dash, rebrand on the way out.
And before that, sometime in the last couple of years, somebody needed Publisher to make a gift certificate. Microsoft killed Publisher in the regular Office subscription, so Jeff had to spin up a partial Microsoft 365 account in a hurry, get one user a license, and move on. He didn't have time to set it up the right way. He told me on the phone, "I didn't do everything correctly. I wanna start over."
So that's the project. A real Google → Microsoft 365 migration, with a simultaneous domain change, on top of a half-baked existing tenant that needs to be undone first. For a hundred-ish users. Run by one IT guy who has plenty of other things to do this week, including security month — they reset every password in the company in April and October.
Jeff's honest take, before we started: "I don't know how to do all of that."
This is exactly the kind of project that AI is unbelievably good at right now. Not because it can do the migration for him. It can't. There are real network changes, real DNS publishes, real licensing decisions that have to be made by a human with skin in the game. AI is good at this because it can be the deeply prepared partner that Jeff doesn't have anywhere else. So that's what we set up.
Why this is a Cowork project, not a ChatGPT project
One of the things I find myself explaining over and over is that there's no single "use AI" answer. There are different shapes of AI for different shapes of problems, and picking the wrong shape will burn you out faster than not using AI at all.
For Jeff's project, the wrong shapes were obvious. He didn't want an autonomous agent — there's nothing to automate yet, and you'd never want to autopilot a DNS cutover anyway. And he didn't want to use AI as a glorified search engine, which is how most people start. Glorified-search works fine when you have one question. Jeff has hundreds of questions, and they all build on each other. Question forty depends on the answer to question seven. Question seventy-eight depends on his current Google licensing, which he told the AI back at question twelve. He needs something that remembers the project.
That's Cowork. Cowork is a desktop AI that works alongside you, holds the entire project in its head, can be made an expert in whatever you point it at, and lets you stay in the driver's seat for every actual decision. It's twenty bucks a month at the entry tier. The tool we used was Anthropic's Claude desktop, which is what I run for all of my own Cowork work.
If you've never set one of these up, the rhythm goes like this:
Install Claude desktop on your machine.
Make a folder on your desktop called something like Cowork. Inside that folder, make one subfolder per project. For Jeff, we made M365 Migration.
Open Claude desktop, point it at that project subfolder. From now on, anything Cowork builds for this project lives in that folder.
Turn on the right skills. Cowork has a "skill creator" skill that lets it build specialized expertise on the fly. We turned that on so it could build itself an expert skill for Microsoft 365 migrations, then save that skill back to the project folder so it would survive future sessions.
Allow it to access the network so it can crawl reference material.
That's about ten minutes of setup. The first thing we asked it to do, after the housekeeping, was three sentences long: I'm an IT manager. I need to migrate from Google Workspace to Microsoft 365, and I'm changing my domain in the middle of it. Become an expert on this whole process — I'm going to give you reference material, then we're going to plan the migration together.
Then we handed it some Microsoft Learn pages as a starting point and told it to crawl out from there.
What it produced on the first pass
About fifteen minutes later, Cowork had written Jeff a Word document called Google Workspace to Microsoft 365 Migration Playbook. It was sitting in his project folder. He opened it, and it was real. Sectioned, written like a runbook, technically dense, specific to his situation. Not a generic AI summary — an actual plan you could hand to a junior tech and say "follow this."
That alone would have been a decent output. But the real value showed up on the second pass.
I asked Cowork: Now go back out, crawl the web on this same topic, and write me a companion document of additional considerations that aren't covered in the playbook. And then I gave it Jeff's specific situation: the legacy domain, the new domain, the half-baked M365 tenant, the fact that they have shared PCs in the back office for the staff who don't have their own desks.
What came back was a fifteen-section companion document that I would, in good conscience, hand to a paying consulting client. Highlights:
From the companion doc
The domain change is the biggest risk factor
Most migration guides assume the same domain follows the user from Google to Microsoft. Jeff is doing both at once. So the doc called out: provision every M365 mailbox with the new @trmarket.com as primary SMTP and the legacy @t-rmarket.com as a secondary alias. Outbound mail rebrands, inbound to the old address keeps working, internal directory lookups don't break. It also flagged that you can't have a domain verified in both Google and Microsoft simultaneously, so the unverify-from-Google → verify-in-Microsoft handoff has to be sequenced carefully.
From the companion doc
A practical DNS cutover sequence
T-minus seven days, drop your TTLs at the registrar to 300 seconds so you have fast rollback. T-minus three days, build the new MX/SPF/DKIM/autodiscover/DMARC records but don't publish them yet. T-minus one day, pilot with three to five users. Cutover day, publish, validate from Gmail and iCloud and a real vendor inside thirty minutes, watch Microsoft Message Trace for four hours. T-plus seven days move DMARC from p=none to p=quarantine. T-plus thirty days, p=reject. With the explicit warning: never leave both Google and Microsoft MX records published at equal priority, your mail will load-balance randomly between the two systems and good luck debugging that on a Tuesday afternoon.
From the companion doc
Licensing for a 25–150 user company
Real recommendations, not just a price list. IT, finance, exec — Business Premium ($22) for the device management and conditional access. Office staff — Business Standard. Shared mailboxes (info@, sales@, support@) — unlicensed up to 50GB. Resource mailboxes (conference rooms) — unlicensed. And it flagged the trap: every Business SKU caps at 300 seats per tenant, and the jump from Business Premium to E3 is a 77% per-seat increase. If you're growing, plan that trigger now. It also warned about Basic Auth SMTP being deprecated, which matters because Jeff has copiers and scanners that send mail with username/password, and they'll all stop working on cutover day if nobody plans for it.
It also covered SharePoint vs OneDrive structure (don't migrate Google Drive verbatim, you'll end up with users who have 800GB OneDrives full of stuff that should have been in SharePoint). What does and doesn't migrate from Google Drive without a third-party tool. Calendar gotchas — only the primary calendar comes over, secondary calendars and conference room bookings don't, mail rules arrive disabled and every user has to re-enable them. Identity decisions — cloud-only, hybrid, or federation, and which one is right at fifty to a hundred and fifty users. Endpoint stuff — uninstall the Google Drive client before installing OneDrive sync or you'll corrupt the cache. The third-party app inventory that nobody ever does and that always causes a post-migration scramble. Customer-facing communications timeline. Backup and compliance. A pilot/wave/cutover plan. A day-one survival kit for end users. And a top-ten list of things that bite people on these migrations, with mail rules silently arriving disabled at number one.
Fifteen sections. Specific. Actionable. Tailored to Jeff's company size and the domain change.
That's the document I would have charged a customer real money to produce, six months ago, by hand, over the course of a week.
What Jeff was actually doing while Cowork worked
This is the part of these stories that gets lost when people write up AI use cases in the abstract. Jeff was at his desk. He was Splashtopped into me so I could click around in his Cowork. His phone was ringing every few minutes — his receptionist needed help with the password to a vendor portal, somebody else needed something else. He was finishing up password resets for security month between Cowork prompts. He was simultaneously poking at Microsoft's pricing pages because he'd realized halfway through that the company had a separate question about device-based licensing for the shared PCs in the back office, which was a real wrinkle the playbook hadn't covered yet.
And every time something came up — the shared-PC question, the Publisher backstory, "we have copiers that send mail" — Jeff just told Cowork. And Cowork rolled it into the plan.
That's the thing I keep trying to explain to people about why Cowork is different from prompting ChatGPT. ChatGPT will give you a great answer to your question. Cowork is doing the project with you. The state of the project, the open questions, the previous answers, the corrections, the gotchas you mentioned in passing forty minutes ago — it's all there.
At one point we ran out of tokens and had to stop for an hour. (You hit usage limits when Cowork is doing heavy work like crawling and writing. Once it knows the project, normal back-and-forth is much cheaper.) When we picked back up an hour later, Jeff didn't have to explain anything. The project was where we left it. We just kept going.
The CLAUDE.md trick
One thing worth flagging because this is where Cowork starts to feel like a real coworker and not a tool: before we wrapped up, I asked Cowork to write itself a project memo. In Cowork-speak, this is a file called CLAUDE.md that lives in the project folder and gets read every time the project is opened. Think of it as the agent introducing itself to its future self.
What it wrote into that file was, in Jeff's own situation: You are an expert assistant on Google Workspace to Microsoft 365 migrations. The customer is migrating from t-rmarket.com on Google to trmarket.com on Microsoft. Existing tenant exists with two licenses and is misconfigured — plan a clean rebuild. Org size is 25–150 users. Multiple shared PCs, device-based licensing under consideration. Reference docs in the working folder include the migration playbook and the companion considerations document. Open decisions: final SKU mix, third-party backup product, Intune scope, customer communication ownership, voicemail-to-email impact.
So the next time Jeff opens the project folder — Tuesday morning, three weeks from now, whenever — Cowork will read that file first and pick up exactly where it left off. He doesn't have to remember what they discussed. The project remembers.
This sounds small. It is not small. The reason most consulting projects lose momentum is that the context gets stale between meetings. Whoever was up to speed leaves, the runbook becomes "ask Dave," and Dave eventually forgets. Project memory is the unsung hero here.
What this is actually saving Jeff
If Jeff had tried to plan this migration the old-fashioned way, here's what was on his plate:
Read four hundred pages of Microsoft Learn documentation. (He has the Office 365 Administration Inside and Out book on his desk. He told me he was trying to find a PDF version because flipping pages while doing the work is a pain.)
Buy a couple of consulting hours from a Microsoft partner to sanity-check the plan. Two grand if he was lucky.
Write the runbook himself.
Build the third-party app inventory himself. Forget half of it. Discover the rest at 11pm on cutover day.
Hope.
What he actually has now is a real plan, written in his voice, specific to his company, with the gotchas called out, with open decisions clearly listed, with a project memory that survives across days. Total spend: twenty bucks a month for Cowork, plus the hour he and I spent setting it up together.
And here's the part that's not just about cost. Jeff is going to be a smarter operator at the end of this project than he was at the start. He's going to read that playbook a dozen times. He's going to ask Cowork follow-up questions. He's going to push back on things that don't match his environment, and Cowork will revise. By cutover day he's going to know this stuff like he wrote it. Because in the relevant sense, he kind of did.
Where this fits in the bigger AI picture
If you read my piece from a few weeks ago, you know I think about AI in three shapes inside an SMB. Agentic workflow processes (like our AI receptionist Arthur, or our level-one tech Alex). Desktop AI working alongside you (Cowork). And AI personal assistants with memory.
Jeff's project is the textbook example of the second shape. There was no good way to handle it with the first or the third. It needed a partner who could become a real expert on a specific problem, hold all the context for as long as the project ran, and let the human stay in charge of every consequential decision. That's Cowork's whole reason to exist.
I'd guess that for most IT-team-of-one operations in SMBs, this is going to be the highest-leverage AI use case for the next year. Not flashy. Not autonomous. Just a partner who showed up the day you needed deep expertise on something you don't have time to master. The migration is one example. The same pattern works for: a security audit. A compliance project. A new ERP rollout. A vendor consolidation. Any project where you'd hire a consultant if you had the budget and don't.
How to start, if this resonates
The honest answer is, install Claude desktop, point it at a folder on your computer, give it a real project, and see what happens. The cost is low and the only way to learn the rhythm is to do it. The Jane example I write about in the six-months piece — taking her two-day monthly invoicing job down to fifteen minutes — that started exactly this way. So did the legal letter to my contractor that referenced specific Washington State construction code. So did Jeff's migration plan. None of those projects had a polished how-to in front of them. We just opened the tool and started.
If you're a VoIP PBX Express customer or you're working with us on Arthur Igloo, you're already in the "free guided AI assistant deployment" tier of our service — that means we'll sit down with you and help you set this up. No charge. Same way we sat with Jeff. If you're a small or mid-sized business outside our customer base and you want a hand getting started with this kind of thing, that's exactly what Arthur Igloo's consulting arm is being built to do. Practical AI deployed by people who already understand how a small business runs.
Either way, the real point of this story isn't the migration. It's that one IT guy in a market in New Mexico just turned a project that was bigger than him into a project he can finish. That's what good AI looks like in 2026. Not magic. Not a science-fiction takeover. Just a really well-prepared partner showing up on the day you needed one.
Want to try this on your own project?
Pick something you've been putting off because you don't have the expertise to get started — a migration, a compliance audit, a vendor evaluation, a system you've been meaning to learn. Open Claude desktop. Give it a folder. Start the conversation. If you'd like a hand, give us a call at (360) 515-2700 or email info@arthurigloo.com.