Two people with the same AI model are like two restaurants with the same ingredients: you won’t get the same meal out of both. That’s where Jon Iskow and I started in Episode 2 of Kumo Coffee Talk. Jon is a ChatGPT guy. I’m a Claude guy. We still landed in the same place.
The short answer: Claude and ChatGPT each win at specific tasks, and you should use them for those. For business work, Microsoft 365 Copilot is the better choice. It works inside your own security boundary, it can see your company’s data the way you can, and it lets you use Claude and OpenAI models instead of making you pick one.
Below is the written version of the conversation, with links to each part of the episode.
Copilot vs ChatGPT vs Claude at a glance
Our take as of October 2026, from day-to-day use and client work:
| Microsoft 365 Copilot | ChatGPT | Claude | |
|---|---|---|---|
| Best at | Work grounded in your company’s email, meetings, files and data | Image generation, quick coding tasks, always-on personal assistance | Technical and business writing, long documents and data sets, long coding tasks |
| Access to your data | Signs in as you and sees what you can see, inside your Microsoft 365 tenant | Through connectors and extensions you set up | Through connectors and extensions you set up |
| Governance | Agent 365, tenant security, audit trail | ChatGPT for Work admin controls | Claude Enterprise admin controls |
| Model choice | Claude, OpenAI and Microsoft models | OpenAI models | Anthropic models |
| Where it fits | The default business AI tool for most Microsoft 365 companies | Personal productivity, creative and visual work | Deep writing, analysis and development work |
Claude vs ChatGPT: where each one wins
Where Claude shines
I moved my personal subscription from ChatGPT to Claude about a year in, and the reason was how structured it is. Where it stands out for me:
- Technical and business writing. Claude writes prose that doesn’t read like formulaic AI output. ChatGPT is a little better at general, casual writing.
- Long, structured inputs. Give it a long PDF or a data set and ask it to pull something out, and Claude does a better job.
- Artifacts. Claude can build a working page, chart or document next to the conversation.
- Long development tasks. For heavy, long-running development work, Claude Code is still the stronger option.
Its weak spot is illustrations. Because Claude takes an HTML approach, it’s great at site designs and slides, but for drawn images I’d go to OpenAI.
Where ChatGPT shines
Jon’s case for ChatGPT:
- Image generation inside the conversation. He uses it for every PowerPoint, plus process diagrams and workflow explanations.
- Codex for small and mid-sized development tasks. Hand Codex a bug or two from Azure DevOps and, in Jon’s experience, it finishes them faster and with less context than Claude Code.
- An always-on personal agent. It watches his email and calendar and acts on its own: it flags a meeting that clashes with school pickup and suggests a new time, works as a second spam filter, and tracks anything he asks it to watch. In his words, it’s like having a chief of staff without running your own server 24/7.
Always-on agents in a business
The always-on idea gets more interesting at work. A common example from our own projects: an Azure app registration secret that expires in a year or two, which nobody tracks until a client emails to ask why something is down. An agent that keeps an eye on dates like that is a real gain.
Microsoft is heading the same way with Copilot. Its agent that runs in its own virtual machine, opening browsers and on-premises systems, starts to look like a replacement for RPA tools such as UiPath and Power Automate Desktop. (We covered that shift in Power Automate vs UiPath in 2026.) Browser extensions already go a long way: with Copilot in Edge, or the ChatGPT and Claude extensions in Chrome, you can have an agent research leads from the CRM tab you already have open, then hand them to a second agent to qualify.
Claude vs Copilot and ChatGPT vs Copilot: why Copilot wins for business
The security reality check
Connecting a personal AI tool to business systems carries real risk. I once gave in to a user who kept asking to connect his own agent to our environment. His first request was “update my leads.” He had created every lead in that CRM during a migration, so the agent updated all of them, and we had to roll it back.
That’s the core difference. Copilot signs in as you. It works through your Microsoft account and sees only what you’re allowed to see, inside your tenant.
Governance is where Copilot pulls ahead
Jon’s point: with Agent 365, Microsoft’s governance for Copilot agents is the best on the market, and ChatGPT for Work or Claude Enterprise don’t match it. That matters more every month. As agents take more actions, and start to get their own identities, licenses and accounts, you need accountability: who did what, an audit log, and a way to change what an agent knows when it gets something wrong.
Keeping AI inside the trust boundary of your tenant, the Microsoft Graph and Work IQ, is what lets a business feel safe putting it to use.
Copilot isn’t a model, it’s the harness
A lot of people tried Copilot once, a couple of years ago, judged it “as a model,” and never came back. That misses what it is now. Copilot is the harness around the models. Inside Microsoft 365 Copilot and Copilot Studio you can choose Anthropic models such as Sonnet and Opus, OpenAI’s GPT models, or Microsoft’s own, which in some cases beat both.
To stretch the restaurant analogy: Claude, ChatGPT and the rest supply the ingredients. Copilot is the building, the service and the white tablecloths, and you decide what comes out of the kitchen and who gets to order it. So the “which is better” debate matters less than it seems. Use Claude where Claude is better and OpenAI where OpenAI is better, with one set of security and governance around both.
Copilot Cowork: why we rolled it out at Kumo
We touch on Copilot Cowork near the end of the episode, because it’s how we made this decision ourselves. Everyone at Kumo can use whatever personal AI they like. But once the work involves business data or customer data, we standardized on Copilot and Cowork, with licensing and spending policies set up for everyone.
Two things stand out so far:
- Better slides from people who hate PowerPoint. Presentations are cleaner and more polished, without anyone beating their head on the keyboard to get there.
- It knows your work context. I can type “Jon and I met with so-and-so yesterday, find the transcript” without naming the meeting. Cowork finds the meeting series and the latest session on its own. I’ve worked with human assistants who would have struggled with that.
Cost: use the right model for the right task
Copilot Studio is moving to credit-based consumption, including for agents that help you build agents. The organizations handling this best are setting a credit strategy now: budgets by role and department, with more spend where it pays off and less where it doesn’t.
The biggest lever is model selection (watch at 22:22):
- Don’t send every task to the smartest model. I leave GitHub Copilot on auto most of the time. You don’t need a frontier model like Fable to build a small HTML widget; Sonnet is fine. Use the bigger model for planning (security, architecture, solution design), then let a smaller model build the pieces.
- Give agents well-defined tasks. One vague sentence like “do this whole thing” burns through tokens. Agents are great at task-oriented work and lose focus when handed open-ended autonomy.
- Budgets can run away fast. One team Jon worked with had used ten times its GitHub Copilot budget only 10% of the way into the period.
A real example: cost vs. accuracy
On a current client project we’re extracting data from tax forms (Form 1040 and its variations) into a financial model. In discovery we compared Azure Document Intelligence, OpenAI’s GPT models and Anthropic’s models on cost, time and accuracy. I’m an Anthropic guy, but the GPT models won here: the more expensive options were only a fraction of a percent more accurate. We started with GPT, and once there are thousands of forms to learn from, we’ll look at building a custom model in Azure AI Foundry.
The lesson: don’t default to the newest, shiniest model. Every model has a cost-benefit trade-off, and Copilot lets you pick per task.
Which should your company choose?
- For business work on company data, choose Microsoft 365 Copilot. If you already run Microsoft 365, it’s the strongest option on security, governance, and access to your data in Microsoft 365, Azure and Dataverse. Jon’s estimate is that it’s the best business AI tool for 80–90% of American companies, and I agree.
- Use Claude and OpenAI models through Copilot and Copilot Studio for the tasks where they’re stronger, instead of running separate tools with separate data access.
- Let people keep personal tools for personal work, and draw a clear line at business and customer data.
- Set your credit and model strategy now. It’s the difference between AI that saves money and AI that blows the budget.
If you’re weighing Copilot Studio against Azure AI Foundry for your own agents, see Copilot Studio vs Azure Foundry vs built-in Copilots. And if you want help picking the use cases, models and guardrails, that’s what our Copilot and AI agent development team does every day.