What happens to a software business when its customers still use the product, but almost never open it?
That's where I've got to with Notion, which I use to keep knowledge, documents and databases. I've had it for years, but now almost never interact with it directly. Most of that work happens through ChatGPT and Claude. Notion still holds the material; I just don't need to go there to work with it, and I never use its AI.
For now, Notion still has a role in that arrangement. But as assistants acquire more capabilities of their own, I'm starting to question how much of the underlying software I'll continue to need. ChatGPT's move into documents, spreadsheets and presentations through Space makes that a fairly immediate question. I'll come back to that.
The attraction of working through one agent becomes clearer when you consider a complex project. Understanding where it stands means bringing together the programme, financial position, procurement and the decisions being made in meetings and emails. A project can look on track in one system while information in another tells a different story. Getting a useful picture means understanding how those things relate, not simply collecting a status report from each application.
Given the necessary access, a universal agent is the obvious place to do that work. It can follow the project across systems rather than require the person managing it to coordinate a separate assistant in each one. That leaves the software vendors with a difficult question: what would make the customer choose their assistant instead?
By universal agent, I mean a general-purpose assistant such as ChatGPT, Claude or Gemini, equipped with the tools to work across applications. My position is that these will ultimately win as the user's main interface unless the specialist vendors enable something those agents haven't been built to do. If Claude can use the same records and tools as the built-in assistant, having access to them is no longer a reason to choose the built-in one.
Making room for someone else's agent
For software vendors, this has already become a choice between accommodating the customer's agent (and ceding control) and risking being left out in the cold. They can restrict agent access, but that just creates friction in the customer relationships they are trying to protect.
Monday.com illustrates the awkwardness. It made a big pitch for native agents in its May AI Work Platform launch, while the same announcement included connections to outside assistants. Its Bring Your Own Agent beta lets external agents participate in the workspace. It is trying to sell the value of its on-platform agents while accommodating customers who would rather bring their own.
Salesforce's Claude integration lets Claude Enterprise users ask about accounts and pipeline using Salesforce data. The current beta is read-only. I see Salesforce trying to retain control of the underlying system while accepting that the customer may prefer to work elsewhere. Salesforce is a big beast with plenty of its own leverage, so it will be interesting to see how much of that control it can retain.
That doesn't mean specialist vendors can only respond by opening the door. Snowflake's Cortex Agents can use external tools, including Salesforce and Jira. They can reach out into other systems themselves. The contest becomes which agent can best handle the wider job, rather than which vendor happens to hold a particular piece of information.
Does the specialist keep its role?
Miro initially looked like a good reason to choose a specialist agent. Its AI can create editable diagrams in a visual workspace, where people can inspect and develop them. The value comes from the tools and interface around the model, often called the harness, rather than from a model with some exclusive ability to understand diagrams.
But Miro's MCP connection lets Claude create native, editable diagrams too. I haven't fully compared it with Miro's built-in agent, so I can't claim the capabilities are identical. Even so, what looked like a reason to adopt another assistant may instead be a useful tool for the one I've already chosen. Miro appears to be responding to the same pressure to let people bring their own agent.
In that arrangement, Miro becomes an app within an app. I can direct the work through my preferred assistant and use Miro's canvas when I need to work visually. The product still has a role, but its relationship with the user changes in much the same way Notion's has for me.
The next question is whether the application keeps that role. We started with ChatGPT playing nicely with Google's documents, presentations and spreadsheets. Now, with ChatGPT Space, those capabilities are being absorbed into ChatGPT itself. For work it can handle directly, there's no longer a need to use Google's tools underneath. What began as integration is becoming substitution.
Where does that evolution end, and which types of software retain their place? I'm not saying ChatGPT will replace every application. But connecting to a universal agent doesn't guarantee a lasting role underneath it. A customer may first stop opening the application, then discover they no longer need it for parts of their work. Then what? That's the existential challenge for software vendors.
What would make me stay?
I'd need a specialist experience that helps me do something my universal agent cannot do as effectively. To provide that, vendors need to rethink how people interact with their software. Most applications were designed around people finding information, filling in fields and moving work through a sequence of screens. If an agent can do much of that, we need to think harder about how value is created.
The person's role shifts towards explaining what they want, examining the work and making decisions. For a complex project, the software might assemble a live picture of progress, identify a conflict and let the person work through the consequences of different responses. The interface needs to support that way of working. That's what I mean by AI-native software, and it requires much more than putting a chat window beside the existing application.
Security is the other strong reason to stay. Microsoft is trying to offer a universal assistant within the security arrangements IT teams already trust. I understand the appeal, but "it's inside Microsoft" shouldn't end the discussion. Is the particular setup demonstrably safer, or simply familiar and already approved? Let's be clear: restricting your best and brightest to using Microsoft's AI tools is severely inhibiting their ability to do great work in 2026. My concern is that businesses are being asked to trade progress for safety without establishing how much safer the approved option actually is, or the extent of the trade-off they're making. That deserves an article of its own.
We already see the same tension in consumer markets. The discussion around Meta's Muse and ChatGPT's dots is part of the wider shift towards agents coming between brands and their customers, with commercial incentives of their own. I'll return to that separately.
For software vendors and consumer brands alike, the question is how to remain relevant and distinctive when someone else's agent handles more of the relationship. For users, once their preferred agent can access the systems they need and do the work effectively, the case for adopting a separate assistant in each application is getting harder to see.