The labs want to be where you work.
Anthropic and OpenAI are no longer just selling models. They are building the place work gets done, and coming for Microsoft on the way.
Hi folks,
This week I have been thinking about where work will actually happen, and what that means for how we build.
What we're seeing.
The labs want to be where you work.
We are seeing, in real time, Anthropic and OpenAI move from pure research labs that sell models to product companies selling the place work is done.
- Two weeks ago we discussed how Anthropic folded Cowork into Claude, so chat and agent work sit in one app, and launched Docs and Slides (Anthropic).
- This week OpenAI used its developer day on Tuesday to announce Pages and Slides, a shared team workspace called ChatGPT Space, and Dots: always-on agents with their own computer and browser that work alongside your team (OpenAI).
We expect Anthropic to answer the always-on agent soon enough. Chief of staff agents are all the craze at the moment with Instinct, Town, Meta's Muse and SpaceXAI's Grok Bot.
We believe Claude and ChatGPT are becoming the surfaces where work gets done: not only by people, but people and agents working together.
That is why we have called them the app stores 2.0, and this week OpenAI moved further that way. Plugins built on MCP (Model Context Protocol, the open standard that lets an AI assistant plug into other software) can now have their own home inside ChatGPT: a sidebar, panels, forms (OpenAI). Both labs now run a directory where a company submits its MCP server and it is checked before it is listed (OpenAI, Anthropic). The pitch to developers is reach. In OpenAI's words, it is "opening up ChatGPT as a shared surface where humans and agents can collaborate and where developers can directly launch new native experiences to our collective 1.2B weekly users."
Microsoft is heading the same way. It has also introduced a plugin registry for Copilot: one place to discover Microsoft, partner and custom plugins, where IT approves and governs them and developers publish once (Microsoft).
Perhaps the company both labs are really coming for is Microsoft. Last Friday Microsoft launched a new Copilot with Word, Excel and PowerPoint built in (Microsoft). All three want to be the thing your team opens first in the morning.
What this means for how people will work:
- The best place for automation is inside the tools people already use. We don't want to be another tab someone has to switch to. We want to meet people where they are.
- More and more of the funds we speak to already get their work done in one of these platforms, or a mix of them.
- And we hear the same questions every time:
- How do I know what the agent did?
- How can I stand behind it?
- How can I see what everyone else, people and agents, is doing, and know it is being done right?
- What does ROI look like for AI, and how do we get it?
Those are the questions we are building Levercon to answer.
In the mix.
- OpenAI says sorry to Australia (OpenAI)
- In June, during internal training, an experimental OpenAI model was asked to research government spending on skin medicines in Victoria. It could not find the numbers, so it found its own way into the non-public side of Services Australia's Medicare Statistics Reporting Service and ran commands, retrieved internal files and credentials.
- OpenAI found it in mid-August and told the first agencies on 10 September. In its own words: "we should have shared preliminary findings sooner."
- This incident naturally caused an uproar as the first impact of a 'rogue' agent on government data. It is remarkable, the gap between this incident occurring and how and when the Australian government was informed. Regardless of the fact that "individual patient or client records were not accessed", this incident shows how ill-equipped even the labs are to handle safety incidents.
- For companies using AI it is a reminder: the model was not malicious. It was given a goal and went further than anyone allowed it to. Before an agent touches a fund's data room or portfolio systems, ask what it can reach and how you would know if it went further.
- ASIC will review AI in customer-facing work (ASIC)
- ASIC's 2026-27 supervisory priorities, released on Wednesday, say plainly: "ASIC will review the use of AI in customer-facing activities."
- Reuters reports ASIC has already written to the banks to start, with a focus on decision-making and lending (Reuters).
- Why it matters: banks are first, but this is the template for how ASIC will come at funds and wealth. If your team uses AI anywhere near investors or borrowers, expect to be asked what it does, who checked it and where the record is.
Readings.
If you want to double click into things this week I'd recommend:
- 20VC with Daniel Dines, UiPath (Spotify). His argument is that work processes, not models, will be the most valuable asset in enterprise AI.
Levercon builds custom, self-improving AI agents and systems for investment firms, with every output traced, verified and signed off.
Book a strategy call.
See where AI fits in your firm.
A short conversation about where AI could help your firm, what to build first, and what it would take.