AI moves into the ledger, the notebook, and the document
A single idea runs through today's announcements: AI is leaving the standalone chat window and moving into the exact places work already happens — the ledger, the notebook, and the document. The other half of the day is data on how companies actually buy that AI, and it points away from betting on one vendor.
If you are choosing tools this quarter, these two threads matter together. The buying data tells you not to over-commit, and the product news tells you where the committed capacity will show up first.
What the token data says about buying AI
Most of the figures you read about which AI model is "winning" come from a shaky base. They are built from consumer app usage, from developer routers, or from surveys where people report what they think they use. SaaStr's write-up of the Databricks AI Index makes exactly that point before offering a different lens: "Most model share data we get comes from consumer apps, developer routers, or surveys" [1]. The index instead looks at actual token traffic from more than 10,000 companies, which is a closer proxy for what firms genuinely run in production.
Three numbers from that report are worth carrying into a buying decision. Open-weight models have climbed to 31% of enterprise tokens. Seventy-two per cent of companies now use more than one model provider. And there is a 281x gap between the median user and the power user. Each of these tells you something practical.
The first says open-weight options are no longer a fringe choice; they now carry roughly a third of real workload. The second says that a multi-provider setup is now the norm rather than the exception, so any tool you buy that locks you to a single model family is working against where the market has already moved. The third — the 281x gap — is the quiet one. It means a small group of heavy users can drive most of your consumption and most of your bill. When you price an AI tool, do not model the average seat. Model the handful of power users, because they will dominate the invoice.
The lesson for a business choosing tools is not "pick the top model." It is "keep the ability to switch," and to plan budgets around a lumpy usage curve rather than a flat one.
The close moves into ChatGPT
The month-end close is the routine where a finance team reconciles accounts, checks that the books agree with reality, and signs off on the numbers for the period. It is repetitive, it is deadline-bound, and it is where small errors get expensive if they slip through. Xero's announcement today puts a slice of that work inside a place many people already sit: it is expanding its OpenAI integration to bring month-end close review to ChatGPT. As the company frames it, "The month-end close routine is one every accountant and bookkeeper knows by heart, and one Xero is helping you tackle with less friction" [2].
The shape of this change is what to watch, not the brand names. Instead of asking a person to open a separate reporting screen, the review step comes to the conversational tool. That lowers the activation cost of a check — you are more likely to run a review when it takes a sentence rather than a login. The trade-off is that the conversation becomes part of your financial control process, so the same rules you apply to any system that touches the books — who can see it, what it is allowed to change — now apply to the chat surface too.
If you are weighing how the money side of your stack fits together, our note on recurring billing and whether you need it covers the same instinct: automate the routine parts of finance, but keep a human sign-off on the edges.
Notebooks that build their own tools
Research notebooks have historically been passive. You collect sources, you read, you summarise. Today's Gemini Notebook update changes what the output can be. Google writes that "Gemini Notebook can now turn sources into interactive sites to help you quickly make sense of complex information" [3].
The concept worth understanding is the shift from a document you read to an artefact you use. When research becomes an interactive site, the reader can navigate it, filter it, and move through it at their own pace instead of scrolling a static page. For a team, that means the person who did the reading can hand over something explorable rather than a wall of notes. The caution is the usual one for generated artefacts: a tool built from your sources is only as trustworthy as those sources, so the review step does not disappear, it moves.
Structure for long documents
Long documents fall apart in predictable ways. A figure is referenced three pages before it appears. A "see table above" points at nothing after an edit. Nobody can find the one chart they remember. Google Docs addressed this class of problem today with three additions — inline captions, cross-references, and tables of captions. The post opens by placing the work in context: "Google Docs makes it easy to create, structure, and collaborate on complex documents across formats, including viewing and editing Microsoft Word (.docx) files directly in Docs" [4].
These features are unglamorous and that is the point. A cross-reference that updates itself when content moves removes a whole category of stale pointers. A table of captions gives a reader an index of every figure. For teams that produce reports, proposals, or anything long enough to get lost in, this is the plumbing that keeps a document honest as it changes. If keeping people aligned on shared material is your concern, the same thinking sits behind our piece on keeping a small team aligned without meetings: structure beats memory.
Dropbox buys into AI infrastructure
Finally, a corporate move rather than a feature. Dropbox completed its acquisition of Tigris Data, saying: "We're excited to welcome the Tigris team as we expand into AI infrastructure for companies and developers through DBX Agent Cloud" [5].
The concept here is that the layer beneath AI products — where data is stored and served to models — is now something the application vendors want to own rather than rent. For a buyer, an acquisition like this signals direction: a storage company is positioning itself as somewhere your AI workloads run, not just where your files sit. The trade-off to weigh is familiar whenever one vendor expands downward into the stack. More of your setup under one roof can mean fewer moving parts, and it can also mean more of your dependence concentrated in one place. Given the multi-provider reality the Databricks data described earlier, that concentration is exactly the thing to keep an eye on.
The thread, pulled tight
Put the day together and the message is steady. AI is becoming a feature of the tools you already run, not a separate destination you visit. Meanwhile the hard numbers say companies are spreading their bets across providers and that a few heavy users drive most of the cost. Choose tools that let you keep that flexibility, price them against your power users, and keep a human in the loop wherever the output touches money or the record. For the shorter version of yesterday's moves, see our briefing for 17 November.
Sources
- [1] 5 Interesting Learnings from the Databricks AI Index: Open-Weight Jumps to 31% of Enterprise Tokens, 72% of Companies Now Use Multiple Model Providers, and a 281x Gap Between Median and Power Users — SaaStr
- [2] Xero expands OpenAI integration, Bringing Month-End Close Review to ChatGPT for Accounting — Xero
- [3] Transform your notes and research into dynamic tools with Gemini Notebook — Google Workspace
- [4] Add inline captions, cross-references, and tables of captions in Google Docs — Google Workspace
- [5] Dropbox completes acquisition of Tigris Data — Dropbox