Almost every important thing your business knows lives in a table. Customers in rows, facts in columns — orders, payments, sign-ups, churn. It is the least glamorous data in the world, and it runs everything. So when a genuinely new kind of AI shows up that is built for tables rather than chat, it deserves more attention than it will get.
What Google just released
On 30 June, Google Research put out something called TabFM. Skip the name; here is what it does. You hand it a spreadsheet it has never seen before — your customers, your columns — and ask it to predict something: who is likely to churn, what a deal might be worth, which sign-ups will convert. It answers straight away. No training project. No data-science team tuning it for weeks. No “come back in a month.”
Until now, that was the hard part. Getting a prediction out of your own tables meant hiring specialists, building a model for your specific data, and waiting. TabFM reads your whole table at once — the examples and the question together — and just answers, the way a chat model reads a paragraph and continues it. For anyone sitting on rows and columns, that is a bigger deal than another chatbot.
Why this matters for ordinary businesses
The promise is simple: the kind of prediction that used to need a project might soon need an afternoon. “Which of these customers should we call first?” stops being a quarter-long data-science request and starts being a question you can ask of the data directly. That lowers the ladder for smaller teams enormously — you no longer need a research department to get value out of your own history.
The catch is the same catch as always
Here is the part a headline will not tell you, and the part that actually decides whether any of this helps you. A model that reads your table is only as good as the table. Feed it customers duplicated three times, half-empty columns, and out-of-date records, and it will predict confidently from nonsense. Clever new tools do not fix messy data — they just make its mistakes faster and more convincing.
So the exciting news lands on a familiar, calming conclusion. The frontier of AI is quietly moving toward the boring, structured data you already own — which means the work that pays off is the work you could start today, with no new model at all. Clean the duplicates. Fill the gaps. Get to one honest record per customer. Do that, and when tools like this become part of everyday software — and they will — your tables will be ready to answer, while everyone else is still cleaning up. The advantage was never the model. It was always the table.