Why Todoist says less AI can deliver more

Gonçalo Silva, CTO of Doist, admits he has no idea where AI is going next. Ask him for a roadmap and he'll tell you he could sketch ten different futures with roughly equal odds of being right. So instead of betting on a single model or vendor, Doist set three rules over a year ago: AI features must be purposeful, private and secure, and built to survive whatever model comes next. That framework is now showing up most clearly in Automations, a new Todoist service currently in closed beta, expected to ship in August or early September.

The mechanics are the interesting part. When a user describes what they want in plain language, an AI model translates that into a manifest — essentially a set of instructions. But once a trigger fires, ordinary code executes the steps, not the model itself. Silva's logic is blunt: a model is excellent at figuring out intent once, but asking it to behave identically the tenth or the hundredth time invites the exact unpredictability you don't want in automation. Running code on a CPU is also just cheaper and faster than calling a model on every execution, something Doist learned the hard way when an early beta with only a few hundred users generated a surprisingly large bill.

Silva calls the broader philosophy "subtraction over addition," and Todoist's recent history backs it up. The company killed its Goals beta on July 13 after deciding the feature, despite real investment, wasn't meeting its bar — a call that drew some user pushback even though Goals had never left beta. Before launching Ramble, its voice-to-task feature, a small team quietly prototyped and discarded at least 18 other AI ideas, including a project-summarization tool. Silva's read on that period is candid: Doist simply wasn't ready to build seriously with AI yet, even as it recognized the technology's power.

To avoid getting locked into one model, Doist runs extensive evaluations — Ramble alone is tested across roughly 18 languages with dozens of scenarios each — partly to catch behavior changes when swapping providers, and partly to hunt for cheaper models that still clear the same quality threshold. That same insistence on expertise shows up internally, too. Doist briefly experimented with letting product managers, designers, and engineers roam freely across each other's work, only to conclude that taste and specialized skill still matter. A product manager, Silva notes, generally can't distinguish a maintainable pull request from a fragile one that merely happens to work. An engineer can.

The pattern across all of it is consistent: AI shortens the distance between having an idea and building a prototype, but Doist is trying hard not to let that speed quietly lower the bar for what actually reaches users.