Moacir Moda challenges the pursuit of perfect AI understanding, proposing instead a validation-centric approach that balances machine efficiency with human oversight, informed by his experiences with meeting summarisation tools and knowledge management systems.
Moacir Moda argues that the real problem with artificial intelligence is not that it is useless, but that it is often trusted to understand too much. In his newsletter, he describes how a flawed meeting summary pushed him to rethink the relationship between human judgement and machine output. Instead of trying to make the system perfectly understand every nuance of his work and personal life, he concluded that the safer approach was to design a process in which AI could be useful while still being allowed to make mistakes.
That change in thinking emerged from his time working closely on product and engineering problems at Tintim, where he says he spent months revisiting older programming habits and speaking frequently with technical adviser Henrique Bastos. Moda presents software development as the cutting edge of practical AI adoption, with new tools, models and development methods appearing so quickly that even experienced teams are still testing what works. From that environment, he says he borrowed a key lesson for his own work as a chief executive: systems should be built around validation, not blind automation.
His own experiment began with a search for a “second brain”, a popular idea in note-taking and knowledge management. He tried existing repositories and built his own, but found those attempts failed until he stopped trying to automate everything at once and focused on the underlying process first. The first useful gains were simple and mechanical: downloading transcripts, organising files and preserving source material. The harder problem was interpretation. Google Meet’s own note-taking tools can already generate summaries and action items, according to Google’s support pages, but Moda says that kind of output can still misread context, such as treating casual pre-meeting conversation as if it were part of the decision-making process. Google says its “Take notes for me” feature stores notes in Google Docs and works only in certain Workspace editions or Google AI plans, while later coverage from Tom’s Guide noted that the feature was expanded to in-person meetings as well.
Moda’s response was to split the workflow into separate stages: ingestion, suggestion and processing. In his description, the AI does the reading and drafting, but it cannot change anything without approval. That structure matters because the factual record remains untouched while the machine produces proposed edits, each tied to evidence that explains why the suggestion exists. Only after review does he accept or reject the proposal. Rejected items are discarded; accepted ones are applied. He says the point is not to eliminate human effort entirely, but to remove the repetitive work that has low value while keeping control in human hands. Guides from notes.so, Workalizer, Notexapp and RecordMeeting all point to a similar gap in Google Meet’s native tools: they can capture text and produce summaries, but they do not always provide the depth of context, consistency or cross-platform flexibility that teams need.
He also structures the system around preserving history. Notes can be revised, but they are not deleted, because the full record is kept in Git. That allows the system to track changes over time and makes it easier for AI to recover context later. He divides his workspace into practical categories, from source, front, project and task to event, artefact and proposal, each serving a different role in turning raw information into something actionable. In one example, an all-hands meeting produced two facts: a sales comment about a recently launched feature and a marketing confirmation about a costume appearance at an event. The AI turned both into suggestions, but Moda accepted one and rejected the other, not because the second was false, but because it was not important enough for his purposes. That, he argues, is the real challenge with AI in day-to-day work: not perfect understanding, but a system that makes errors visible, contains them and still speeds up the work that matters.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
