She is looking for a voice memo she recorded three months ago about pricing the new offer. She remembers the gist, a comment about the third tier, something about a competitor’s structure. She does not remember the date, the title, or whether it was before or after the trip to Manchester. Her phone shows three hundred and forty-one unlabelled audio files, sorted by date. She scrolls. She gives up. She will rebuild the thinking from scratch.
This happens in every owner-managed firm I have worked with, and the diagnosis is almost never “you should be tidier”. The capture system worked. The retrieval system did not. And the response a generation of productivity writing reached for, build a tidier folder tree, classify everything at capture time, learn Notion properly, is exactly the wrong fix.
This is the front-of-the-Automate-quadrant post in the AI for your own work cluster, sitting alongside auto-summarising every meeting and the AI calendar-prep one-page brief. It is the post that says, before you pick a tool, change the order of operations.
What is the founder’s AI filing system?
The founder’s AI filing system is a three-step pipe. Capture without classifying, let AI transcribe and tag retrospectively, retrieve via search rather than by remembering where you filed it. Voice memos, files, forwarded emails, and meeting transcripts all land in one inbox. A model reads each one, applies a small set of tags, and indexes the contents. You ask the system questions. The folder tree disappears, because you no longer need it.
The shape borrows from the EAD-Do framework recast for AI Automate quadrant, the work that should run with no founder hand on the wheel. Apple’s on-device Voice Memos transcription, Otter.ai for live meeting capture, and Mem.ai or Notion AI for tagging cover the practical stack. The change that matters is the order, capture first, classify later, search always.
Why does the tidy folder approach keep collapsing?
Because tidy folders impose a decision at the worst possible cognitive moment, the second after you had the thought. Tiago Forte’s PARA framework asks you to classify each note as a Project, Area, Resource, or Archive item before you save it. In a client meeting, on a school run, between two calls, that decision costs more attention than the note is worth. The note does not get filed. Within three months the system is dead.
Sönke Ahrens, in How to Take Smart Notes, argued the same point about academic writing more than a decade ago. The friction of classification at capture time is what kills the practice. Niklas Luhmann’s Zettelkasten worked not because the slips were perfectly classified but because they were densely linked and indexed, the entry-points emerged from the corpus rather than being imposed on it. AI is the cheap version of that mechanism, applied to a founder’s working week.
Where do you actually meet this in your week?
You meet it four places, and the same logic applies to each. The voice memo recorded between meetings, the file dropped into a Drive inbox, the meeting transcript captured by Otter or Granola, the email forwarded into a workspace. Each of these is a frictionless capture event. The model reads the content, summarises in a paragraph, and applies four tags. Client or stakeholder, topic, decision-status, follow-up flag. That is the entire taxonomy.
The four-tag rule is the one piece of discipline the system asks for, and it is deliberately small. Bigger taxonomies degrade because tagging discipline drifts under operational pressure, the AI starts inferring tags inconsistently because the founder has fed it inconsistent examples. Four tags is enough to answer the questions a founder actually asks, what have I decided with this client, what follow-ups am I behind on, what topics have come up more than three times this quarter. Beyond that the noise outweighs the signal.
A worked example. A founder records a five-minute memo after a Tuesday client call. The model transcribes it, summarises in eighty words, and applies tags: Client-Acme, Topic-Scope, Decision-Pending, Follow-Up-Pending. Three months later, preparing for a renewal conversation with the same client, she searches “Client-Acme AND Decision-Pending” and surfaces every open question across all formats, voice, email, meeting transcript, file. The act of searching is what replaced the act of filing.
When does this approach earn its slot, and when should you ignore it?
It earns its slot when retrieval has become the bottleneck, when you can feel that the thinking is in there somewhere but you cannot get to it cheaply. Owner-managed firms with high context-load per client, professional services, agencies, consultancies, advisory practices, are where the payoff is largest. The compounding effect over twelve to eighteen months is real, the searchable archive becomes institutional memory that survives staff turnover and onboards new team members faster.
Ignore it in three cases. If you are already running a tight Notion or Obsidian system that you actually maintain, do not replace working machinery with a different working machinery. If your work is contractually bound by data-sovereignty rules that cloud transcription would breach, run on-device only or wait until your platform of choice supports local processing. And if your week is genuinely simple, fewer than ten clients, predictable rhythm, the cost of setting up the system exceeds the friction it removes. Honest assessment beats tooling enthusiasm here.
There is also a UK GDPR layer that owner-operators commonly under-think. The moment you record a client’s voice, you are processing personal data. Get explicit consent at the start of any recorded call, the line is one sentence and the client almost always says yes. Choose a transcription vendor whose data processing agreement says they do not train models on your content, Anthropic’s Claude for Work and Notion AI for Business both qualify, many consumer tiers do not. Set a quarterly deletion routine before the recordings accumulate to the point where deleting them is a project of its own.
What sits next to this in your toolkit?
Three companion practices make the filing system sturdier. The auto-summarising meeting habit, where every internal call produces a tagged transcript, is the largest feeder into the archive. The AI calendar-prep one-page brief, where the model assembles the morning’s context from prior tagged notes, closes the loop back to the next live conversation. And a standing personal prompt library keeps the tagging consistent across weeks, because the same prompt produces the same output shape.
Beyond these three, two adjacent ideas are worth holding lightly. Semantic search, the kind that understands “what did the client say about timeline in March” without you having to remember the exact words, is now a default feature in Mem, Notion AI, and Obsidian’s Smart Connections plugin. And the four-tag taxonomy is itself revisable, you may find after a quarter that decision-status and follow-up collapse usefully into a single status field, or that topic needs to split between domain and function. The taxonomy is a working tool, not a fixed architecture, and the right tweaks emerge from the searches you actually run.
The point worth ending on is the one the brief opened with. Stop trying to be Notion-perfect. The system is what does the filing, and the founder’s job is the work the system cannot do, the judgement, the relationship, the call. If you would like a hand thinking through which tools fit your firm and where the GDPR seams sit, book a conversation.
