Last month, I tried to build five things.
- A Slack automation that read, triaged, and drafted responses to messages across my workspace.
- A tax assistant that retrieved my forms and pre-filled the relevant fields.
- A LinkedIn pipeline for my own personal brand.
- A Facebook Marketplace scanner that flagged listings matching my criteria.
- A Google Photos cleaner that freed up storage.
None of them shipped. Not because the AI couldn't do the work. It could. Because I couldn't get the AI to the data.
Five projects. One month. The same wall.
I don't think it stays this way. The use case is too obvious and the capability is spreading too fast.
The assumption underneath everything
The APIs, permission models, and Terms of Service that govern modern software were designed around one assumption: that the entity requesting programmatic access is an organisation building a product for other people. A company goes through a verification process, signs terms of service, gets approved, and gets access. That model made sense when "building on top of a platform" meant shipping to thousands of users.
It doesn't account for what I was doing. One user. My own data. No distribution. Just trying to make my own tools work.
The personal builder, automating their own life on their own data for their own use, simply wasn't in the design space at all.
Once I saw that, the five failures sorted themselves into three distinct tiers.
Fig. 1 — Three tiers. One root cause: built for orgs, not people.
Tier 1: No API, and automation is a ToS violation
Facebook Marketplace and LinkedIn both live here. No public API exists for personal use, and both platforms explicitly prohibit automated access in their Terms of Service. Meta's terms state you may not "access or collect data from our Products using automated means." LinkedIn's User Agreement bans "bots or other automated methods to access the Services." Tools that automate LinkedIn exist in a grey zone, but LinkedIn reported blocking 97.1% of fake and automated accounts in the first half of 2025 before they were even flagged. Account restrictions and permanent bans are real outcomes.
There's no compliant path here. The platforms simply decided personal programmatic access isn't a use case they support.
Tier 2: API exists, but the ownership model blocks you
Slack lives here. I wanted to automate reading, triaging, and drafting responses to messages across my workspace, including writing back. The moment you need to update or post messages on behalf of a user, bot-scoped tokens aren't enough. Bot tokens can only act in channels they've been explicitly invited to, and they can't post as a human. To read and respond across the full workspace, an app needs user-scoped permissions installed org-wide. That requires workspace admin approval and exposes the app's permission scope to the entire organisation: a significant barrier for what is, at its core, a personal tool.
Slack's Terms of Service clarify the deeper issue: it's a workplace tool intended for businesses, not personal use. The legal Customer is your employer, not you. Even if you navigate the permissions, you're building personal tooling on top of data access rights you don't actually own.
Tier 3: API exists, but the capability was removed or was never there
Google Photos is the sharpest example. I wanted to free up storage: find duplicates, delete old screenshots, reorganise. The Google Photos API exists, but it never supported deletion of items from a user's existing library. Then, as of March 2025, Google removed the scopes that let any third-party app read your existing library at all. The Library API now only manages content your own app uploaded. A personal storage tool has no supported API path, and never really did.
The tax assistant sits in the same tier, different mechanism. I wanted to retrieve my tax forms and have AI pre-fill the relevant fields: the kind of work that takes hours and is almost entirely mechanical. No API was ever exposed. Browser automation was the only path: not calling an endpoint, just impersonating a user clicking through a UI that can change any day. I ran it twice. It broke on the third try.
What gen AI changes about this
For most of the last decade, these friction points were annoying but survivable. If you couldn't automate something, you did it manually.
Gen AI shifts that calculation. The reason to build these personal tools now isn't just convenience. It's that AI can do genuinely useful work on your data if it can actually reach it. The Slack automation I had in mind wasn't just inbox triage. It was an agent that reads context, drafts in my voice, and keeps me out of Slack almost entirely. That's not a productivity tweak. That's a different way of working.
Multiply that across five tools, and the value of agentic access to personal data goes from "nice automation" to "meaningful leverage" fast. The platforms haven't moved at all. And the tools to build around them are now available to almost anyone.
Why I think this changes, and when
This is a hypothesis, not a prediction. But two things make me think the current state doesn't hold.
The first is that the capability to build personal tooling is now effectively universal. A year ago, wiring up an AI agent to read and respond to Slack messages required real engineering. Today it's a Claude Code session or a Make.com workflow. The number of people who can build what I was trying to build is expanding fast, and it cuts across every tier. Someone blocked from automating their LinkedIn can write a scraper. Someone blocked from the Google Photos API can pull their library through Google Takeout and run a local script. The wall doesn't disappear, but the routes around it multiply.
The second is that enforcement is structurally difficult in ways that matter. Take Facebook Marketplace: nothing stops someone from building a public website that aggregates listings using the same data that's already publicly visible, presented differently. LinkedIn's grey zone exists precisely because the platform can't cleanly distinguish a human clicking through a profile from a script doing the same thing at human-like speed. The ToS is clear but the boundary is fuzzy, and it gets fuzzier as AI tooling gets better at mimicking natural behaviour.
Regulatory tailwinds help at the margin. The EU Data Act came into force in September 2025, establishing that data you generate is not the platform's to lock away. It's currently focused on IoT devices, not social platforms, but the principle expands. MCP (Model Context Protocol), open-sourced by Anthropic in late 2024 and donated to the Linux Foundation in December 2025, is a standardised protocol for connecting AI agents to data sources without custom per-platform integrations. It's pointed mostly at enterprise systems today, but the architecture is the same one that would work for personal data.
I apply what I think of as the six-month rule: something fundamental in AI shifts roughly every six months. Not incremental updates, but actual renegotiations of what's possible. I've been watching that cycle long enough to know that the assumptions baked into today's permission models can go stale faster than the platforms expect. I'm keeping an eye on this one specifically.
Fig. 2 — From platform gatekeeping to personal-first access
What I'm watching
For now, I'm accepting the friction. Browser automation where there's no API. Workarounds where ToS makes a compliant path impossible. Manual steps where the platform hasn't caught up. The workflows are worth having even in imperfect form.
What I keep coming back to is the banking analogy. A decade ago, a custom report from your bank meant calling them. Then online banking happened, and you could download a CSV yourself. Nobody called that a revolution. It was just the obvious next step once enough people wanted it and one bank decided to offer it.
The idea that automating your own data requires you to act like a software company, filing for API review, signing enterprise terms, building a product for other people just to access your own activity, will start to look as strange as needing a business account to download your own bank statements.
The gap cost me several evenings and weekends across five projects. The capability to close it is already in the hands of anyone willing to look. It's just a matter of when the path of least resistance becomes the compliant one.