watts.it.com // daily AI micro-learning
Tools & connectors connectorsdata freshnessefficiencyworkflow 2026·06·08 · 4 min · evergreen

Connectors vs. file upload: when to plug into your live data and when to just drag a file in

Overview

This is the foundational choice for the whole pillar: do you wire the AI into a live data source, or do you just drag a file into the chat? In June 2025 OpenAI shipped connectors for Google Drive, SharePoint, OneDrive, Dropbox and Box to paid users, and by December 2025 had folded them into a broader “apps” experience; Anthropic launched its connector directory in July 2025, and by April 2026 it had grown past 200 integrations. Plugging in is now a daily decision, not a setup task. By the end of this you’ll be able to make that call in seconds, on a real task, with a rule you can defend to a sceptical reviewer.

The content

The obvious read is that a connector is just a fancier, slower upload — same data, more clicks, so why bother. That gets the difference exactly backwards.

A connector keeps a live link to the source. The model queries the system at the moment you ask, so it sees the current state and, crucially, respects the permissions already set on that source — connectors follow your organisation’s existing access controls rather than creating a new copy outside them. An upload does the opposite: it drops a frozen snapshot into the conversation. That copy is yours, it’s fast, it works offline from the source, and it never silently changes under you.

Neither is better. They fail in opposite directions, and that’s the whole decision.

Reach for a connector when freshness is load-bearing and the work recurs: a pipeline report you pull every Monday, a policy doc that keeps getting edited, anything where yesterday’s version is a wrong answer. The cost is that a live link is only as current and as scoped as the source — if the underlying file is stale or your access is broader than it should be, the model inherits both problems.

Reach for an upload when the task is one-off, or when you specifically want a fixed reference: the exact contract draft you’re marking up, a dataset you need to stay still while you reason over it, a document you’d rather not wire a standing connection to. The cost is that the moment the source moves, your copy is wrong and won’t tell you.

So the lens is three questions. Does the answer have to be current (freshness)? Will I do this again (one-off vs recurring)? And how sensitive is the source — am I comfortable with a standing link, or do I want a copy I control (sensitivity)? Live and current points to a connector. One-off, fixed or sensitive points to an upload.

Try it

Don’t take this on faith — run it against something on your plate right now. Pick a task you’d reach for AI on today and score it before you decide how to feed it the data:

I'm about to give an AI assistant some data for this task: [describe the task in one line].

Score it on three axes, briefly:
1. Freshness — does the answer have to reflect the current state of the source, or is a snapshot fine?
2. Frequency — is this one-off, or will I repeat it on a schedule?
3. Sensitivity — am I comfortable with a standing live link to this source, or do I want a copy I control?

Then recommend: live connector or one-off file upload? Name the main risk of your choice.

Where this breaks: if you can’t actually answer axis 3 honestly — if you don’t know what a connector would expose or who set the permissions on that source — don’t connect it yet. A live link inherits whatever access the source already grants, so “I’m not sure what’s in there” is a reason to upload a vetted file, or to stop and check, not to plug in and hope.

Additional reading

Editor’s note

The freshness axis gets all the attention, but it’s the sensitivity one that bites. A connector inherits whatever the source already grants, so the model can read things you’d forgotten were shared. For example: wire up a team drive to answer one quick question, and you’ve also handed the model a half-drafted meeting agenda. Connect for recurring work using well-understood and current materials; attach/upload for anything else.

signed-off-by: Luke Topfer <editor> · 2026·06·08