All Things AI · Season 3 · What You Can Build · Episode 4
You can point a model at the folder already on your computer and it will work from what is in it. The export, the report, the notes, the twelve documents attached to a deal. No pasting, no uploading one file at a time, and no summarising things by hand first.
The blank window is the bottleneck
Every agent who has used AI seriously has hit the same wall. The model is capable, your question is reasonable, and the obstacle is entirely mechanical: everything it needs to answer is sitting in files, and getting it in there is a copy-and-paste job you are not going to do.
So you paste one document instead of six. You summarise the rest by hand, which means the model is now working from your summary rather than the material. And the answer comes back shaped by what you happened to include, which is the one variable you were trying to remove.
The fix is not a bigger paste. It is giving the model access to the folder.
What file access actually looks like
There are three levels of this and the difference matters, because most agents have only met the first one.
Upload
You attach a file to a message. It works, it is fine for one document, and it does not scale past about three. Everything has to be selected by hand every time, and the model forgets it existed at the end of the session.
A folder it can search
You point a tool at a directory and it indexes what is in there. Now you can ask a question and it finds the relevant material itself, across everything, without you choosing which files matter in advance. This is the level where it stops being a chat window.
A folder it can read and write
The model can open files, work on them, and save results back. This is what turns a question-answering tool into something that completes a job: read these, produce that, put it here. It is a genuinely different category of software and it is the one most agents have never seen.
What changes when it can see everything at once
The obvious gain is that you stop pasting. The real gain is that the model can answer questions that span documents, which is the shape of almost every question worth asking.
A question about one document is usually a question you could answer yourself by reading it. A question across twelve documents is the one you have been avoiding, because answering it means holding twelve things in your head simultaneously and you have appointments.
That is where this earns its place. Not summarise this page, but tell me where these two documents disagree, and what is in this set that is not in that one, and which of these mentions the thing I care about.
How you set it up
The setup is closer to organising a drawer than to configuring software.
- Put the material in one folder. This is the actual work, and it is the step people skip. A tool pointed at a tidy folder is transformative. Pointed at a downloads directory with four years of everything in it, it is noise.
- Use a tool built for folder access rather than one built for chat. Season 2 covers which ones, and the distinction is exactly this capability.
- Give it read access first and live with that for a week before you give it write access. Reading is safe. Writing is powerful and worth easing into.
- Keep the folder current. A model reading a stale folder produces confident answers about a situation that has moved on, and it has no way to know that.
The habit that makes it work
Ask about the material, not about the world.
The reason this setup is reliable when a chat window is not is that you have given the model a source. When you ask what does this set of documents say about the timeline, it has somewhere to look and its answer can be checked against that somewhere. When you ask what is the usual timeline for this, it is generating from training and you have no source at all.
The same tool does both, and it will not warn you which one is happening. Getting into the habit of asking file-shaped questions is what makes the difference between a tool you can trust and a tool that occasionally embarrasses you.
Where this goes wrong
Three failures, and the first is by far the most common.
Pointing it at a mess. The single biggest determinant of whether this works is folder hygiene, which is unglamorous and is also the entire job. Four years of unsorted downloads produces answers assembled from documents that are outdated, superseded, or belong to someone else entirely, and the answer will be delivered with complete confidence.
Forgetting the folder is a moment in time. Documents get replaced and the old versions stay in there. A model has no sense of which file is the current one unless the folder tells it, and the most reliable way to tell it is to not keep the old one.
Giving write access before you trust it. It can save over things. Read-only for the first week is not caution for its own sake, it is how you find out what it does before it can do it to your files.
What to ask it first
The first session is where most people decide whether this is useful, so it is worth spending it on the right kind of question.
Start with something you genuinely do not know the answer to and would have to read to find out. Not a test question. A real one. The whole value of this is answering questions that were previously too expensive in time to bother asking, and you will not feel that on a question you could have answered yourself in a minute.
Good first questions look like: what is in this set that contradicts anything else in it, what is the earliest date mentioned anywhere in here, which of these documents refers to the same matter, what appears in one version and not the other, and summarise everything here that relates to one specific subject.
Then check one answer by hand. Open the document it cited and confirm. Doing that once, deliberately, tells you more about how far to trust the setup than a month of casual use, and it establishes the habit of asking where a claim came from.
What it grows
The gap between having information and using it, which in this business is enormous.
Most agents are sitting on far more material than they can read. Reports, exports, notes, records of past deals, correspondence. The information is there. What has never existed is the time to go through it, so the overwhelming majority of it is never consulted after the week it arrived.
This is the first tool that makes accumulated material an asset rather than storage. That changes what it is worth keeping good records for, which changes how the business operates over years rather than weeks.
What it costs
Free if you run it against a local model, and one ordinary subscription if you use a paid service. The genuine cost is the hour it takes to organise the folder, which is a cost you were carrying anyway in a less visible form.
Questions agents actually type
Can AI read files on my computer?
Yes. Tools built for folder access will index a directory and answer questions across everything in it, rather than requiring you to attach documents one at a time. Some can also write results back into the folder.
What is the difference between uploading a file and giving AI folder access?
Uploading is one document, selected by hand, forgotten at the end of the session. Folder access means it searches everything you have pointed it at, finds the relevant material itself, and can answer questions that span many documents at once.
Why does AI give wrong answers about my documents?
Usually because the folder contains superseded versions alongside current ones, and the model has no way to tell which is which. Folder hygiene is the single largest factor in whether this works, and it matters more than which tool you choose.
Should I let AI write to my files?
Eventually, but not on day one. Read access is safe and delivers most of the value. Give it a week of read-only use so you understand its behaviour before it is able to save over anything.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.