What leaves the computer, what stays, and who can see what.

Written for the office manager, not the auditor. We claim no certification we do not hold.

Last updated 30 October 2026. About Has'an Employee, our desktop app.

What stays on the computer

Has'an works on the accountant's own Windows computer, and this part of the work never leaves it.

By default the screen is read on the computer: the accessibility layer first, then Windows OCR, then a small vision model on the same machine.
Watch-and-learn recordings and the guides written from them stay in your own data folder. Frames are deleted once they are described. If you type into a password field while recording, that window and its frames are dropped.
Chats, documents and learned jobs stay on the computer until you delete them in the app, unless you share a job to the company library.
Voice typing runs on a local Whisper model by default, and audio is never written to disk.
With a local main model, no model call leaves the computer at all.

A cloud model is used only if you say yes, at install time or later in Settings. You can take the yes back in Settings at any time.

With the cloud model on, the text the model needs for the job goes through your company server to OpenAI: your message, text read from the screen and text from the files the job opens.
Screenshots, scanned pages and voice recordings are sent only with a separate consent for each.
Before a screenshot is sent, Turkish ID and tax numbers, IBANs, card and phone numbers and email addresses are removed from it. Names, addresses and amounts are not. A screenshot that cannot be masked is not sent.
Requests go out with OpenAI's storage option turned off. How long OpenAI keeps data is set by OpenAI's own policy, not by our code.
Web search is off in every mode.

Every action with a consequence stops and asks. The list of what Has'an may touch is sealed, and only the user can change it.

Programs Has'an may act on are listed in a sealed allowlist. A list edited by hand is treated as invalid, and the screen tool refuses to act.
Confirm words such as 'print', 'send' or your own words stop the tool before the click and ask in the mini window.
Writes outside the job folder, elevated commands and network access go to an approval, not straight to the system.
Every screen action is written to a ledger on the computer, so what Has'an did can be read back later.
You can stop Has'an at any moment with the Stop button.

Each company runs its own server, or uses ours. It connects the app to the model, meters use, ships updates and holds the shared job library.

It keeps each person's name and email, the computer's name, the app version, active time per day, which sections were opened, and for every model call the time, model, token counts, cost and duration.
It does not keep messages, documents, screen text or images. Model calls pass through in memory, and request bodies are not written to its logs.
Each person has a monthly limit. Users see a percentage, never a model name or a price; amounts appear only in the admin panel.
The computer signs in with a token kept in Windows Credential Manager. The server stores only a hash of it.
Updates are signed. The app installs only what verifies, and only when nothing is running or unsaved.
Jobs shared to the company library go through admin review before anyone else can add them.

Besides us, two companies touch Has'an data, each for one job. Offices that run their own server leave Railway out.

Who else processes it
Company	What it does	Where
OpenAI	Runs the cloud language model, and screen reading and voice typing when you consent to them	United States
Railway	Hosts the company server we run: accounts, usage counts, updates and the job library	Railway's cloud. Ask us for the region.

OpenAI
Runs the cloud language model, and screen reading and voice typing when you consent to them
Where: United States
Railway
Hosts the company server we run: accounts, usage counts, updates and the job library
Where: Railway's cloud. Ask us for the region.

Questions about a specific deployment, or a request under KVKK Article 11: write to us and name the program and the office size.