SFTP
Not everything has an interface. A great deal of business still moves as files on a schedule, and that data is as useful as any other.
Not everything has an interface. A great deal of business still moves as files on a schedule, and that data is as useful as any other.
What SFTP holds, and why that matters here
SFTP file exchange remains standard for batch data between organisations — supplier catalogues, bank statements, government returns, partner reports. It is unglamorous, reliable, and holds data that frequently exists nowhere else in an accessible form.
For an assistant, the relevant case is answering from data that arrives as files rather than through an interface, which is common in exactly the industries where interfaces are scarce.
- Scheduled files — data arriving on a defined cadence.
- Structured exports — reports and extracts from partner systems.
- Catalogues — product and price data from suppliers.
- Statements — financial data delivered as files.
- Delivery state — whether an expected file arrived.
Questions Elbi can answer from files delivered over SFTP
Did the file arrive?
Whether an expected transfer completed, which is a constant operational question.
What is in the latest file?
Content from the most recent delivery.
What is the price for this?
Catalogue data that only exists as a periodic file.
Has the statement been received?
Delivery confirmation for financial files.
When did the last one arrive?
Transfer history and cadence.
Is anything missing?
Gaps against an expected schedule.
How Elbi reads from an SFTP drop
- A dedicated account with a key, not a passwordYour team creates an SFTP account for the assistant on your server and authorises an SSH key for it. Key-based authentication leaves no shared password to leak or reuse, and removing the key from the account ends access.
- Confined to the agreed directoriesThe account is confined to the directories in scope — the nightly export folder, the supplier catalogue drop — and, on OpenSSH, the SFTP subsystem can run in read-only mode so the account cannot upload, rename or delete anything.
- Arrival is checked against a scheduleFor each expected file you agree a name pattern and a window: the settlement file by early morning, the price list every Sunday. "Did it arrive?" then has a precise answer — present, late or missing — rather than a guess based on whatever happens to be in the folder.
- Half-written files are left aloneA file that is still being uploaded looks complete to a directory listing. Agreeing the convention your sender uses — a temporary name renamed on completion, or a companion file marking it done — keeps answers from being drawn out of a partial export.
- The asker's role decides the answerA statement file may be within the connection's reach and still be answered only for finance. The person asking is identified through your own sign-in, and what they are told follows their role, not the account's directory access.
- Files are read to answer, then let goWhat a question needs is read from the current file at the moment it is asked. Files are not accumulated into a store of our own; the history stays on your server, under your housekeeping rules.
Keys, directories and what stays yours
| Direction of access | Read-only. The account can list and download within its directories and do nothing else, enforced on the server by read-only SFTP mode or file permissions rather than by agreement alone. |
|---|---|
| Who can see what | Directory scope decides which files the connection can reach; the asker's identity decides which answers they receive from them. |
| Credentials | An SSH key you authorise on a dedicated account. Removing the key, or disabling the account, ends access at once and from your side, without our involvement. The server's host key fingerprint is recorded at setup so the connection can tell your server from an impostor. |
| Where data goes | Values drawn from a file answer the question in front of the assistant. The file itself is not kept, message content is destroyed within 24 hours — normally as soon as the conversation closes — and no outside AI service sees any of it. |
| Audit | Your SFTP server's own logs record each login and, with a more detailed log level, each file opened. Those logs, held on your side, are the primary record of what was read. |
| If it is unavailable | If the server refuses the key, cannot be reached, or the expected file is absent, the assistant says which of those it is and offers a person. A missing file is reported as missing, never answered from an older file as though it were today's. |
What the SFTP connection replaces
The most common operational question here is simply whether a file arrived. Batch processes fail quietly, and the failure is often discovered downstream when something else does not work. An assistant that can confirm arrival turns a discovery into a check.
The second is content. Catalogue and price data delivered as files frequently has no other interface, which means answering an ordinary question requires someone to open a file. That is exactly the kind of task worth removing.
The third is schedule awareness. Knowing that a file is late, rather than discovering it through a downstream failure, is a meaningful operational improvement in file-driven industries.
Where scheduled file exchange is still the norm
Distribution
Supplier catalogues and price files.
Banking
Statement and reconciliation files.
Government
Regulatory returns and periodic submissions.
Logistics
Partner data exchange on a schedule.
Getting value from scheduled file drops
Write down, for every file, who sends it, what it is called, when it should arrive and who depends on it. That register is the core of the connection, and in many organisations it exists nowhere else — the knowledge lives with whoever set up the transfer years ago.
Content questions depend on format. A clean CSV with a header row is straightforward; fixed-width exports, spreadsheets with merged header rows and files whose columns change without notice are not. Agreeing a stable layout with the sender does more for answer quality than anything done on the assistant's side.
Encrypted files need a decision. Banks and payment partners often encrypt files with PGP before sending, and answering from their content would mean holding a decryption key. Arrival questions need no key at all, so connecting for arrival first and deciding about content separately is a sensible order.
What connecting an SFTP server involves
- Build the file registerList every file in scope with its sender, name pattern, format, expected arrival time and the people who rely on it. The register sets the scope and doubles as documentation you may not have had.
- Create the account and directoriesSet up a dedicated account confined to the relevant directories, in read-only mode where your server supports it, and separate from any account a sender uses to upload.
- Authorise the key and record the host keyAdd the connection's public key to the account and record your server's host key fingerprint for the connection, so each end can verify the other.
- Agree completeness signalsConfirm how each sender marks a file as finished — a rename, a marker file or a fixed completion time — and what counts as late for each feed.
- Test the bad daysRun it against a missing file, a late file, an empty file and a file whose layout has changed. Each should produce a clear answer and, where it matters, a person.
- Connect one feed firstBegin with the file people ask about most, often a daily export or a supplier price list, and add the others once arrival and content answers for that one are trusted.
Personal data in files that arrive on a schedule
Files exchanged over SFTP often carry more personal data than any single question needs: a settlement file lists every payer, a partner export holds full customer records. Under the Saudi Personal Data Protection Law, answers should draw only on the fields a question requires, and feeds that consist mainly of personal data deserve a narrower scope than catalogues or price lists.
The data sits wherever the server sits — on your premises, with a hosting provider, or on a partner's server you collect from. Record which, alongside the file register, so a reviewer can see where each file lives as well as who may read it.
Because files are read rather than accumulated, retention stays where it already was: with your server's housekeeping. The connection does not create a second archive that needs a deletion schedule of its own.
What organisations typically see
How file-transfer questions are handled elsewhere
Managed file transfer products are the established answer to "did it arrive?": they track transfers, retry failures and alert when an expected file is late. Where one is already in place, its records may be a better source for arrival questions than the directory itself.
Cloud platforms now accept SFTP straight into storage — AWS Transfer Family and the SFTP support in Azure Blob Storage are two examples — which lands files in buckets or containers where other tools can pick them up.
Many teams still rely on a scheduled script that checks the folder and emails someone when a file is missing. It works until the person on the email list changes role, and it answers only the one question it was written for.
The recurring failure is an account with more access than a reader needs, typically the sender's own upload credentials reused for convenience. A reader that can upload or delete is able to overwrite the very file it is meant to report on.
Common questions about the SFTP connection
Only those in scope. Which directories and file types are agreed at deployment.
Yes, against an expected schedule — which turns a downstream failure into an early check.
No. Files are read to answer questions, not accumulated — the assistant does not keep its own copy.
Your SFTP server, through an account and key you authorise. It reads the agreed directories and nothing else on the server, and removing the key ends access without going through us.
Arrival and timing can be checked without decrypting anything. Answering from the content of an encrypted file would mean holding its decryption key, which is a separate decision for each feed and one to settle during deployment.
No. The account is read-only, ideally enforced on the server itself. Moving, renaming and deleting files stay with the processes that own the transfer.
Connect it to your own systems
A working assistant against your own records, in both languages.