Beta 0.16.0 · updated 2 October 2026

Built for legal and accounting work, and tested on it.

Use RangerIO with your own files from day one. This page explains what the team tests before every build, how your files are protected, and the few things we are still fixing, so nothing surprises you.

What we test on

Every build is checked against files like the ones you work with, in the formats you use.

Accounting and finance

General ledger detail, trial balances, AP aging, bank transactions and statements, supplier invoices, expense claims, FX rates, forecasts and close memos.

Legal

Master services agreements and amendments, NDAs, contract registers, policy handbooks, authorizations and email threads.

Audit and compliance

Audit request lists, audit findings, control registers, compliance findings and M&A due-diligence checklists.

HR and payroll

Employee rosters, payroll registers, timesheets, departures and access events.

Sales and customers

Sales pipelines, rep performance, orders, project revenue and customer churn.

Mixed questions

Questions across several files at once, follow-up conversations, and tables and documents selected together.

Formats: CSV, Excel workbooks with several sheets, PDF, Word, JSON, text and HTML, from 6 rows to 121,020 rows.

Thousands

of questions asked and graded through the QA process, on finance, legal, audit, HR and sales files

1,000s of hours

of testing and validation, of answers and of privacy protection

Every build

re-checked against a fixed release question set before it ships, together with the automated security checks

Certification

RangerIO is going through industry security certifications

Your files and your privacy

What happens to a file from the moment you import it.

Checked several times before anything is sent

RangerIO sends only the parts a question needs, with names, emails, phone numbers, addresses, account and card numbers and IDs swapped for placeholders, then puts them back in the answer you see.

It scans for personal data at several points: when a file is imported, while the question and its data are prepared, and once more on the final request, just before it leaves your computer. Reports and other analysis go through the same final check. If a detected value would still go out unmasked, the request is blocked and nothing is sent.

You can see exactly what was sent

Under each answer, Cloud request lists what was replaced. View lineage opens the full record in Settings, under Security & Audit.

Business content such as figures and clause wording is sent as it is, so the model can answer. The AI provider is the one you connect, under your own account and key. RangerIO is going through industry security certifications.

Using your own files

Start with files you know well, so you can judge the answers.
Leave out files with medical details for now, such as personal injury files, medical leave or accommodation records. Health terms inside documents are not masked yet.
Keep client details out of bug reports: describe the file and the question instead of attaching the file. The guide's last section shows how.

Known limitations

You don't need to report these. Do tell us if one happens more often than described, or looks worse.

Fixed in next build Open After UAT By design

Privacy

Accuracy of certain health and medical terms inside documents miss maskin.
A diagnosis written in a contract note or a leave record is sent as it is.

leave out files with medical details.
Open
Some ordinary values are masked as if they were personal.
A city can be read as a person's name, and some invoice numbers, policy codes and amounts get masked. Answers about those values can be less precise.

rephrase, or ask about the record rather than the value.
By design
Ordinary words can trigger a "which person?" question.
Words such as "bank", "expense" or "payroll" are sometimes taken for a person's name.

pick the right option or rephrase.
Open

Answers

The AI reads large tables as a sample.
Tables up to about 170 rows are read whole. Larger ones are sampled, and the answer says so. Calculations always use every row.

ask big-table questions as a total, count or average.
By design
Ask a filtered total in full, and leave the file name out.
"What is the total Gross for period 2026-06?" is calculated exactly. Adding the file's name ("…in the payroll register") or a month name ("…for June 2026") sends it to the AI with a sample instead, which may decline.

ask again in full, without the file name.
After UAT
A one-word follow-up ignores the conversation.
"total" on its own is answered as a new total of the whole table.

write follow-ups as full questions.
Open
A word that is also a column name is used as that column.
"How many lines for each Department?" on a ledger with a Line column counted line numbers, not rows.

if a count looks far too small, ask "how many entries…" instead, and let us know.
Open
A column with a year in its name isn't used for calculations.
"Total support tickets in 2024" doesn't match a column named support_tickets_2024, so it goes to the AI instead: 60 seconds against 0.3 in our test, and on some files the AI says it cannot compute it.

use the column name exactly as it appears (type @ to pick it).
Open
The AI can still get arithmetic wrong from complete data.
For example, picking the wrong month as the lowest from a correct table.

use Validate metrics, and report the question with the figures.
By design
Some answers mention internal labels.
Such as "Chunk 6" instead of a section number, or "dataset statistics".
Open
"Which documents…?" over many documents can list only some.
In one test it named 3 of 5.

select fewer documents, or ask about each one.
Open
A question about a named person in a roster can miss their row.
note the question and report it.
Open

Speed

Several files full of personal details are slow.
A question across five such files takes about 100 seconds, because each value is replaced before sending. It is slow, not broken.

select only the files you need.
Open
The first question on a large table waits for its privacy check.
About a second per 1,000 characters of table. Later questions on the same table are quick.
Open
Reasoning models stop after 150 seconds of silent thinking.
The screen says so (E020) and suggests another model.

ask once more; if it repeats, switch models.
By design

Files

Figures in PDF tables are read as text, not as a table.
RangerIO reads the figures in a PDF's tables, and the import message names any table it could not read. A total, count or average over a PDF table is worked out by the AI from that text, not calculated exactly.

if you have the same figures as a spreadsheet or CSV, import that for exact calculations.
Open
Workbooks are sometimes counted sheet by sheet.
The file picker can say "2 selected" for one workbook, and selecting a whole workbook can add "could not be included in this answer" even when its sheets were used.

select the sheets you need, and check the answer.
Open
Scanned PDFs and very large free-text files are refused.
A scanned PDF (a picture of text) is refused with E042; the message explains how to import it again with OCR. A file with a great deal of long free text is refused with a suggested way to split it.
By design
A very long document import can stop without a message.
Most likely on slower computers, when a document takes more than 20 minutes to check.

note the file's type and size, and report it.
Open

Found something that isn't on this list?

Report it, with the question, the answer, and a description of the file (type, size, rows and columns). A refusal or failure with no reason and nothing to do next is a blocker: tell your test lead the same day.