What is protected, and how
A rewrite that changes a number is worse than no rewrite. So the figures SayAlign recognises in every result are compared with yours before it can replace your text, and most results are also read by an AI review for what they commit you to.
Fact Guard
Before a result can replace your text, SayAlign extracts the facts from both — amounts with their currency, dates and times, percentages, quantities with their units, and invoice, order or reference numbers — and compares them. It runs in your browser, on every result, on both plans.
What happens depends on what you asked for:
- Rewrites. A changed amount, date, time, percentage, quantity or reference blocks Replace for that result. So does a new amount, date, percentage or reference you did not write. A new quantity, or a figure that was dropped, shows a warning.
- Replies. Only figures that appear in neither the message you are answering nor your notes are flagged. A new amount, date, percentage or reference blocks Insert; anything else warns.
- Situations (Push back, Negotiate and the rest), written from your notes. A different number is often the point — a negotiation proposes one — so the guard warns and never blocks, and figures your message leaves out are not reported.
When Replace is blocked, Copy stays available: you can take the text and fix it yourself, or press Try again.
Languages. SayAlign writes in whatever language your message is in (Output: Same language), or in English. Fact Guard compares amounts with a currency, percentages and numeric dates (15.10.2026) in any language. Figures that depend on a word are recognised in fewer languages so far: a date written with a month name (“15 October”, “15 Ekim”), a quantity with a unit word (“500 units”, “500 adet”) and an invoice or order number after that word are read in English and Turkish, and invoice and order numbers also in Polish. In other languages those figures are not compared yet, so check them yourself. Across a language change, “Fatura 4471” and “Invoice 4471” are the same invoice; a pair the guard cannot line up warns rather than blocks.
What Fact Guard does not do
It compares figures. It does not judge meaning, tone or intent, and a figure it does not recognise is not compared at all.
Meaning review
After most results — rewrites, replies and situations — the original and the result are read once more by an AI model, looking for five things: a position reversed, a position softened until it no longer reads as a decision, a promise or offer you did not make, work or a next step assigned to you, and a deadline you never gave. If it finds one, the result carries a warning that quotes the words concerned.
- It warns and never blocks. Replace waits for it at most a second and a half.
- It is not run on Fix grammar, and not on Concise or Friendly unless the language changed.
- If the review cannot run, you see no warning. No warning is not a guarantee — it is a second reading, not a proof.
- It uses neither your actions nor your checks.
Before you send
Below the actions, the panel checks the message you are writing. Instantly, in your browser and without limit: a leftover placeholder such as [AMOUNT], an empty salutation, an invoice number that is missing, a date that does not exist or has passed, and a request with no date. On request, “Check what this commits me to” sends the whole draft for an AI reading that lists what the message commits you to and notes whether it reads clearly. It changes nothing in your message. Each one uses a check: 5 a month on Free, 250 on Pro.
What the prompts aim for
The instructions behind every action tell the model to keep your position, to add no facts, promises or courtesy lines you did not write, and to keep who did what. That is what the product is designed to do; it is not something a check can prove for every message. Read the result the way you would read anything before sending it.
Your text
| When you use | What is sent |
|---|---|
| Rewrite, or a situation such as Push back | The text you selected. |
| Reply | The message you are answering — the part you selected, or the latest message in the conversation when you start from the reply box — and your notes, if you wrote any. |
| Meaning review (after most results) | The text that action started from and the result it produced. |
| Check what this commits me to (only when you press it) | The whole draft you are writing. |
| Fact Guard and the instant checks | Nothing. They run in your browser. |
Text is sent to SayAlign’s service and on to our AI provider, OpenAI, only to produce the result. SayAlign does not store it. OpenAI may keep request data for a limited period to detect abuse, under its own API policy, and does not use it for training. Your inbox list, contacts and other conversations are not sent.
We keep metadata about the action — which one you chose, whether it succeeded, how long it took — and never the content.
The extension keeps your session and panel position in browser storage. It has site access to Gmail, Outlook on the web, LinkedIn, the mail pages of common webmail services and SayAlign’s own service. On any other site it runs only after you turn it on there from its toolbar button, and your browser asks you first. It never asks for your history, your cookies or access to every site.
