What it can see
- Only what the signed-in person can see. Every lookup the assistant makes runs with that person’s permissions. A sales executive’s assistant cannot read salaries; an HR manager’s cannot read the bank book unless their role allows it.
- When it searches records by meaning, results are filtered by the same permissions at the moment of the search.
- “Ask about this page” sends the page’s help text and the fields you can see on it; on an open record, the values of those fields.
- AI agents run on a schedule or when a record is saved, as a user an administrator chose in advance, with that user’s permissions.
What it can do
- It proposes; a person confirms. When the assistant drafts a change, such as a new task or an updated contact, it shows the proposed change on screen. Nothing is saved until the same person presses Confirm, and a proposal left for 30 minutes expires.
- The AI buttons on each page work the same way: what they read from a bill, an order or a photo becomes a draft you check.
- One exception is yours to choose: if the owner switches on auto-post for Read a bill (off by default), a bill that passes every check is posted with its scan attached.
- It answers questions about your business in plain words, for example “which customers have not paid for more than 60 days?”
How answers are shown
- Answers appear as the assistant’s, separate from your records.
- They link to the records they came from, and “Why this answer?” shows what was looked up. On our demo company (23 Oct 2026), “How many open leads do I have?” got “You have 31 open leads”, followed by the filter it used: leads whose status is not converted, junk or unqualified.
- Where AI puts a page’s figures in words, any figure that is not on the page is flagged.
- People can mark an answer good or bad; the administrator sees these.
What leaves your system
- The question, the earlier messages of the conversation and the records needed to answer it, never the whole database.
- Personal details are masked. Phone numbers, e-mail addresses, PAN and Aadhaar numbers are replaced with placeholders before text is sent, and put back in the answer you read. This is on by default.
- Photos and scans cannot be masked, so an administrator can stop the AI reading them.
- AI Settings shows, under “Where your data goes”, where text is sent for the provider you set up, and says so when the AI runs on your own network.
What is recorded
- Every AI call is logged with what it cost. The AI activity log is kept for 180 days, AI agent runs for 90 days and usage records for 400 days.
- Conversations can be kept private to the person who had them, administrators included, and deleted after a set number of days. Switching privacy off is kept in the settings history.
- Each person sees their own AI use for the month in My settings, and can delete all their conversations there.
- What people type is treated as data, never as instructions to change the assistant’s rules. Attempts to trick it into ignoring permissions or leaking data are part of our test suite.
Who pays for AI use
- MyRecords Cloud: a monthly AI budget is included when the AI Assistant is switched on. If you need more, we raise it and bill the extra use.
- Your own server: you use your own AI provider key and pay the provider directly, or buy prepaid AI credits from us.
- Limits you set: a monthly AI budget in rupees, after which the AI pauses until next month or until you raise it, and, by default, 120 AI requests an hour per person.
On your own server, the AI Assistant needs a working AI provider key: Google Gemini, Vertex AI on Google Cloud, or any OpenAI-compatible service. Without one, everything else in MyRecords works as normal.
Limits
- AI answers can be wrong. Check figures that matter, such as tax amounts, against the report itself before you act on them.
- The assistant does not file anything with the government.