Does AI train on your accounting data? A Malaysian answer
Written by the Lejar team
Most AI tools train on what you type unless you turn it off, and several large providers changed their consumer terms in that direction during 2025. Accounting data is the worst possible thing to hand to a model on those terms: it carries your margins, your customer list, your payroll, and your bank balances. This page answers the question directly for Lejar, names every provider that touches your data, and gives you the questions to put to any other vendor.
Does Lejar train AI models on your books?
No. Your ledger, documents, and chat prompts are never used to train an AI model, not by Lejar and not by the providers Lejar sends requests to. The AI providers we use are contractually barred from training on your content, and we do not run any training of our own on customer data. If that ever changed it would be a change to the privacy policy, which is public and dated.
Where your data actually goes when you use an AI feature
Only the content for that one request leaves the database, and only when you trigger an AI feature such as scanning a receipt, importing a bank statement, or asking the Ask lejar assistant. Browsing your own reports sends nothing to an AI provider. The full sub-processor register, with each provider role and a link to its data processing addendum, is on the security page.
| Provider | Role | Position on training |
|---|---|---|
| Anthropic | AI document processing: receipt and statement extraction, Ask lejar chat | Does not train on your content. SOC 2 Type I and II, ISO 27001:2022, ISO/IEC 42001:2023 |
| Cloudflare | AI gateway routing and Workers AI embeddings | Does not train on your content. ISO 27001:2022, ISO 27701:2019, SOC 2 Type II |
What Lejar does learn from you, and how to delete it
There is one thing that does learn, and it is worth understanding because most vendors are vague about it. When you correct a draft, for example telling Lejar that a particular supplier belongs to a particular account, that correction can become an explicit rule stored inside your organisation. It is used only to prepare your own future drafts. It is not shared with other customers, it is not pooled, and it is not used to train any model. You can view, export, and delete these rules yourself in Settings.
That is the honest distinction to hold on to. A system that remembers your preferences inside your own account is not the same thing as a system that folds your books into a model other people will use. Ask any vendor which of the two they are doing, because both get described as the software learning from you.
What "does not train on your content" actually means
It is a contractual commitment in a data processing addendum, not a technical impossibility. That is worth saying plainly rather than implying something stronger. What makes it checkable is that the addendum is published, the provider is audited against named standards, and the register naming that provider is public and dated. A vendor that will not tell you which AI provider it uses cannot give you any of that.
The questions to ask any accounting vendor
- Which specific AI provider processes my documents? A vendor that answers "our AI" rather than naming a company cannot show you a data processing addendum.
- Is training on customer content barred by contract, and can I read that contract?
- What leaves the database, and on which actions? "Everything, continuously" and "the one document you scanned" are very different products.
- Does the AI write to my ledger on its own, or does a person approve every entry?
- If the software learns from my corrections, is that stored in my account only, or pooled across customers?
- Can I export and delete what it has learned, without asking support?
- Where is the data stored, and who at the vendor can read it?
Why this matters more for accounting than for most software
A bookkeeping system holds the things a competitor would most like to know and a regulator is most entitled to see: pricing by customer, supplier terms, salaries, and the bank position. It is also data you are obliged to keep and be able to produce. Those two facts together are why "where does it go" deserves a straight answer before you upload anything, and why the answer should be written down rather than given in a sales call.
Malaysian context: PDPA
Lejar is operated from Kuala Lumpur by Iterate Ventures Sdn. Bhd. and handles personal data under the Personal Data Protection Act. The privacy policy sets out what is collected, the legal basis, how long it is kept, and the rights you can exercise, including access, correction, and deletion. Read it alongside the sub-processor register on the security page rather than on its own, because between them they answer the where and the who.
Frequently asked questions
Does Lejar use my accounting data to train AI?
No. Your ledger, documents, and chat prompts are never used to train an AI model, by Lejar or by its AI providers, who are contractually barred from training on customer content under published data processing addenda. That is a contractual commitment backed by audited certifications, not a technical impossibility.
Does Anthropic train on the documents Lejar sends?
No. Anthropic is listed in the Lejar sub-processor register as not training on your content, under a published data processing addendum, and is certified to SOC 2 Type I and II, ISO 27001:2022, and ISO/IEC 42001:2023.
What data leaves the database when I scan a receipt?
The content of that request only, and only because you triggered an AI feature. Viewing your own reports and ledger sends nothing to an AI provider.
Lejar says it learns from my corrections. Is that training?
No. A correction can become an explicit rule stored inside your own organisation and used only for your future drafts. It is never pooled across customers and never used to train a model. You can view, export, and delete these rules in Settings.
Can the AI post entries to my ledger by itself?
No. The AI drafts, and a person approves and posts. Nothing reaches the ledger without a human action.
How do I check a vendor is telling the truth about AI training?
Ask which AI provider processes your documents, then read the published data processing addendum from that provider yourself. A vendor that will not name its provider cannot evidence the claim.