Agents prepare, people approve: keeping control of an AI native ERP
The first question every owner and every auditor asks about an AI that works in the books is the right one: who is in control? The answer should be designed into the system, not promised in a brochure.
The fear is reasonable
An owner who hears that an AI will prepare purchase bills, match the bank and draft customer reminders is right to ask what stops it booking the wrong thing, paying the wrong supplier or telling a customer something untrue. Finance is the one part of a business where a confident mistake is worse than no answer at all.
The answer is not to trust the AI more. It is to give it no more power than a careful junior accountant on their first day: it may prepare anything, and it may finish nothing without a person.
The agent prepares; a person reviews and posts. Every rule that applies to a person, rights, approval limits, segregation of duties and the audit trail, applies to the agent too.
6 controls an AI native ERP should have
The agent's output is a draft: a bill, a receipt, a journal, a reminder. It reaches the ledger, the tax portal or a customer only when a person with the right to do so approves it. No exceptions for small amounts.
If a payment above 5,00,000 needs a second approver when a person prepares it, it needs one when the agent prepares it. The agent does not get its own, looser rules.
When a salesman asks a question, the agent answers from that salesman's customers only. Rights are checked on every answer, not just on every screen.
The person who approves an agent prepared supplier bill should not be the person who approves its payment, exactly as if a clerk had typed it.
What the agent read (the PDF, the bank line), what it prepared, what it changed, who approved it and when. The trail must not be editable, by the agent or by anyone else.
An answer or a prepared document says how each number was arrived at and opens the entries behind it. A figure nobody can trace does not belong in the books.
One supplier bill, end to end
A supplier emails a PDF bill for copper wire. Here is what happens in a well designed system.
| Step | Who | What the audit trail records |
|---|---|---|
| Reads the PDF | Agent | The file, when it arrived, what was read from it |
| Prepares the bill | Agent | Supplier, items, quantities, rates, GST, matched to the purchase order and goods receipt; any difference flagged |
| Checks the supplier's GST filing | Agent | Whether the invoice appears in the input tax credit statement, and if not, that credit is held |
| Reviews and posts | Accountant | Who posted, when, and any change made to the agent's draft |
| Approves the payment | Owner or finance head | A different person from the one who posted the bill |
The agent did 3 of the 5 steps and none of the 2 that change the books or move money.
Why this matters more in India
Indian books carry obligations that make traceability non negotiable. Input tax credit depends on what the supplier filed, not on the bill alone. Electronic invoices and eway bills are reported to government systems as they happen. Accounting software used by companies must keep an audit trail of every change that cannot be switched off. An agent that works inside these books has to leave the same kind of trail, so that the business, its auditor and the tax officer can all follow it.
Questions to ask before you let any AI near your books
- Can the AI post, file or send anything on its own? The right answer is no.
- Do my approval limits apply to what the AI prepares? The right answer is yes, unchanged.
- Can a salesman use the AI to see figures he cannot see on screen? The right answer is no.
- Can I see, for any posted entry, whether the AI prepared it and who approved it? The right answer is yes, in 1 click.
- Does the AI compute the figures, or does the accounting engine? The safe answer is the engine.
How Niyantrak ERP is built
Niyantrak ERP, the AI native ERP built in India by Niyantrak Consulting LLP, follows each of these rules. Its AI, Controller, prepares documents and answers questions; a person posts. Approval limits, rights per screen and per action, and segregation of duties apply to Controller's work as they do to a person's, and every action it takes is in a tamper evident audit trail. That is how finance work gets automated without losing control. The cloud edition is coming soon.
Questions and answers
Is it safe to let an AI work in the accounting books?
Yes, if it can only prepare and a person approves. Approval limits, rights, segregation of duties and an audit trail must apply to the AI exactly as they apply to a person.
What is maker checker for AI in accounting?
The AI is the maker: it prepares a bill, a receipt or a journal. A person with the right authority is the checker who reviews and posts it. The AI never posts on its own.
How does Niyantrak ERP keep control of its AI?
In Niyantrak ERP, Controller prepares and a person posts. Approval limits, rights per screen and action, and segregation of duties apply to Controller, and every action is in a tamper evident audit trail.
Can AI automate finance work for a small manufacturer?
Yes. Reading supplier bills, matching the bank, reconciling input tax credit and drafting customer reminders can all be prepared by an AI agent, with a person approving the result.
Want to check how control would work in your business?
Niyantrak Consulting LLP builds control into systems for a living, and built its AI native ERP the same way. We are glad to walk through approvals, rights and the audit trail with your team or your auditor.