AI & ERP

How should an SME use an AI assistant inside its ERP without losing control?

Updated 26 October 20266 min readNaffo product team

How can a small business use an AI assistant inside its ERP safely?

Separate what the assistant may read from what it may change. Use a read-only mode for questions about sales, stock and dues. Let a second mode draft entries, but save nothing until a person confirms, and only with the same permission the normal form needs. For multi-step jobs, have the assistant stage each write for review. Start with questions, check answers against reports, and give drafting rights to roles that already make those entries.

Key takeaways

Keep questions and changes apart: a read-only mode for answers and a drafting mode for entries.
An AI draft should become a record only after a person confirms it, with the permission the normal form already requires.
Multi-step agent work is safest when every write is staged and visible before anything is saved.
A daily brief and a suggestions feed are most useful when they point to records a person can open and check.
Roll out by role and check AI answers against standard reports before relying on them.

Why an ERP assistant needs modes

An assistant that can answer "who owes me money" can also be asked to "record a payment from that customer". The first is harmless; the second moves money. Mixing both in one open-ended chat makes it hard for a user to know when the assistant is only looking and when it is about to change the books. Clear modes solve that by making the intent visible before anything happens.

Naffo AI has two modes. Ask is for questions and analysis and does not write. Do drafts entries for you to confirm. Where agent support is enabled, Do can hand a multi-step job to an agent that stages each write for review. A company can also connect Claude, ChatGPT, Gemini or another MCP client to the same data through the AI connections guide. The same account permissions apply there; changes made over MCP are logged and can be reverted, rather than confirmed one by one.

Ask: questions about your own records

Ask answers questions about sales, stock, dues, purchases and production from the company's own records. Where attachments are enabled, you can add a file such as a supplier statement or a price list and ask about it. Ask does not create or change records. Good first questions are ones you can check against a report you already trust.

Which customers have been overdue for more than 60 days, and how much do they owe?
Which products are below their minimum stock at the main warehouse?
What was today's milk intake by centre, compared with yesterday?
Which purchase bills are due to MSME suppliers this week?

Do: drafts that wait for your confirmation

Do turns a request into a draft: a sales order, a purchase entry, a receipt, or a procurement step. The draft shows the party, items, quantities and amounts it has picked, and nothing is saved until you confirm. A draft only becomes a record if your user has the same permission the normal form needs. A sales clerk who cannot record a fund transfer in the form cannot do it through the assistant either.

What each mode can do in Naffo AI.
Mode	Reads data	Writes data	Confirmation
Ask	Yes, within the user's access	No	Not needed
Do	Yes, within the user's access	Drafts changes for review	Required before saving, with the form's permission
Do with an agent (where enabled)	Yes, within the user's access	Stages each write in a multi-step job	Each staged write is reviewed before saving

Multi-step jobs, where agent support is enabled

Some jobs take several steps, for example checking stock for an order, drafting a purchase for the shortfall and then drafting the order itself. Where agent support is enabled, Do can hand a job like that to an agent, which breaks it into steps, does the reading itself and stages every write for review instead of saving it. You see the full set of proposed changes and confirm them. Agent support is enabled per deployment and can be limited to particular companies; ask during onboarding whether it is on for yours.

Briefs, suggestions and reading documents

Not every use of AI starts with a question. The AI daily brief summarises the business each day. Where the suggestions feed is switched on, it proposes actions for production, restock, collections, pricing, expiry and payments, each with an impact estimate you can open and check. Both are prompts for a person to look at the records, not decisions taken on your behalf.

Reading documents is the other everyday use. Naffo can scan screenshots or visiting cards into CRM leads, draft a quotation from a reference quote without copying old prices, and read a photo of a milk collection slip into rows a clerk checks before saving. In each case the output is a draft or a set of review rows, and a person saves it.

Roll it out by role

01Start with Ask for owners, admins and accountants
These are the roles Ask is open to. Use questions that map to existing reports, and compare answers for two weeks.
02Give Do to roles that already make the entries
A purchase clerk drafts purchases; an accountant drafts receipts. Keep the same approvals that apply to the normal forms.
03Use the agent for repeatable jobs, where enabled
Pick jobs with clear steps, such as stock checks and follow-up drafts, and review the staged writes each time.
04Review the audit trail
Check what was saved, edited and by whom, and adjust roles if the assistant is being used outside its intended scope.

For external AI clients, read how Claude connects to an ERP. To see the assistant on your own data, book a demo.

Frequently asked questions

Can Naffo AI save an entry without asking me?

No. Ask does not write at all. Do prepares drafts, and where agent support is enabled the agent stages its writes, and nothing is saved until a person confirms it. A draft also needs the same permission the normal form requires, so the assistant cannot do more than the signed-in user could.

Is the in-app AI the same as connecting Claude or ChatGPT?

They are two ways in. Naffo AI runs inside the app with Ask and Do modes, plus an agent where enabled. External clients such as Claude, ChatGPT or Gemini connect through MCP, where changes are logged and can be reverted. Both work on the same company data under the signed-in user's permissions.

What should a business check before relying on AI answers?

Compare the assistant's answers with standard reports for a few weeks, especially for receivables, stock and tax figures. Keep the accountant's review for anything filed or paid, and use the audit trail to see which records were saved and by whom.

This page also answers

Can an AI assistant create invoices in an ERP?
What is the difference between an AI chat and an AI agent in business software?
How should AI permissions work in an ERP?

Run this inside one system

naffo.tech is an all-in-one business management and manufacturing platform for Indian SMEs: versioned recipes and BOMs, batch-wise material issue under FEFO, yield, by-product and reason-coded wastage capture, in-process and finished-goods QC, batch traceability to dispatch, GST invoicing with e-invoice and e-way bill, and double-entry accounting that closes live.

Book a demo See the modules Estimate savings and a suitable software budget