Skip to main content

Perspectives · Guide

AI agents for ERP and CRM: where the change gets approved

AI agents can now change ERP and CRM systems directly: add fields, post entries, merge customer records. The question is not whether the agent can make the change. It is where the change is tested, who approves it and how it is undone. Put those three in place before any agent gets write access.

Author
M. K. Elango · Founder, elan1
Published
Reading time
4 min read
Hands at a keyboard beside a clipboard holding a printed report.

ERP and CRM systems are where a business keeps its commitments: what it owes, what it is owed, what it has promised its customers. That is what makes them attractive to agents and dangerous for them. A wrong answer in a chat window is embarrassing. A wrong change in the ledger is a restatement.

Two kinds of agent work

  • Changes to the system: a new field, a revised workflow, a tax template, a price list. One change affects every transaction after it.
  • Changes in the system: a purchase bill drafted, a journal raised, a lead updated, two customer records merged. Each change affects one transaction, many times a day.

Both need approval, but in different places. A change to the system needs testing against real data before anyone approves it. A change in the system needs a rule, and a person for the exceptions.

How ERP changes go wrong

ERP changes rarely fail on their own terms. They fail in the interactions: a discount applied after tax, a stock valuation method changed mid-year, a new ledger that the GST return does not pick up. Reading the change will not show you these. Replaying last month's transactions through it will. Our illustrative case of the discount field that broke a GST return walks through one example.

How CRM changes go wrong

CRM failures are quieter and often legal. A merge that combines two customers' consent records. A discount committed beyond the seller's authority. A campaign sent to people who withdrew consent, which under India's DPDP Act is a question of purpose and consent, not just etiquette. The checks for CRM are about consent, discount authority and merges, as much as data quality.

Where approval belongs

  1. On a copy: every change is built and tested on a copy of your system, never on the live one.
  2. Against your own data: the trial balance must still tie, stock must still reconcile, the GST return must still generate, access must not widen and the audit trail must stay intact.
  3. Before the commit: a person you name reads the result and says yes or no. The agent never holds the live keys.
  4. With a way back: every change can be reversed for a set period, and the reversal is tested too.

What agents should never do in an ERP

Some actions should be outside an agent's reach entirely. In elan1 Work, the agent will never make a payment, change bank details, post into a closed period or amend a filed return. It can prepare an e-invoice; a person files it. Each agent has its own identity in the ERP, never a shared integration user, so the software's edit log names it.

Which systems, and when

elan1 Work starts on Frappe/ERPNext and Twenty CRM, with Zoho CRM next. SAP and Salesforce are on the roadmap, and we will announce dates when they are firm. Tally and Zoho Books can be read for migration, never written to. Each of these carries its status on the elan1 Work page: designed or planned, never a promise dressed as a fact.

Key takeaways

  1. Separate changes to the system, which need testing on real data, from changes in the system, which need a rule and a named approver.
  2. Test every change on a copy against your own ledger, stock, tax, access and audit trail, then let a named person commit it.
  3. Keep payments, bank details, closed periods and filed returns out of the agent's reach, and keep a tested way back.

Product capabilities named here carry their current status (available, in pilot or designed) on each product page and in Trust and security. Figures marked “Design target” are specifications, not measurements.

About the author

M. K. Elango

Founder, elan1

M. K. Elango founded elan1 in Chennai to put AI agents to work inside regulated businesses, with a named person answerable for every action that commits them. He writes about pricing, approvals and what it takes for an auditor to trust an agent.

Meet the leadership team : M. K. Elango

Intelligence you can answer for.

Start with one process. Add the next when it holds.