An AI agent workforce in operations: what the agents do, and where a person must decide

It is not a chat window on your database. An agent in our framework is a small program tied to one event or one clock. One wakes when a purchase order is approved. Another runs every night. A third runs only when a buyer presses a button and asks. Each is switched on per agent and per business, and a business can hold an agent to a tighter limit than we set, never a looser one.

What an agent produces is a proposal. It carries the finding, the numbers behind it and the record it is about. A fixed rule checks every proposal before it is stored, and a proposal type with no rule behind it is refused. Nothing an agent writes becomes a business record on its own, and a proposal nobody acts on expires after seven days.

These are jobs the agents in Neauron Intelligent Core, our ERP, are built to do, each switched on only when a business chooses. Each is work a person could do and rarely has time for.

1. Reading documents. A supplier's bill is classified, read, matched to the vendor by GSTIN, checked against the purchase order and the goods received, and left either as a prepared draft or as a case that needs attention, with the reason. A suspected duplicate always goes to a person. A second agent reads the quotation a vendor attached to a bid and reports where it disagrees with what the vendor typed, such as 1,095 a tonne in the form and 1,195 on the quote.
2. Forecasting. Each night an agent reads every project with an approved budget and raises the ones whose cost forecast has moved past a line. Below 15% complete it reports a range, never a single figure, because an early confident forecast is the one that loses everyone's trust. Another agent turns the night's reorder suggestions into draft requisitions, one per warehouse, ready when the buyer arrives. The should-cost modeller revises market benchmarks monthly and says so when there are too few data points to mean anything.
3. Exceptions. One rejected receipt is a local fact. The third short delivery from the same vendor in a month, across two sites, is a pattern no single storekeeper sees. An agent counts across receipts and names it. Another checks, the moment someone raises a requisition, whether the same material is already sitting in a different warehouse.
4. Approvals. The payment-run assembler builds the week's run from settlements already approved, and flags the ones a person should read first. It recomputes no figure, holds read access only, and cannot start a payment. It also refuses to run at all until the bill-matching agent has a full quarter of history with no rejected proposal.

The rule on our site is short. A person approves every move that touches money. In practice these decisions never belong to an agent:

1. Releasing a payment. Above a set value, it takes two people.
2. Submitting a requisition or awarding an order. Both commit budget.
3. Raising a debit note or a return against a supplier. These carry legal weight.
4. Serving a contract notice. An agent can say the notice clock is running and draft the words. A person decides to serve it.
5. Changing a budget, a baseline or a published forecast.
6. Changing which AI model an agent uses, or how much it may spend.

When a person accepts a proposal, that person writes the business record, under their own login, through the screen they would use anyway. The agent service has no way to write it for them. So the answer to "who approved this payment?" is always a name.

Every agent run is logged with what it cost, in rupees. Each agent has a daily spending ceiling, checked before a model is called rather than after. Personal data is replaced with tokens before any prompt is built, and the count goes on the run log, so anyone can check whether personal data left the system.

Then we count. For each agent, what share of its proposals did people reject? An agent earns more room only on evidence, eighty per cent of its drafts approved unedited over four weeks, and loses it after a reversed action.

One agent watches the others. Each day it looks for an accepted proposal with no business record behind it, an agent proposing when it was only allowed to observe, a spending ceiling breached, and an agent switched on that has not run in a fortnight. Its findings go to whoever owns the process. Each agent's flags, drafts and savings are reported monthly, and one that is not paying for itself is switched off rather than defended.

1. Can you show me the list of agents, each with its one job, and switch one off while I watch?
2. Which decisions can an agent never make, and where is that enforced in the code rather than in a policy?
3. When I accept a proposal, whose login writes the record?
4. What does each agent cost a month, and who sees that figure?
5. What happens to personal data before it reaches a model?

A builder who answers all five with a screen rather than a slide has probably run agents in production.

Receivables, bank reconciliation and month-end close are next, and none of them runs today. For receivables we would draft the follow-up on overdue items and explain each deduction, with write-offs left to finance. For the bank we would propose matches between bank lines and open items. For the close we would assemble the reconciliations and draft the journals, with posting and tax still done by rules.

Start with one process and one number, such as days from goods receipt to a matched bill. Put one agent on it in draft mode for a month and read its proposals every week. If people accept most of them, give it more room. If they don't, switch it off. We build agents on the same framework for businesses in India, the UAE and South Africa, and each one follows these rules from its first day.