Skip to main content
Business Processes Need More Than Intelligence: AI Can Propose, ERP Must Execute
Home / Blog / Digital Architecture & ERP
Digital Architecture & ERP

Business Processes Need More Than Intelligence: AI Can Propose, ERP Must Execute

A smart recommendation is not the same as an executable business transaction. An AI model can tell you that a supplier invoice looks correct. It cannot, on its own, post the payable, lock the period, prevent a duplicate payment and leave the evidence an auditor will ask for.

This is Part 2 of the series on why AI will not replace proper ERP. Part 1 set out the thesis: AI can improve a person’s output, and ERP makes that output part of a controlled, repeatable, integrated business process. This article looks at what a process actually needs.

What does a real business process require beyond intelligence?

A real business process needs roles, approvals, consistent master data, exception rules, a system of record and an audit trail. Intelligence helps decide what to do. These six elements make sure it is done correctly, by the right person, and can be proven afterwards.

  1. Roles and responsibilities. Someone is named as the requester, the approver, the receiver and the payer. The system knows which is which.
  2. Approvals and segregation of duties. The person who raises a purchase order should not also approve it and pay it. Controls exist because honest people make mistakes and dishonest people exploit gaps.
  3. Master data consistency. One supplier record, one product record, one chart of accounts. Every document draws from the same source.
  4. Exceptions and escalation. When a price is above tolerance or a delivery is short, the process routes the problem to someone with authority instead of failing silently.
  5. A system of record. There is one place where the official version of each document lives, with its history.
  6. Traceability and audit evidence. Who did what, when, on whose authority, and what changed.

None of these is a feature of intelligence. They are features of governance, and governance is what turns a good idea into a business act.

Where does AI genuinely help in a process?

AI is strongest at the messy, human-facing edges of a process, where the input is unstructured and judgement is useful.

  • Classifying incoming documents and emails.
  • Extracting fields from supplier invoices, delivery orders and receipts.
  • Predicting demand, lead times or late-payment risk.
  • Recommending the next best action to a person.
  • Answering questions from staff about policy or status.

These are valuable. They remove drudgery and improve decisions. But notice what each one produces: a proposal, a classification, a prediction. None of them is a commitment.

Where must ERP lead?

ERP must lead wherever an action changes the official state of the business: posting a journal entry, reserving stock, creating a purchase order, enforcing a credit limit, versioning a record and preserving accountability.

Process step AI can ERP must
Supplier invoice arrives Read the PDF and extract supplier, date, lines and totals Validate against the purchase order and receipt, apply tax, post the payable, flag duplicates
Customer orders Summarise the enquiry and suggest alternatives Check credit and stock, apply pricing rules, create the sales order, schedule delivery
Stock runs low Forecast demand and suggest a reorder quantity Create the purchase request within budget, route approval, track receipt
Employee claims expenses Read the receipt and propose a category Apply policy limits, route to the manager, post to the right account
Month end Draft the commentary on variances Enforce period locks, reconcile, and produce the reports from posted data

A scenario: the supplier invoice that looks right

A distributor receives a supplier invoice by email. An AI tool reads it in seconds and reports that the supplier, the total and the lines all look correct. Useful. Now the process takes over.

The ERP finds the matching purchase order and the goods receipt. The invoice quantity is 200 units but only 180 were received. The ERP does not post the full amount. It places the bill on hold, notifies the purchaser and the warehouse, and records the discrepancy. A week later the supplier sends a corrected invoice. The ERP recognises that the supplier, reference and amount are close to an earlier bill and warns of a possible duplicate before any payment is released.

The AI reduced the typing. The ERP prevented the loss. If the AI had been allowed to post directly, the company would have paid for 20 units it never received, quickly and confidently.

How do approvals and segregation of duties work in practice?

Segregation of duties means that no single person can start, approve, receive and pay for the same transaction. It is one of the oldest controls in business, and it is where an AI shortcut is most tempting and most dangerous. The table below shows a simple authority matrix for purchasing. The amounts are illustrative; set your own.

Step Who Control the ERP enforces
Raise a purchase request Any department user Budget code is required; the requester cannot approve their own request
Approve the order up to S$5,000 Department head Approver must differ from requester; approval is time-stamped
Approve the order above S$5,000 Director Routed automatically by value; escalates if not actioned in time
Receive the goods Warehouse Receipt quantity is recorded against the order
Record the supplier bill Accounts Three-way match against order and receipt
Release payment Finance manager Cannot be the person who recorded the bill

Now ask where an AI assistant fits. It can draft the request, suggest the supplier and read the bill. It should not be the approver, and it should not hold the combined permissions of requester, receiver and payer in one login. Giving a single agent broad access because it is convenient recreates, in software, the exact fraud risk that segregation of duties was invented to prevent.

Why does the system of record matter for audit and compliance?

A system of record is the one place where the official version of each document lives, together with its history. It matters for three practical reasons.

  1. Evidence. Auditors, tax authorities and banks ask for records, not explanations. In Singapore, GST-registered businesses and companies are generally required to keep accounting and tax records for a number of years, commonly five. Confirm the current requirement with your accountant, but the principle is stable: records must exist, be complete and be retrievable.
  2. Structured data. Singapore’s move towards e-invoicing through the InvoiceNow network, which is built on the Peppol standard, rewards businesses whose invoices already exist as clean structured data. A system of record produces that naturally. A folder of AI-drafted documents does not.
  3. Dispute resolution. When a customer says they were promised a price, or a supplier says an order was changed, the record shows what was agreed, when and by whom. Without it, the dispute becomes one person’s word against another’s.

An AI model can help find and explain a record. It cannot be the record, because it does not keep one in the sense an auditor means. Its output varies from one run to the next and its memory is whatever is in the conversation.

A second scenario: the order that arrives by chat message

Many Singapore SMEs take orders by WhatsApp. An AI assistant can read a message such as a customer asking for forty cartons of a product for Friday, identify the product, and draft a sales order. That is a real time saving, and it is exactly the kind of use we encourage.

Now follow the transaction. The ERP checks whether the customer account is active and within its credit limit. It checks whether forty cartons are available or whether part of the stock is reserved. It applies the customer’s agreed price, and if the customer asked for a discount in the message, it flags that the discount needs approval. It schedules the delivery run and creates the pick list. When the goods leave, it creates the delivery order and the invoice, and later matches the payment. If the customer then changes the quantity to thirty cartons by another message, the ERP records the change against the same order rather than creating a second one.

The AI did the reading and drafting. The ERP did everything that makes the order a reliable commitment. If the company skipped the ERP and let the assistant manage orders from a chat thread, the first conflicting message would produce an unresolvable argument.

Why are exceptions the real test of a process?

Any process looks good when everything goes right. A process proves its worth in the exceptions: the short delivery, the price above tolerance, the customer over their credit limit, the invoice with no matching order, the stock count that does not agree. Each of these needs an owner, a route and a record.

Exception AI can ERP must
Short delivery Summarise the gap and draft a note to the supplier Hold the bill, notify the purchaser, record the shortfall
Price above agreed tolerance Explain the likely cause from history Block auto-approval and route to the buyer
Customer over credit limit Draft a polite message to the customer Place the order on hold until released by an authorised person
Invoice without a purchase order Suggest a likely match Refuse to post until a valid order or an approved exception exists

Notice the pattern. In each case, AI makes the human response faster and better informed, while the ERP makes sure the exception cannot be silently ignored. A business that automates the happy path with AI but leaves exceptions to chance has only automated the easy part.

What breaks without ERP

  • Three-way matching is done by eye, so short deliveries are paid in full.
  • Duplicate invoices slip through because nobody can search all supplier history quickly.
  • Approvals are given by chat message and cannot be proven later.
  • The AI tool’s extraction is stored in a personal folder, not in the books.
  • When the auditor asks who approved the payment, the answer is a guess.

Why can the AI not simply be given the keys?

The honest answer is that AI output is probabilistic. It is usually right, and that is precisely the danger. A process that is right ninety-nine times in a hundred will, at volume, produce a steady stream of errors, and errors in finance, inventory and compliance compound. Deterministic controls exist to catch the one in a hundred.

The industry itself is cautious about this. Gartner predicted in June 2025 that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value or inadequate risk controls. The third reason is the one an ERP addresses most directly. Controls are not an obstacle to automation. They are what make automation safe enough to keep.

What AI cannot own in this process

  • The decision that a payment is authorised.
  • The posting of the accounting entry.
  • The definition of what counts as a valid supplier.
  • The record that proves the three-way match happened.
  • The accountability if the payment was wrong.

The key message: AI can propose. ERP must execute, record and govern. In Part 3 we look at the asset underneath all of it, the database.

Frequently asked questions

Can AI post accounting entries directly?

It can prepare them, but posting should pass through the ERP’s rules: valid accounts, open periods, tax logic, matching and approval. The AI proposes the entry and the ERP decides whether it can be posted.

What is three-way matching and why does it matter?

It compares the purchase order, the goods receipt and the supplier invoice before payment. It stops overpayment for goods not received and is one of the oldest and most effective financial controls.

Does adding AI to invoice processing remove the need for approvals?

No. It reduces data entry, not accountability. Approval limits and segregation of duties still apply, and in many companies matter more when volume rises.

Which processes are best for a first AI pilot?

High-volume, low-risk ones with clear data, such as document classification and invoice extraction, where the ERP validates and posts and a person reviews exceptions.

What is segregation of duties?

It is a control that prevents one person from controlling every stage of a transaction, such as requesting, approving, receiving and paying. It reduces both error and fraud. The same principle should apply to AI agents.

Can an AI agent act as an approver?

It can prepare the information an approver needs and can pre-check a request against policy. The approval itself should come from an accountable person or from a rule the business has explicitly defined and can audit.

A question for you: which step in your payables process would hurt most if it were wrong once every hundred invoices?

Sources

By Willie Loo, Managing Director, Mxgsoft Pte Ltd and Moxogo Pte Ltd, Singapore. Part 2 of 9 in the series Why AI Will Not Replace Proper ERP.

Two companion articles extend this series: why agentic tools cannot replace enterprise data governance and audit trails, and how personal AI workflows differ from cross-department ERP control.

Next step: pick one document type, such as supplier invoices, and write down every check that happens between receipt and payment. Each check is either a control that belongs in your ERP or a gap. Related reading: Where to Put the Human in the Loop.