
AI Makes Employees Faster. ERP Makes Businesses Work.
An employee can now write a customer proposal in ten minutes. That is real progress. But a proposal is not a sale, and a sale is not a business. Who approves the discount? Which stock is reserved? Which tax code applies? Which ledger entry results, and who can prove all of it to an auditor next year?
This is the first article in a seven-part series on why AI will not replace proper ERP, and why it will make ERP more important. The thesis is simple and we will repeat it in every article:
AI can improve a person’s output. ERP makes that output part of a controlled, repeatable, integrated business process.
AI changes how people work. ERP changes how a business operates. The winning model is not AI instead of ERP. It is AI embedded in a governed ERP process layer.
What is the difference between individual productivity and business productivity?
Individual productivity is how fast one person finishes a task. Business productivity is how reliably a whole organisation turns demand into delivered, invoiced and collected revenue. AI mostly improves the first. ERP is the system that determines the second.
A generative AI tool is excellent at the work of a single mind: drafting, summarising, analysing, translating, brainstorming, writing a formula or a query. It makes one employee faster. What it does not provide, by itself, is a shared data model, transaction integrity, process rules, permissions that follow roles, integrations between departments, and an audit trail that stands up to scrutiny. Those are exactly the things an ERP exists to provide.
| Capability | AI (the capability layer) | ERP (the operating backbone) |
|---|---|---|
| Core strength | Reasoning, drafting, summarising, predicting, recommending | Master data, transactions, workflow, permissions, audit trail |
| Unit of work | A task or a conversation | A business document moving through a process |
| Nature of output | Probabilistic and variable | Deterministic and repeatable |
| Who is accountable | Nobody, until a person adopts the output | A named role with recorded approval |
| Memory | The conversation or the prompt | The database of record |
Why does faster individual work not create a faster business?
Because work in a company is a chain, and a chain moves at the speed of its slowest link. When one step accelerates, the queue simply forms at the next step.
This is the productivity trap. If a salesperson produces five quotations in the time it used to take to produce one, the finance approver, the warehouse planner and the credit controller now face five times the volume with the same hours in the day. Review, approval, integration and decision-making become the bottleneck. The gain at the front of the chain is absorbed as congestion further down.
Research points the same way. A paper covered by MIT Sloan in April 2026, Chaining Tasks, Redefining Work, models work as sequences of interdependent tasks and argues that the value of AI depends on how those tasks are arranged, not only on how well AI performs any single one. The researchers also found that each handoff between AI and humans adds review and coordination cost, and that organisations see the gains only after they adapt their workflows and build capability. In other words, AI adoption is an organisational design problem as much as a technology purchase. ERP is where much of that organisational design lives in software.
A scenario: the five-minute quotation
A sales executive at a Singapore trading company pastes a customer enquiry into an AI assistant and receives a polished quotation in five minutes. The wording is excellent. Now the real questions begin.
- Is this customer within their credit limit, or are two invoices overdue?
- Does the price follow the current price list, the customer’s contract rate, and the volume tier?
- Is the stock actually available, or is it already reserved for another order?
- Which GST treatment applies, given the delivery address and the item?
- Does a discount of twelve percent need a manager’s sign-off?
- When the customer accepts, what triggers picking, delivery, invoicing and payment follow-up?
An ERP answers every one of these as part of creating the quotation, and enforces the answers when the quotation becomes an order. The AI wrote a good document. The ERP made it a controlled commercial commitment.
How should you measure individual speed against business speed?
The easiest way to see the gap is to put the two kinds of measure side by side. Most AI pilots report the left column. Most owners care about the right column.
| Individual measure (what AI moves) | Business measure (what ERP makes visible) |
|---|---|
| Minutes to draft a quotation | Days from enquiry to confirmed order |
| Emails answered per hour | Share of orders delivered on time and in full |
| Reports produced per week | Working days needed to close the month |
| Hours saved per employee | Cost per transaction and error rate |
| Tasks completed per person | Days from delivery to cash collected |
A company can celebrate the left column for a year while the right column does not move. That is not a failure of the AI tool. It is a sign that the process around the person cannot absorb the extra speed. Before approving any AI purchase, ask your finance head to nominate one right-hand number that the purchase should improve, and to name where it is measured today. If the answer is a spreadsheet maintained by one person, you have found your first ERP task.
A week in the life: where the speed disappears
Consider a ten-person trading company that gives every staff member an AI assistant.
- Monday. Sales drafts twelve quotations before lunch, up from four. The sales manager is pleased.
- Tuesday. Four customers accept. Each order must be keyed into the accounting package because the quotations live in documents, not in a system. The admin clerk, who did not receive an assistant, now has triple the typing.
- Wednesday. Two orders cannot be fulfilled because the same stock was promised twice. Nobody could see the double promise, because reservations are tracked in a shared sheet that was last updated on Friday.
- Thursday. One large order is held for credit approval. The approver is travelling. The order waits.
- Friday. The owner asks for the week’s margin. Finance needs two days to reconcile quotations, orders and supplier costs, because they sit in three places.
The company did produce more paperwork. It did not produce more delivered, invoiced and collected revenue. The extra speed was absorbed by re-keying, stock conflicts, approval delays and reconciliation. Every one of those is an ERP problem, and none of them is solved by a better assistant.
What is the strongest case for the AI-first view?
A fair reading of the opposing view makes the argument stronger, so here it is at its best. Supporters of the AI-first position say that traditional ERP is rigid, expensive to configure and unpleasant to use. They point out that AI can now read an email, understand intent, call a software interface and complete a multi-step task, so the screens and forms of ERP may fade into the background. Why train staff on menus when they can simply ask? And if the AI can talk to the database, perhaps a lighter system of record is enough.
There is real truth here. The user interface of business software will change, and natural-language access will become normal. Configuration, data loading, reporting and drafting will all get easier. We expect ERP to feel much less like a form-filling exercise within a few years.
But notice what the argument concedes. Even in the AI-first picture, the AI calls a system that holds the records, applies the rules and keeps the history. That system is the ERP, however it is presented to the user. The question is therefore not whether a system of record is needed, but how good it is. A thin, loosely governed one gives the AI poor data to reason on and weak limits to act within. A well-designed one makes the AI both more useful and safer. The AI-first view is right about the interface and wrong about the foundation.
Three ways this goes wrong, and how to avoid them
Before committing to any AI initiative, ask what total failure would look like, then design against it. Three failure modes appear again and again.
- AI islands. Each department adopts its own tools, holds its own copies of data and produces its own numbers. This is the shadow AI problem we describe in Shadow AI: The Governance Blind Spot. The mitigation is a short approved-tools list and a rule that business data lives in the ERP, not in a chat window.
- Speed without capacity. The front of the process accelerates and the back cannot keep up. The mitigation is to map the whole chain first and strengthen the slowest link before adding speed upstream.
- Tools before outcomes. The company buys capability and then looks for a use. The mitigation is the right-hand-column test above: no purchase without a business measure it is expected to move.
How would you explain this to a ten-year-old?
Imagine a school canteen. AI is a very fast cook who can make a sandwich in one minute. The ERP is the canteen system: the list of who ordered what, who has paid, what bread is left, and who is allowed a free lunch. A fast cook helps. But if nobody keeps the list, the sandwiches go to the wrong children, the bread runs out unnoticed, and no one can say who paid. A canteen needs both the cook and the list. That is the whole argument of this series.
What breaks without ERP
- The AI-written quotation lives in email and a spreadsheet, so nobody else can see the commitment.
- Credit limits, price rules and stock reservations are checked by memory, not by the system.
- Two salespeople promise the same stock to two customers.
- Approval happens in a chat message that cannot be reconstructed for an audit.
- Finance re-types the order into the accounting package, creating errors and delay.
What can AI do, and what must ERP do?
The cleanest way to hold the distinction is to split every task into a part that needs judgement and a part that needs control.
| AI can | ERP must |
|---|---|
| Draft the quotation and tailor the wording to the customer | Apply the correct price list, tax and currency, and record the quotation against the customer |
| Predict which customers are likely to pay late | Enforce credit terms, block or hold orders, and run the dunning workflow |
| Summarise a long supplier email thread | Keep the purchase order, receipt and bill linked as one traceable record |
| Suggest a reorder quantity | Create the purchase order only within budget and approval limits |
| Answer a manager’s question about margin | Hold the single, reconciled set of numbers the answer is drawn from |
What can AI not own?
Some responsibilities cannot be delegated to a probabilistic tool, however capable. Keep this checklist visible when anyone proposes replacing a system of record with a chat assistant.
- Legal accountability: a company, not a model, signs contracts and answers to regulators.
- Financial postings: journal entries need deterministic rules, period locks and reconciliation.
- Master-data ownership: someone must be responsible for the one correct version of each customer, product and supplier.
- Approvals: authority limits belong to roles, recorded with time and identity.
- Compliance evidence: auditors ask for records, not for the recollection of a chatbot.
- Transaction integrity: an order is either fully committed or not at all, never half-saved.
Where does this leave the AI-first argument?
The case for AI-first is not foolish, and it deserves a fair statement. Its supporters note that AI can now read documents, call tools and complete multi-step tasks, so why maintain rigid screens and forms? Gartner expects the shift to be rapid: in August 2025 it predicted that 40 percent of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5 percent in 2025. We agree with the direction.
But look at what that prediction describes. The agents are arriving inside enterprise applications. They are features of the system of record, not a replacement for it. The more actions an AI can take, the more a business needs a governed place for those actions to land. That is why we argue that AI raises the importance of ERP rather than lowering it. The next article, Business Processes Need More Than Intelligence, looks at what a real business process requires beyond a smart answer.
Frequently asked questions
Will AI replace ERP software?
Not in any form that keeps a business controllable. AI will change how people interact with ERP, through natural-language search, drafting and automation of routine steps. But the shared data model, transaction rules, permissions and audit trail still have to live somewhere, and that is the ERP.
If my staff already use AI tools, do I still need ERP?
Yes, if you want that individual speed to become business speed. Without ERP, faster drafting just moves the bottleneck to approval, stock, finance and delivery, where work still depends on spreadsheets and email.
Is ERP only for large companies?
No. Growing SMEs often feel the pain earlier, because they outgrow spreadsheets before they can afford errors. A modular system such as Odoo or MoxogoERP lets a smaller company start with sales, inventory and accounting and add more later.
Where should an SME start with AI and ERP together?
Start with a well-defined process that already runs in the ERP, such as supplier invoice handling, and add AI to read and classify the documents. Keep the posting, matching and approval inside the ERP.
What is the difference between an AI copilot and an ERP module?
A copilot helps a person think and write. An ERP module runs a business function, such as sales, inventory or accounting, with shared data, rules and records. A copilot can sit on top of a module and make it easier to use, but it does not replace the rules the module enforces.
How do I know if my company is losing the benefit of AI to its processes?
Look for symptoms such as staff re-typing data between tools, orders waiting for approvals, disagreements about whose numbers are right, and AI-drafted documents that nobody can trace to a transaction. If you see two or more, the process is absorbing the speed.
A question for you: where in your business is AI making individuals faster while the process still cannot absorb that speed?
Sources
By Willie Loo, Managing Director, Mxgsoft Pte Ltd and Moxogo Pte Ltd, Singapore. Part 1 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: list the three processes where your people complain most about re-typing data. Those are your best candidates for ERP integration before any AI is added. If you want a second opinion, talk to us about a Resource Master Audit.


