September 4, 2026

AI for the Over 40 – Week 33: Why AI Won't Replace ERP (And What It Actually Threatens)

One of our salespeople asked me to join an upcoming prospect meeting. The prospect had a question they wanted answered before moving forward with an ERP evaluation: Why should we invest in an ERP system if AI is going to replace it? 

My first reaction was to laugh. Not at the prospect, at the question. After 32 weeks of working intensively with AI, studying its capabilities and limitations, building frameworks for when to use probabilistic versus deterministic systems, and documenting every breakthrough and failure along the way, the idea that AI would replace ERP felt like asking whether calculators would replace accounting. 

But then I caught myself. 

I have spent the last eight months writing about exactly this kind of moment, the moment where confidence outruns analysis. I explored this in the paradox of expertise, where I confessed that I had spent years giving clients the Microsoft answer, not because I had rigorously evaluated alternatives, but because the habit of confidence had outlasted the conditions that justified it. I wrote about the difference between lying and bullshitting, and how the most dangerous form of bad advice comes from people who genuinely believe they are right but have not bothered to check. 

Was I doing it again? Was my dismissal based on 32 weeks of earned understanding about what AI can and cannot do, or was it based on the fact that I have spent 23 years selling ERP? 

I sat with that question longer than I expected, and what I found convinced me that both things can be true. My reasoning is sound, and I owe it to the prospect, and to you, to show the work rather than just deliver the conclusion. 

So let me do what this series has always tried to do: take a question seriously, examine it honestly, and share what I actually found. 

The strongest case for “AI kills ERP” 

Before I defend ERP, let me give you the best argument against it, because if I cannot articulate why smart people are asking this question, I have no business answering it. 

The pitch goes something like this. Modern AI can read unstructured documents such as invoices, purchase orders, contracts, and emails. It can understand business context. It can apply rules. It can make decisions. It can execute actions across systems. 

If AI can read a purchase order, match it to available inventory, apply your pricing rules, generate the appropriate journal entry, and post it, why do you need a rigid, expensive ERP system built on relational database architecture that has not fundamentally changed in 30 years? 

The supporting evidence is real. AI already handles natural language queries against business data. It automates document processing. It manages approvals. It generates reports. It writes code. Startups are pitching AI-native business operations platforms that promise to do everything ERP does without the implementation pain, the consultant fees, or the change management headaches. 

The argument is essentially this: ERP is the mainframe of the 2020s. Everyone knows it will eventually be replaced. The only question is when. 

I understand why that argument is compelling, and I understand why a prospect evaluating a significant ERP investment would want someone to address it directly before writing the check. 

Here is why the argument falls apart, and why the real answer is more interesting than either yes or no. 

What ERP actually is 

Before we can ask whether AI replaces ERP, we need to agree on what ERP actually is. And I think most people, including many who use ERP every day, confuse the thing they interact with for the thing itself. 

Strip away the vendor marketing and the implementation jargon, and ERP is really three things. 

The first is a database, a structured, relational store of every transaction your business has ever executed. Orders, invoices, payments, shipments, journal entries, and inventory movements are all connected, all indexed, and all queryable. This is your system of record, the single authoritative source of what actually happened in your business. 

The second is a set of rules, the business logic that determines how transactions are processed. When you post a sales invoice, the system does not just record it. It applies posting groups to determine which general ledger accounts are affected, calculates tax based on jurisdictional rules, updates accounts receivable, adjusts the customer’s credit limit, and flows cost of goods sold to the income statement. These rules enforce consistency. The same type of transaction produces the same accounting treatment every time, regardless of who initiates it or when. 

The third is an interface, the pages, menus, fields, reports, and workflows that humans use to interact with the database and the rules. This is where you enter a purchase order, run an aging report, or approve a journal entry. 

That is it. Database, rules, interface. 

Now, here is what becomes immediately clear once you see it this way: when people say AI will replace ERP, they are really talking about the interface, and they are largely right. AI is already transforming how humans interact with business systems, and that transformation will accelerate dramatically. Natural language will replace menu navigation. AI will handle document intake. Conversational queries will replace report builders. 

But nobody, not the startups, not the vendors, not the consultants, is seriously arguing that businesses can operate without a structured database of financial transactions. Nobody is arguing that we can eliminate the rules that ensure consistent, auditable accounting treatment. Nobody is suggesting that tax calculations, posting groups, or revenue recognition schedules should be handled probabilistically. 

The “AI replaces ERP” argument, once you break it down, is really “AI replaces the interface layer of ERP.” That is not a prediction. It is already happening. It is also not a threat to ERP, it is an enhancement that makes ERP dramatically more accessible. 

The interface was always the weakest part of ERP anyway. It is what drives training costs, consulting fees, user frustration, and change management budgets. If AI can replace all of that with natural language interaction, ERP does not become less valuable. It becomes more valuable to more people, more often. 

But let me be thorough about why the database and the rules cannot be replaced, because the reasons go deeper than most people realize. 

Five walls AI cannot walk through 

1. The deterministic problem: your general ledger doesn’t need judgment 

In deterministic versus probabilistic systems of this series, I spent weeks researching the distinction between deterministic and probabilistic systems. It is the most important concept for understanding why AI and ERP serve fundamentally different purposes. Deterministic systems, like ERP, operate on Boolean logic: if this, then that, with the same input producing an identical output every time. When a general ledger entry posts, there is zero tolerance for ambiguity, and the system cannot guess debit and credit amounts. 

Probabilistic systems, like AI, calculate probability distributions. They excel at judgment, context, and handling unstructured information, but identical prompts can produce different responses. That is not a bug, it is the feature that makes AI valuable for tasks that require flexibility and interpretation. 

The question is not whether AI is powerful. It is. The question is whether probabilistic reasoning can replace deterministic processing for financial transactions that require legal, regulatory, and audit compliance, and the answer is no. Not because AI is not good enough yet. Because the fundamental architecture is wrong for the job. Your general ledger does not need judgment. It needs certainty. 

2. The system of record problem: what would a world without one look like? 

Here is a thought experiment. Forget about ERP for a moment and ask a simpler question: what would it actually look like to run a business without a system of record? 

Your CFO needs to close the books at month end. Where is “the truth”? If AI processed transactions throughout the month, interpreting invoices and making judgment calls about categorization while applying business rules probabilistically, where do you go to get a definitive answer about what happened, not what AI thinks probably happened, but what actually happened? 

Your bank wants to evaluate your creditworthiness for a line of credit and asks for financial statements. Statements are prepared from what, exactly, a collection of AI decisions that may or may not be reproducible? A conversation history? 

Your company gets audited, and the auditor wants to trace a revenue transaction from the source document to the financial statement. They need to see the original order, the shipment, the invoice, the payment, and how each step flowed through your general ledger according to consistent rules, and they need to see that the same rules applied to this transaction applied to every other transaction of the same type. Where does that audit trail live if not in a system designed specifically to maintain it? 

Tax authorities want to know how you calculated your sales tax liability for the third quarter, and the answer needs to be reproducible, because if you are audited in three years, you need to be able to demonstrate that your calculations were correct at the time they were made, not that AI would probably reach the same conclusion today. 

The “AI replaces ERP” pitch does not just require better technology. It requires reinventing the entire concept of verifiable financial truth, and it requires every regulatory body, tax authority, auditor, bank, and investor to agree that probabilistic answers are acceptable substitutes for deterministic ones. That is not a technology problem. It is a civilization problem. 

3. The context problem: the most expensive transaction in history 

This one is personal. After 32 weeks of working intensively with AI, I have learned more about context windows, token consumption, and the practical limits of AI infrastructure than I ever expected, and that experience reveals something about the “AI replaces ERP” argument that I have not seen anyone else point out. 

Consider what AI would need to know to process a single, routine purchase order in Business Central. Not a complex one, just an ordinary purchase order with three line items from a domestic vendor. 

To process it correctly, AI would need, in its context, the vendor’s master data, such as payment terms, currency, default locations, and posting groups, along with the item master data for each line item, including unit of measure, costing method, inventory posting group, general product posting group, and tax product posting group. It would also need the posting group matrix that determines which general ledger accounts get hit when this combination of vendor and item transacts, the tax setup for this jurisdiction, the dimension rules that might require cost center or department codes, the approval workflow rules that determine whether this purchase order amount triggers an approval requirement, the warehouse setup and whether this is a directed put-away location, the currency exchange rate if applicable, the current inventory levels if the system needs to check availability, any blanket order this purchase order draws against, and the vendor’s purchase price agreements or discount structures. 

That is not an exhaustive list. That is a simplified list for a routine transaction. 

Now consider that Business Central handles all of this through relational database lookups, indexed queries that retrieve exactly the data needed in milliseconds, using table relationships that have been defined once and apply consistently forever. The context is not loaded into memory for each transaction. It is structured into the system architecture. The posting group matrix does not need to be understood; it is executed. 

For AI to do the same work, every bit of that context would need to be in the conversation, and every transaction would require loading the equivalent of the entire relevant portion of your business configuration, just for one purchase order. A mid-sized company might process hundreds or thousands of transactions per day, including purchase orders, sales orders, inventory transfers, journal entries, payment applications, and bank reconciliations, each requiring its own full-context load and consuming tokens, while also requiring AI to understand relationships that ERP simply enforces. 

I have spent weeks hitting the practical limits of what AI can hold in context, which is one of the reasons I have been exploring infrastructure concepts like MCP servers and RAG systems. The idea of loading enough business context for AI to reliably process even routine transactions at enterprise scale is not just impractical today, it is architecturally absurd. You would be asking AI to rebuild, from scratch, in every conversation, what ERP maintains permanently in structured tables. 

It is the difference between asking someone to memorize your entire tax code versus looking it up in a reference book. The reference book does not need to be smart. It needs to be structured, indexed, and always available. That is what ERP is. 

4. The transactional integrity problem: What happens when step three fails? 

In ERP, I documented a critical architectural requirement that most people never think about until something goes wrong: transactional boundaries. When you post a sales shipment in Business Central, a cascade of things happen simultaneously. Inventory quantities decrease, cost of goods sold flows to the income statement, accounts receivable updates, and the customer’s available credit adjusts. If the shipment involves multiple warehouses, bin contents update, and if there is serial or lot tracking, item tracking entries are created. 

All of these things happen as a single atomic transaction. Either all of them succeed, or none of them do. If step four of a six-step process fails, the system rolls back steps one through three. Your inventory never shows as shipped, while your financials show it still in stock, and your accounts receivable never shows an amount owed, while your general ledger shows no corresponding revenue. 

This is not a feature, it is a fundamental requirement called ACID compliance, standing for atomicity, consistency, isolation, and durability. It is what prevents your business data from ending up in an inconsistent state when something goes wrong, and things go wrong constantly, from network interruptions to data validation failures, concurrent user conflicts, and timeout errors. 

AI has no concept of transactional boundaries. It does not do atomic commits, and it does not roll back. If you asked AI to process that same shipment and it successfully updated inventory but failed to post the financial transaction, you would have a discrepancy between your physical inventory and your financial records, and you might not discover it until your next physical count or audit. 

The “AI replaces ERP” pitch implicitly assumes that every transaction succeeds. Anyone who has spent time in enterprise systems knows that assumption is fantasy. The value of ERP is not just processing transactions correctly; it is maintaining data integrity when transactions fail. 

5. The shared state problem: whose version of truth are you in? 

Here is one more that is easy to overlook until you think about what it actually means to run a business with multiple people making decisions simultaneously. When a warehouse worker in Denver ships an order, the inventory change is immediately visible to the sales rep in Chicago who is checking availability for a different customer. The credit manager in New York can see the updated receivable, and the CFO pulling a cash flow report sees the projected payment. All of them are looking at the same data, in the same state, at the same time. 

ERP manages shared state across the entire organization in real time. It is the reason two salespeople cannot accidentally promise the same last unit of inventory to different customers, assuming the system is configured correctly, and it is the reason the warehouse and the general ledger agree on what is in stock. 

AI conversations are stateless by nature, and each session starts fresh. There is no persistent, shared, organization-wide state that every user sees simultaneously. Even if you connected AI to a database, you would be building an ERP system, just with different branding. 

The shared-state problem is the most underappreciated aspect of ERP. It is not glamorous, and nobody gets excited about concurrent user management and record locking, but it is what allows fifty or five hundred people to operate on the same set of business data without creating chaos, and it is completely absent from every AI-native business platform pitch I have seen. 

What this actually means 

So the prospect’s question, why should I buy ERP if AI will replace it, runs into not one objection but five: deterministic processing, system of record, context economics, transactional integrity, and shared state. Any one of these would be sufficient to argue that AI enhances ERP rather than replaces it. Together, they make the case overwhelming. 

But if I stopped here, this would just be a defensive article from an ERP consultant protecting his revenue stream, and I have spent 32 weeks trying to be more honest than that. 

The prospect’s intuition is not entirely wrong. Something fundamental is changing. They are sensing a real shift; they are just misidentifying what is being replaced. AI is coming for the interface layer, and while the pace of adoption will be governed as much by people and change management as by technology, the economic implications are significant. 

To understand what is actually happening, I need to introduce you to a 160-year-old economist who predicted all of this. 

The Jevons paradox 

In 1865, William Stanley Jevons, an English economist, published a book titled The Coal Question that documented something no one expected. Steam engines had become dramatically more fuel-efficient. Common sense said England would use less coal, since better engines using less fuel per unit of work should mean lower coal consumption. Simple math. 

Except the opposite happened. Coal consumption exploded. 

More efficient engines did not reduce demand for coal; they made coal-powered applications economically viable in contexts nobody had imagined. Factories that could not afford to run on steam suddenly could. New industries emerged, and new uses for steam power were invented. The per-unit savings were overwhelmed by the sheer volume of new applications. 

Jevons captured the principle: when you make a resource dramatically more efficient to use, you do not use less of it. You use more, sometimes exponentially more. 

This pattern has repeated throughout history. More fuel-efficient cars did not reduce total miles driven; people drove more. Energy-efficient refrigerators did not reduce energy consumption; people bought more refrigerators. Cheaper data storage did not reduce storage needs; we stored exponentially more data. 

You might have encountered the concept recently without knowing the name. When Chinese AI startup DeepSeek released a model that matched American competitors at a fraction of the cost, Microsoft CEO Satya Nadella invoked the Jevons paradox in response. His argument was that cheaper, more efficient AI would not reduce demand for computing; it would make AI viable in so many new contexts that total demand would skyrocket. 

Stanford economist Erik Brynjolfsson has been applying it to AI and jobs — specifically to the labor market. He identifies three conditions that must all be true for Jevons’ paradox to hold: the technology makes workers more productive at the task, that productivity translates into lower effective costs for the output, and demand is elastic enough that lower costs trigger significantly more consumption. 

Brynjolfsson’s favorite example is airline pilots. Jets made pilots dramatically more productive, allowing them to fly farther and faster while carrying more passengers. Did that mean we needed fewer pilots? No. Cheaper air travel meant more people flew, demand exploded, and we needed more pilots, not fewer. 

Now apply those three conditions to ERP. 

The Jevons paradox applied to ERP 

Condition 1: Does AI make people more productive at using ERP? 

Absolutely. Natural language queries against business data, AI-assisted reconciliation, automated document processing that feeds structured data into ERP transactions, and Copilot summarizing customer history before a sales call are all capabilities that make ERP more productive to use. 

In the Intake-Engine-Bridge framework, I built a framework I call Intake-Engine-Bridge, with AI as the intelligent intake layer that handles unstructured input, ERP as the deterministic engine that enforces business rules, and connectors as the bridge to external systems. That architecture makes the entire system dramatically more productive than ERP alone. 

Condition 2: Does that productivity lower the effective cost of getting value from ERP? 

This is the critical insight. The cost of getting value from ERP has never been primarily the software license. It is the training, the consulting hours, the change management, the user frustration, the reports that take three days to build, and the data that sits in the system but requires an expert to extract. 

AI drops all of those costs. When a warehouse manager can ask a question in plain English instead of learning how to build a filtered view in Business Central, the effective cost of that insight approaches zero. When an accounts payable clerk can photograph an invoice and have AI extract the data instead of manually keying every field, the cost of clean data entry drops dramatically. 

Condition 3: Is demand for ERP capability elastic? 

This is the question that determines everything. I believe the answer is overwhelmingly yes, and my evidence comes from my own experience just two weeks ago. 

Below-the-line work: Jevons’ paradox in my own inbox 

I’ve previously written about a category of work I called “below-the-line” work — analysis that would be genuinely valuable but that nobody can justify the hours to perform. I had years of weekly compliance reports sitting in my email as attachments, dozens of spreadsheets. A colleague asked how many times they had submitted late time over the past year. The data existed, and the analysis was straightforward, but I simply was not going to do it. The effort of downloading each attachment, opening each spreadsheet, consolidating the data, and building a pivot table could not be justified against everything else competing for my time. 

AI dropped the effort threshold. It took one hour, including building a Power Automate flow to extract the attachments and having Claude analyze the consolidated data, for an analysis I never would have performed. 

That is Jevons’ paradox at the individual level. I did not do less analysis when AI made it easier. I ran an analysis that wasn’t happening at all. The demand was always there, latent, suppressed by friction. Remove the friction and the demand surfaces. 

Now multiply that across every person in an organization who has questions they have stopped asking because the ERP is too hard to navigate, the report is too complex to build, or the data requires consolidation across too many sources. AI does not reduce the need for ERP data. It unlocks demand for ERP data that was always there but could not be economically accessed, and that unlocked demand requires more data in the system, more accurate data, more integrated processes, and more robust business rules, in other words, more ERP capability, not less. 

The evidence was already in this series 

What strikes me now is that I have been documenting Jevons’ paradox for 32 weeks without knowing its name. 

In revolutionizing solution design, a client project estimated at 60 hours was prototyped in 2 hours using AI-assisted development. The result was not less development, it was that a project nobody would have approved at 60 hours became obvious at two. AI made previously uneconomical development suddenly viable, which increased demand for Business Central customization. 

I introduced the concept of operational debt — problems people have stopped calling problems. The diagnostic questions I developed are essentially an inventory of latent demand, since every question people have stopped asking becomes askable when AI drops the friction. 

I built an Outlook-to-Todoist integration that did not exist, not because I was doing task management badly, but because I was not doing it at all, at least not in a way that worked. AI made it possible to build what didn’t exist, which meant more capability consumed, not less. 

The pattern is consistent: every time AI reduced the cost of doing something with my tools, I did more with those tools, not less. I never once found myself thinking, “Now that AI is helping, I need fewer systems.” I found myself thinking, “Now that AI is helping, I can finally use my systems the way they were designed to be used.” 

When the paradox does not apply, and why honesty matters here 

Jevons’ paradox is not a universal law. It has a critical precondition: demand must be elastic. When it is not, efficiency gains do not produce more consumption; they just eliminate jobs. 

Brynjolfsson’s cautionary example is farming. Tractors and modern agriculture made farmers dramatically more productive, but we can only eat so much food, and demand for food is inelastic. The result was not more farming, it was fewer farmers. Agricultural employment dropped from 40 percent of Americans in 1900 to less than two percent today. 

So the honest question becomes whether ERP is more like coal or more like food. I believe it is like coal: the business appetite for better data, faster decisions, deeper analysis, and more granular insight is effectively unlimited, currently suppressed only by friction. But I want to be transparent about why this matters. The answer determines not just whether ERP survives, but what happens to the people whose jobs revolve around it. 

The part nobody wants to talk about: jobs 

Jevons’ paradox predicts that total demand for ERP capability increases. It does not promise that every existing role survives unchanged. 

Some work gets compressed. Manual data entry, basic report generation, routine reconciliation, and simple inquiry responses, the kind where someone logs into Business Central, navigates to a page, reads a number, and reports it back, are the kind of work AI absorbs almost entirely, and it is already happening. 

Some work expands. Analysis and interpretation, exception handling, process design, and strategic decision-making informed by previously inaccessible data all fall into this category, along with analysis nobody had time for — work that was not happening because the effort threshold was too high. This is the new work that emerges when friction drops. 

Some work emerges entirely. AI-ERP integration architecture, designing the intake layers for unstructured-to-structured workflows, building the prompt frameworks that help non-technical users extract value from business data, and managing the complex interplay between probabilistic AI systems and deterministic business systems are roles that barely exist today. They will be essential within five years. 

The transition is not painless, and I will not pretend otherwise. The person whose job is 80 percent data entry and 20 percent judgment calls is in a fundamentally different position than the person whose job is 20 percent data entry and 80 percent judgment calls. Jevons’ paradox says total demand for ERP work increases, but the composition of that work shifts dramatically toward judgment, analysis, and design. 

This is why I have argued throughout this series for building AI superhumans rather than waiting for AI agents. Enhancement over replacement is not just a philosophical position; it is a practical strategy for navigating exactly this transition. The organizations that invest now in building their people’s AI literacy are the ones whose employees will be positioned for the expanded and emerging work. The organizations that wait until compression is obvious will scramble. 

What my prospect actually needs to hear 

Here is what I plan to say in that meeting. The question is not whether AI will change ERP, because it will. The technology already exists for natural language to supplement menu navigation, for AI to handle the intake of unstructured information, and for reports to be generated conversationally. But technology capability and organizational adoption are two very different timelines. We are still dealing with people, and people change at people speed, not technology speed. The interface will evolve significantly, but not overnight and not without deliberate investment to help people make the transition. 

But the engine underneath, the database that stores your financial truth, and the rules that ensure it is processed consistently, does not go away. It cannot be because the technology is not ready, but because the regulatory, legal, and financial requirements for deterministic processing are fundamental to how businesses operate. 

The analogy I would offer is that AI replacing ERP is like suggesting that, because Siri can make phone calls, we no longer need cell towers. The interface changed dramatically. The infrastructure did not go anywhere. 

What AI actually does is make ERP more valuable by dropping the friction cost of extracting that value. A 160-year-old economic paradox predicts exactly this: make a resource more efficient to use and you do not use less of it, you use dramatically more. 

So yes, invest in ERP, but invest simultaneously in AI literacy for the people who will use it, because the organizations that combine robust ERP infrastructure with AI-literate employees will not just survive this transition. They will discover capabilities they did not know they had, answer questions they had stopped asking, and extract value from their business data that was always there, just waiting for the friction to drop far enough that someone would finally look. 

The bottom line 

The prospect’s question, “Why should I buy ERP if AI is going to replace it?”, contains a prediction that gets the direction exactly backward. AI is not going to replace ERP. AI is going to make ERP so much easier to use that organizations will demand more from it than they ever have before. The friction that has been suppressing the value of business data is about to drop dramatically, and when it does, the organizations with robust ERP infrastructure and AI-literate employees will have an overwhelming advantage over those with neither. 

The real question is not whether to invest in ERP. It is whether you are simultaneously investing in the AI literacy that will allow your people to unlock the value that has been sitting in your systems all along, waiting for someone to finally ask the question. 

A 160-year-old economist already told us what happens when you make a powerful resource easier to use. We just were not listening. 

This post is part of my “AI Over 40” series. It first appeared on LinkedIn: Why AI Won’t Replace ERP (And What It Actually Threatens) 

Read more AI and Copilot blogs.

Trending Posts