In a previous post, we argued that AI kills best-of-breed SaaS. Too many specialized tools, too much fragmentation, the AI starved of context.
Some readers pushed back: "That is why ERPs exist. One system. One database. Problem solved."
They are half right. ERPs were built to solve fragmentation. But they solved it by creating a different, arguably worse, problem.
AI kills ERPs too. For different reasons.
The ERP promise
The pitch was compelling. Stop buying separate tools for finance, procurement, HR, inventory, and operations. Buy one system that does all of it. One vendor. One database. One contract.
SAP, Oracle, NetSuite, Microsoft Dynamics. They built empires on this promise. And for a long time, it worked well enough. The data was in one place. The reports were consistent. The CFO could see a single number.
The price was rigidity.
The customization trap
Every company operates differently. Your approval workflow has four steps. Your competitor has two. Your procurement process routes through legal. Theirs does not.
ERPs handle this with "configuration." In theory, you adjust the system to match your business. In practice, you adjust your business to match the system.
SAP implementations take 6 to 18 months. The average mid-market ERP project costs $150k to $750k. Half of that is consultants configuring the system to approximate your actual workflow. The other half is training your team to work the way the ERP expects them to.
And the moment you upgrade to the next version, half of your customizations break.
This is the customization trap. The ERP is rigid by design. It has to be, because it serves thousands of companies with the same codebase. Your specific business logic is a rounding error in their product roadmap. You file a feature request. It ships in 18 months, or never.
Why AI breaks the ERP model
ERPs assume that business processes are stable and predictable. You define the workflow once, configure it in the system, and it runs the same way for years.
AI assumes the opposite. AI is useful precisely because processes change, exceptions happen, and context matters more than rules.
An AI agent that can read your entire business data and act on it needs flexibility. It needs to understand that this particular invoice should be flagged because the client changed their payment terms last week, even though the standard workflow would auto-approve it.
An ERP cannot give the AI this flexibility. The processes are frozen in configuration. The data model is locked into the vendor's schema. The "customizations" are brittle layers on top of a system that was not designed to be modified.
Ask an AI agent to do something that falls outside the ERP's configured workflow and it hits a wall. The data is there, but the system will not let the agent act on it in a way the configuration did not anticipate.
The ERP has the data. But it has it in a cage.
The cost of the cage
ERP vendors charge for the cage and then charge again for the key.
Need a custom report that the standard module does not support? That is a consulting engagement. Need to change how approvals route for one department? That is a configuration project. Need to integrate with a tool the ERP does not support natively? That is middleware, plus a maintenance contract.
The per-seat costs are just the beginning. SAP charges $100+ per user per month. Oracle is in the same range. Then you add implementation, customization, training, and the annual maintenance fee that is typically 20% of the license cost.
A 50-person company on a mid-market ERP is spending $200k+ per year, all-in. And they are still working around the system instead of with it.
The irony: companies adopted ERPs to reduce the cost of fragmentation. They ended up paying more for rigidity than they ever paid for scattered tools.
What AI actually needs
The argument for ERPs was "one database, one system." That part was right. AI does need unified data. We wrote about this in The AI Bolt-On Fallacy: an agent that can only see one silo is useless.
But AI does not need the rest of the ERP package. It does not need frozen workflows. It does not need 18-month implementation cycles. It does not need a $500/day consultant to change an approval rule.
What AI needs is:
A shared database where every record is accessible and the schema is flexible enough to evolve with the business.
Permissions that are structural, not configured through a vendor's admin panel. Role-based access defined once, enforced everywhere, for humans and AI agents alike.
The ability to build exactly what you need, not configure the closest approximation of what you need from a menu of pre-built modules.
An audit trail that logs every action, human or AI, at the data level.
The third way
Best-of-breed gave you perfect tools that could not talk to each other. ERPs gave you one system that could not bend.
There is a third option: shared infrastructure that you build on.
Instead of buying an ERP that does everything badly, you build the exact internal tools your company needs. A procurement tracker that matches your actual approval workflow. A billing system that handles your specific invoicing rules. An inventory tool that reflects how your warehouse actually operates. Not a configuration compromise. The real thing.
Every tool shares the same database. The same authentication. The same permissions. The same audit trail. The AI agent sees all of it and can act across all of it, because there are no frozen workflows blocking it.
This is what RootCX provides. One shared PostgreSQL database, SSO, role-based permissions on every resource, immutable audit logs. You build your internal tools with Claude Code, Cursor, or RootCX Studio. You deploy with one command. Every app inherits the infrastructure.
The difference from an ERP: you own the code, you own the data model, and your AI agent can act on any record without hitting a configuration wall. The difference from best-of-breed: everything shares one database, so the AI sees the full picture.
The new math
An ERP costs $200k+ per year for a 50-person company, and you spend the first 6 months configuring it to be 80% of what you need.
Building on shared infrastructure, you spend the first week shipping the exact tool you need. The second week, you ship a second tool that reads data from the first. By month two, you have an operational stack that matches your business, not a vendor's idea of your business.
The code is yours. The data is yours. The infrastructure is open source. If you outgrow it, you modify it. No consultant required. No 18-month upgrade cycle.
ERPs solved fragmentation by replacing it with rigidity. AI needs the opposite: unified data with flexible tools. The ERP era served its purpose. It is time to build what comes next.
RootCX is the infrastructure for what comes next. Open source, self-hostable, free to start.