ERP software, short for enterprise resource planning, is the system that runs a company's finance, payroll, inventory and supply chain from one place. It is often the single biggest technology purchase a company makes, and it has a reputation for going wrong. Independent studies have long put ERP failure or serious cost overrun rates somewhere between 50 and 75 percent, depending on how strictly failure is counted.
The most visible recent example is Birmingham City Council in the UK. Its move from an old SAP system to Oracle was originally budgeted at a fraction of its final cost, and the project is now set to cost more than seven times the earlier estimate, with the system still not fully working five years past its planned start date. The council declared itself effectively bankrupt partway through the rollout, a direct result of the ERP project spiraling out of control.
This is why ERP work is so dangerous for the people in charge of it. Chief information officers already have the shortest average tenure in the executive suite, staying roughly a year less than other top leaders, and a failed or delayed ERP project is one of the clearest ways a technology chief loses their job.
Into this already risky situation, every major ERP vendor is now pushing artificial intelligence that can act on its own. SAP calls its version Joule, Oracle has rolled out agents across its Fusion applications for finance teams, and Workday has expanded its Illuminate agents to handle HR and finance tasks. These tools are meant to approve invoices, flag unusual transactions, forecast cash flow, and let employees ask plain questions instead of running reports manually. Industry forecasts suggest this is not a minor feature: analysts expect a meaningful share of everyday finance decisions to be made without direct human involvement within the next few years.
The problem is that none of this works unless a company's underlying data is already in order. An AI agent making a purchase order decision or flagging fraud is only as good as the records it can see. If inventory counts are wrong, if customer records live in three different formats, or if departments have never agreed on what a single number means, the agent either freezes or acts confidently on bad information. That is a more dangerous failure than a human making a slow mistake, because it happens at speed and without anyone checking first.
This changes the calculation for any company planning an ERP project or upgrade. The AI features in the sales pitch are not something you simply turn on. They are the reward for years of unglamorous work cleaning up data, standardizing processes, and training staff to question what an automated system tells them. Companies that skip this step and jump straight to the AI layer are setting up the next Birmingham-style story, just with a faster, more confident failure mode.
The winners here are the vendors and the consulting firms that get paid either way, whether the project succeeds or turns into a multi-year rescue effort. The companies that actually gain an edge will be the ones willing to spend the next year on data housekeeping before they spend a dollar on AI agents. Everyone else is buying a more expensive way to repeat old mistakes.