A lot of company leaders assume that old software is what's stopping them from using AI well. New research points the other way: the biggest source of failed AI projects is not old systems, it's the attempt to rebuild everything from scratch.
A widely discussed study from MIT looked at hundreds of company AI projects and found that most of them never produced a measurable financial return. The more useful finding sat one layer deeper. AI tools that companies built entirely on their own failed roughly twice as often as AI tools bought from an outside vendor and plugged into the company's existing data and systems. Buying and connecting beat building from zero.
Separate research backs this up from a different angle. RAND, a well known research group, found that more than eight out of ten AI projects fail to deliver what they promised, which is about double the failure rate of ordinary technology projects that don't involve AI at all. A survey of over one thousand five hundred companies by Publicis Sapient found a similar gap: most companies now use AI regularly, but only about one in ten call it truly core to how the business runs. Lots of companies have adopted AI. Very few have made it actually matter.
The math on tearing out old systems entirely is not kind either. Replacing a core banking system, the kind of software that runs accounts and transactions at a bank, typically costs between fifty million and two hundred million dollars and takes three to five years to finish. Roughly one in four of these replacement projects fails outright, and half of the ones that finish don't deliver the benefits they were built for. A well known case outside banking makes the same point: IBM and a major cancer hospital spent five years and sixty two million dollars trying to build an AI tool to help doctors choose cancer treatments, only to shelve the project before it was ever used on a real patient.
There's a practical reason old systems hold value that a new competitor cannot easily copy. Most of a company's most useful information, things like old contracts, claims, invoices, and customer records, sits in messy formats never meant for a computer to read. Estimates put this unstructured information at eighty to ninety percent of everything a typical company holds. A new company starting fresh doesn't have this history at all. An old company that finally organizes it has something a startup cannot buy or recreate quickly.
None of this means every old system should stay forever. If software is no longer supported, the vendor is out of business, or the law requires something the system truly cannot do, replacing it is the right call. But for most companies, the smarter and far cheaper move is adding AI in layers on top of what already runs the business, rather than betting the company on a full rebuild that the numbers say is more likely to fail than succeed.