AI is the biggest bet most enterprises are making right now. Most of them are losing it, and the model has almost nothing to do with why. AI works best at very narrow definition of scope and iterated to find ways the process breaks so AI can learn to handle exceptions by itself.
MIT’s Project NANDA looked at 300 deployments and found 95% of enterprise generative AI pilots produced zero measurable return. RAND put the number at 80% across 2,400 initiatives, roughly twice the failure rate of ordinary IT work. The numbers are staggering but in some ways not unexpected. When buzzword implementation comes from top down, projects get rolled out without the full foresight on what goals truly are.
The failure lies in scope, not technology. The world has seen the shift with coding, but software programmers have always led with frontier technology. The tools service their day to day needs and now they manage ideas and convert the ideas into instructions which LLM then go write code around. Other departments now have to go about considering what their role is, and what their thought process is and document to similarly use LLMs to gain efficiency in their day to day work.
How companies can be more successful is by having each employee identify one painful, manual, or very tedious part of their workflow that is repetitive and then break it down into steps and describe the process to solve it. This is the instruction manual for LLM to then build a process around gathering, cleaning and analyzing the data to produce an effienct process that can show the return on AI.
Companies that identified a specific workflow to automate actually saw success. LegalZoom put generative AI on legal document drafting, a repetitive task with clear quality criteria and a baseline to measure against. Samsara did the same with fleet operations paperwork. Both saw measurable gains inside 90 days. DXC and Rimini Street started with a single defined workflow before letting agents touch anything else.
Back-office document automation, procurement, and risk review are where the returns land, not the customer-facing chatbot that needs to handle millions of potential user variations. Data readiness is the key barrier with only 12% of companies saying their data is clean enough to use. A narrow workflow lets you fix the data for that one task instead of the whole company.
So if you are scoping an AI project this quarter, narrow it down to just one workflow for the department. For most companies the whole point is an efficient bottom line, and that is the line between the 5% that pay off and the 547 billion dollars that did not.