Why AI projects fail to deliver ROI — and how to avoid it
MIT researchers reported in 2025 that around 95% of organisations they studied were getting no measurable return from generative AI. The reasons are rarely the technology itself. They are choices about what to automate, how to measure it and how to roll it out.
The headline numbers
MIT's Project NANDA based its 2025 report on a review of more than 300 public AI initiatives, 52 interviews and 153 survey responses from senior leaders. It found most organisations were not seeing measurable returns, and that only a small fraction of custom enterprise tools reached production. Independent studies of individual tasks, meanwhile, regularly find real productivity gains (see what the research says). The gap between the two is where projects fail.
Failure 1: Starting with the tool, not the task
"We bought licences for everyone" is not a project. Without named tasks and owners, usage spreads thin and value is impossible to measure. Fix: start with one task, one team and one measurable outcome. See which tasks to automate first.
Failure 2: No baseline
If you do not know how long a task took before, you cannot prove it is faster after — and self-reported savings are unreliable. In METR's 2025 trial, experienced developers believed AI had sped them up by about 20% when it had actually slowed them down by 19%. Fix: measure time per task for one to two weeks before the pilot.
Failure 3: AI bolted on, not built in
A chatbot in a separate browser tab means copying information in and out. The time saved writing is lost moving data around. Fix: connect the AI step to where the work already happens — the inbox, the CRM, the document system — and redesign the workflow around it.
Failure 4: Ignoring the jagged frontier
AI performs very well on some tasks and poorly on similar-looking ones. When people apply it everywhere, errors on the weak tasks erode trust and cancel out gains elsewhere. Fix: test task by task, and set clear rules for when output must be checked.
Failure 5: Unmanaged adoption
Tools that are optional, unexplained and unsupported get used by enthusiasts and ignored by everyone else. Fix: train people on the specific workflow, appoint a champion, and track usage. Plan on 70–90% adoption, not 100%.
Failure 6: Costs that were never in the plan
Integration, data cleanup, usage charges and maintenance often appear only after go-live. Fix: include them up front — our list of hidden costs is a good checklist.
Failure 7: Freed time that goes nowhere
Saving two hours a week per person is worth nothing to the business if those hours are absorbed without anyone noticing. Fix: decide in advance what the time is for — more capacity, faster response, less overtime, delayed hiring — and track that outcome.
A simple success checklist
- One task, one owner, one metric
- Baseline measured before the pilot
- AI step built into the existing workflow
- Rules for when a person must check output
- All costs in the business case
- A plan for how freed time will be used
- A decision date: scale, adjust or stop
Sources
- MIT Project NANDA, "The GenAI Divide: State of AI in Business 2025", as reported by Virtualization Review (Aug 2025)
- METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (July 2025)
- Dell'Acqua, McFowland, Mollick, Lakhani et al., "Navigating the Jagged Technological Frontier", Harvard Business School working paper (2023), as reported by The Harvard Crimson