Research analysis
Analysis: the root causes behind failed AI projects
AI Cubed
August 9, 2026
8 min
RAND's report on the root causes of failure for artificial intelligence projects identifies a consistent pattern: projects fail because the problem was misunderstood, the data was inadequate, the team chased the latest technique rather than the useful one, the infrastructure to deploy and sustain it was missing, or the problem was not solvable with AI in the first place. That research is RAND's.
Our analysis takes each cause and turns it into a check we run during diagnosis — because every one of them is cheaper to catch before a build than after.
Cause one: the problem was misunderstood
The most common version we encounter is a project defined by output rather than decision — 'summarise the intake notes' rather than 'reduce the time between enquiry and qualified booking'. The first has no failure condition; the second can be measured, argued with, and improved.
Our check: state the decision that changes, who makes it today, and what the current cycle time is. If those three cannot be written in a paragraph, the project is not ready.
Cause two: the data was inadequate
Data problems are almost never discovered in a workshop; they are discovered when something tries to act on the data. Free-text fields where a picklist was assumed, duplicate records, a field everyone stopped filling in eighteen months ago.
Our check: pull two weeks of live records and run the intended logic against them by hand. The failure rate on that sample is the honest starting point.
Cause three: novelty over utility
A significant share of automation work does not need a model at all. Routing, validation, scheduling, and reconciliation are frequently rule problems, and rules are cheaper to run, easier to audit, and far easier to explain when something goes wrong.
Our check: attempt the rule-based version first. Reach for a model only where judgement or unstructured language genuinely sits in the path.
Cause four: the infrastructure to sustain it was missing
Building is the small cost. Monitoring, error handling, credential rotation, version changes in the upstream systems, and someone to answer the alert — that is where sustained value lives, and it is the line most commonly missing from a proposal.
- Where do failures surface, and who sees them?
- What is the manual fallback while a failure is being fixed?
- Who reviews the system quarterly against its original objective?
Cause five: the problem was not solvable this way
Some problems are organisational, contractual, or commercial, and no system will resolve them. Saying so early is the highest-value thing a diagnostic can produce, and it is why our first engagement is a diagnosis rather than a build.
Frequently asked questions
This page is AI Cubed's own analysis of independently published research. The underlying study belongs to its publisher, credited and linked in full below.
Sources
- The Root Causes of Failure for Artificial Intelligence Projects — RAND Corporation
Start here
See where your operation is losing time.
Twenty minutes with an operator, not a salesperson. We'll name the one bottleneck costing you the most — and tell you whether it's worth fixing with software at all.
Book your free 20-minute consult→20 minutes · video call · no preparation needed