Why AI projects fail, and how to make artificial intelligence pay for itself
The studies disagree on the number but agree on the diagnosis: what sinks an artificial intelligence project is rarely the model. It is the wrong problem, the data that isn't there, and the routine nobody changed.
In short
- Independent studies show that most AI projects do not deliver the expected result: RAND cites a failure rate above 80%, BCG found only 4% of companies generating substantial value, and Gartner predicted that at least 30% of generative AI projects would be abandoned after proof of concept.
- The main cause is not technical. Under BCG's 10-20-70 rule, success depends 10% on algorithms, 20% on technology and data, and 70% on people and processes.
- The seven most common causes: a poorly defined problem, data that can't support the promise, a pilot with no path to operations, AI outside the real workflow, adoption ignored, unrealistic return expectations, and no owner or governance.
- AI pays for itself when it is applied to a process the company already understands, with reliable data, a business metric set in advance, and people in charge of the decisions that call for judgment.
- The safe plan has five moves: pick the problem by its value, check the data, design the workflow with the people who will use it, pilot against the business metric from day one, and only then scale.
What the research says, without the alarm
The numbers vary because each study measures something different, but they all point the same way.
| Study | What it found | What it names as the cause |
|---|---|---|
| RAND Corporation, 2024 | Estimates that more than 80% of AI projects fail, twice the rate of IT projects without AI. Based on interviews with 65 experienced data scientists and engineers | A misunderstood problem, insufficient data, focus on the technology instead of the problem, weak infrastructure, problems too hard for AI |
| BCG, October 2024 | Only 22% of companies have moved past proof of concept and generate some value. Just 4% generate substantial value | The 10-20-70 rule: 10% algorithms, 20% technology and data, 70% people and processes |
| Gartner, July 2024 | A prediction that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025 | Poor data quality, inadequate risk controls, escalating costs and unclear business value |
| MIT NANDA, 2025 | 95% of organizations with no measurable return on generative AI | Tools that do not learn from context or fit into the workflow |
The historical parallel helps. In 2020, before the generative AI wave, BCG had already shown that 70% of digital transformations fall short of their objectives, for the same reasons: vague strategy, absent leadership, fragile governance. Artificial intelligence did not create this pattern. It made it more expensive and easier to see.
The seven causes that sink an AI project
1. The wrong problem was chosen
The project starts from "we need to do something with AI", not from a business pain with an owner and a value attached. RAND puts this cause first: technical teams and leadership do not share the same understanding of the problem to be solved. The result is a technically correct model that answers a question nobody asked.
The sign: nobody can say, in one sentence, which business number will change if the project works.
2. The data can't support the promise
Duplicate records, incomplete histories, information scattered across personal spreadsheets, different rules in each department. AI learns from what exists. On bad data, it produces wrong answers that look precise, which is worse than having no answer.
The sign: the first stage of the project is "gather the data", and it never ends.
3. The pilot has no path to operations
Proofs of concept are built to impress: hand-prepared data, an isolated environment, a dedicated team. None of that exists in the real operation. Without a budget, an architecture and an owner defined from the start for the next phase, the pilot becomes a slide deck.
The sign: the project has a date for the demo and no date for going into use.
4. The AI was left outside the real workflow
A tool in another tab, one that requires copying and pasting, loses to habit. This is what the MIT NANDA report describes as the main barrier: solutions that do not fit into the work or learn from the company's context. The intelligence has to show up inside the system the person already uses, at the moment they decide.
The sign: people tried it, liked it, and a month later nobody opens it.
5. Adoption was treated as a detail
Buying licenses is not transformation. Every change brings some perceived loss: of mastery, of autonomy, of relevance. If the people who will use it had no part in the design, were not prepared and do not see what they gain, the project meets a quiet, polite resistance. This is where the 70% in BCG's rule lives.
The sign: the training was a video sent by email.
6. The expected return is unrealistic
"We want to see a return in 90 days" pushes teams toward cases that are showy and shallow. Productivity gains are real, but hard to convert directly into financial results, as Gartner itself notes. Without a baseline measured beforehand, not even success can be proven.
The sign: the project's metric is model accuracy, not a business indicator.
7. No owner, and no governance
Who answers when the AI gets it wrong with a customer? Who approves the use of personal data under privacy laws such as the GDPR and Brazil's LGPD? Who tracks usage costs, which grow with volume? Without those answers, legal stops the project the day before launch or, worse, nobody stops it.
The sign: the questions about risk come up for the first time in the final approval meeting.
Where AI pays for itself
The pattern in projects that work is almost the reverse of the list above. They apply AI to processes the organization already understands, with high volume, available data and a clear business metric.
How to prioritize use cases: business value on one side, data readiness on the other.
Three traits show up again and again in the cases that pay off:
- The AI is inside the workflow. The suggested reply appears on the support screen, the conversation summary goes straight into the CRM, the forecast sits on the dashboard leadership opens every Monday.
- People are in charge. What can run on its own does, under clear rules. The decisions that call for conversation and judgment stay with people, and the AI works for them.
- The gain has an owner. Someone from the business, not from technology, answers for the indicator the project promised to move.
A plan to get AI out of the pilot
- ProblemA business pain with an owner, a value and a metric
- DataDo the sources for this problem exist, and can they be trusted?
- WorkflowDesigned with the people who will use it, inside the current systems
- PilotShort, in the real operation, measuring the business indicator from the start
- ScaleGovernance, costs, team readiness and continuous improvement
First the people and the problem. Then the process and the data. Then the technology.
1. Pick the problem by its value, not by the technology
List the pains leadership already knows. For each one, estimate the value at stake and identify who answers for the result. If nobody wants to own the indicator, the problem is not a priority, however interesting the technology.
2. Check the data before you promise
For the chosen problem, check whether the data exists, where it lives, who maintains it and how good it is. This check takes days and saves months. If the data can't support the case, the right project is to fix it, and that is already progress: the same foundation will serve revenue operations, the dashboards and the next automations.
3. Design the workflow with the people who will use it
Sit next to the person who does the work today. Where the AI comes in, what it suggests, what the person confirms, what happens when the system is wrong or doesn't know. Decide also where the solution lives: adapting an off-the-shelf tool or building your own component is a strategy choice, and we cover it in the guide on custom or off-the-shelf software.
4. Pilot in the real operation, against the business metric
A short pilot, with real users, real data and a comparison group when possible. Measure the baseline first. Count every cost: integration, usage, human review, team preparation. The pilot ends with a decision: scale, adjust or stop. Stopping early on a case that doesn't pay is also a result.
5. Scale with governance and follow the adoption
Define owners, usage limits, handling of personal data, and monitoring of quality and cost. Prepare the teams before the switch, not after. And stay close: adoption is measured weeks after launch, when the novelty has worn off and the old habit tries to come back.
Before you approve the next AI project: ten questions
- Which business indicator will change, and by how much?
- Who, outside the technology team, owns that indicator?
- What is the baseline as measured today?
- Does the necessary data exist, and does anyone trust it?
- On which screen, in which system, will the person find the AI?
- What happens when the AI is wrong, and who notices?
- Did the people who will use it take part in the design?
- Is there a budget and an owner for the phase after the pilot?
- Have the legal and privacy risks been assessed?
- Would simple automation rules solve the same problem for less?
If more than three answers are "we don't know", the project is not ready to start. And finding that out beforehand is the best return an hour in a meeting can give.
Frequently asked questions
What is the failure rate of AI projects?
It depends on the study and on what counts as failure. RAND Corporation (2024) cites estimates that more than 80% of AI projects fail, twice the rate of IT projects without AI. BCG (2024) found only 22% of companies moving past proof of concept and 4% generating substantial value. MIT NANDA (2025) reported 95% of organizations with no measurable return on generative AI, a figure whose methodology has been challenged. The consensus is in the diagnosis: most projects never reach a business result.
Why do so many AI projects get stuck in the pilot?
Because the pilot is designed to prove that the technology works, not that the business improves. It runs on hand-prepared data, outside the real systems, and without the people who would use the solution every day. When the time comes to integrate, train teams and take on operating costs, there is no budget, no owner and no metric to justify going on.
How do you measure the ROI of an AI project?
By setting a business metric, not a technology metric, before the project starts: customer response time, hours freed up in a routine, error rate, conversion, retained revenue. Measure the baseline, run the pilot with a comparison group when possible, and count every cost, including integration, operations, human review and training. Model accuracy is a technical indicator, not a return.
Where should a company start with AI?
With a process the organization already understands well, with high volume, reasonably clear rules, available data, and a cost of error that is low or easy to contain with human review. Common examples: sorting and answering frequently asked questions, summarizing and logging conversations, classifying documents, forecasting demand. Messy processes should be fixed before they get AI.
Are generative AI and automation the same thing?
No. Automation executes defined rules: if this happens, do that. It is predictable and cheap to audit. Generative AI interprets language, produces text and handles cases the rules did not foresee, at the cost of some unpredictability. The best results come from combining them: automation for the workflow, AI for the points that call for interpretation, and people on the higher-risk decisions.
Do I need all my data in order before starting with AI?
All of it, no. The data for the problem you chose, yes. A well-scoped project only needs quality in the sources it uses, which makes the cleanup feasible. Waiting for a perfect database paralyzes the company, and ignoring data quality produces wrong answers that look precise.
From reading to practice
AI where it pays for itself. Never as decoration.
At Dynamis Works, reliable data comes before any promise, and no project ends at delivery: we stay with the adoption until the new way becomes the normal way. If your organization wants to take artificial intelligence out of the pilot and into results, let's talk.
Talk to Dynamis Works
