Why 42 per cent of enterprise AI projects are abandoned
New research from S&P Global Market Intelligence found that the share of companies abandoning most of their AI initiatives before they reach production has jumped from 17 per cent to 42 per cent in a year. Sidharth Mukherjee, Chief Strategy and AI Officer at CCI-Startek, sets out where projects break, in the data layer and in the organisation, and why stopping was often the right call.
The figure that has travelled furthest this year is a stark one. S&P Global Market Intelligence's Voice of the Enterprise: AI and Machine Learning, Use Cases 2025 study, published on 30 May 2025 and based on 1,006 midlevel and senior IT and line-of-business professionals across North America and Europe, found that the proportion of companies abandoning the majority of their AI initiatives before production had surged from 17 per cent to 42 per cent year on year. On average, respondents said 46 per cent of projects were scrapped between proof of concept and broad adoption.
One qualification matters before the number is used. The sample is drawn from organisations already investing in AI, so the 42 per cent is a failure rate among AI-active enterprises, not a cross-section of all companies. With that established, the more useful question is mechanical: what actually breaks, and when. We put that to Sidharth Mukherjee, Chief Strategy and AI Officer at CCI-Startek, a managed-services firm that describes its model as human-augmented AI. The exchange was conducted in writing.
Where does the 42 per cent figure come from, and what does abandoning most AI initiatives actually mean?
The figure is from S&P Global Market Intelligence's Voice of the Enterprise: AI and Machine Learning, Use Cases 2025 report, published on 30 May 2025. The survey covered 1,006 midlevel and senior IT and line-of-business professionals across North America and Europe, all from organisations that are actively investing in AI, so it is not a general business sample. It skews towards organisations already engaged with AI, which means the 42 per cent is a failure rate among AI-active enterprises rather than a cross-section of all companies. In the study's own words, the percentage of companies abandoning the majority of their AI initiatives before they reach production has surged from 17 per cent to 42 per cent year on year, with organisations on average reporting that 46 per cent of projects are scrapped between proof of concept and broad adoption.
How much of the failure traces back to the data layer, meaning quality, lineage, integration and access, and how much to the organisation?
Informatica's CDO Insights 2025 survey of 600 data leaders found that 92 per cent were concerned that their organisations are continuing to accelerate AI adoption even as they discover underlying problems with both data and organisational readiness. The top obstacles preventing generative AI initiatives from moving from pilot to production broke down as data completeness, quality and readiness at 43 per cent, technology maturity at 43 per cent, people at 35 per cent, process at 35 per cent and regulation at 34 per cent. Those were multi-select questions, not a forced ranking, so the figures do not add to 100 per cent, and a single failed project can be attributed to both data and people at once.
A data-quality example: we ran a pilot to automate first-draft proposal generation, training the model on past winning proposals. In the pilot the outputs looked strong. In production the model consistently produced proposals that reflected the firm's positioning from three years earlier, with outdated service lines, deprecated pricing and capabilities that were no longer relevant. The cause was data quality. The proposals sat in a shared drive with no version control, no metadata tagging by date or outcome, and no governance over what went into the training set.
An organisational example: a client wanted to reduce customer churn. The team built a churn-prediction model that identified at-risk customers accurately. In implementation the customer success team did not find it useful, because they already knew which accounts were at risk. What they needed was a recommendation engine for intervention, what to do about an at-risk account and when. The model solved the wrong problem. The data was fine. The failure was organisational: no one ran a structured discovery process with the end users before scoping the build.
Pilot to production is where most of these die. What specifically breaks in that transition?
The thing I see underestimated most often is integration complexity in a production environment. In the pilot, data is pulled manually or through a clean export, fed to the model, and the results are reviewed in a dashboard or spreadsheet. In production, the model has to read from a live system, write decisions back to a different live system, and do both within the latency tolerances. The integration work, the APIs, authentication, data contracts, error handling and retry logic, often turns out to be a larger scope than the model build itself, and it is usually not budgeted at the start of the pilot.
CCI-Startek sells implementation, so “the problem is implementation” is also a description of the service. Does the argument hold if you sold nothing at all?
CCI-Startek positions itself as a human-augmented AI services company. We do not focus only on implementation, we focus on end-to-end managed services, orchestrating AI agents and a human in the loop to reach the customer's outcome. Engagements start with discovery and joint solutioning, then implementation, then ongoing managed services. That said, the argument holds regardless of whether an organisation uses an external partner or does everything internally. Successful AI implementation needs more than a working model. It needs the right use case, clear business outcomes, integration into live workflows, organisational readiness and effective change management. Whether those capabilities sit inside the organisation or come from a partner is secondary. If the fundamentals are not in place, even technically strong AI is unlikely to deliver.
Here is the uncomfortable alternative. What share of that 42 per cent were right to abandon, because the technology genuinely did not work for the use case?
In most cases, stopping the pilot was the correct call at that point, notwithstanding the root causes I have described. If I had to put a number on it, I would say that in 60 to 70 per cent of cases it was the right decision.
If implementation is the binding constraint, why do the organisations with the deepest engineering talent and the biggest budgets report much the same failure rates?
Because engineering capacity and organisational readiness are two different things. Implementation is less about the size of the engineering team and more about organisational readiness and change management: whether the use case was chosen because it was genuinely high value or because it was politically safe and technically tractable, whether success was defined in business-outcome terms before the build started, whether the workflow the model would sit in was redesigned or the model was simply dropped into an unchanged process, and whether the people whose jobs would change were involved in the design or just presented with the output. Deep engineering talent can solve hard technical problems, but engineering capability alone cannot address those organisational challenges.
What is an AI initiative that should never have been started, and if the abandonment rate is unchanged in two years despite everyone following the implementation advice, what will you have misdiagnosed?
The board of a B2B services firm wanted its AI team to build an agentic orchestration platform, for the simple reason that all their competitors appeared to be building one. In most cases those initiatives were more for optics, and what looked like a competitor's product was really a white-labelled solution from a third party. The team should have done the research, benchmarked the competitor solutions and looked at where the agentic AI platform market is heading. Thousands of venture-backed startups building similar platforms suggests that in two years these will be commoditised and cheap to buy. The better move was to go back to the board with the research and an informed recommendation to partner or buy rather than build. Starting with the right problem statement matters before millions are spent on execution.
As for what I might have misdiagnosed: a possible gap is skills availability, and organisations trying to do everything themselves. Choosing the right technology and platform, and the right implementation partner, is also key, and a way to hedge against the risk of failure.