The number that shows how fast this shifted
Synthetic data's share of enterprise testing rose from 14 percent in 2024 to an average of 25 percent in 2025, according to the Capgemini, Sogeti, and OpenText World Quality Report, which surveyed more than 2,000 respondents. That near-doubling in a single year reflects a broader pattern worth noting on its own: 35 percent of organizations surveyed now generate more than a quarter of their test data synthetically, meaning this has moved well past the experimental stage at a meaningful share of enterprises.
The underlying driver is straightforward. Generative AI has compressed synthetic data generation that once took weeks into a matter of seconds, removing the primary friction that used to keep synthetic data as a niche technique reserved for specific edge cases. Once that friction disappeared, adoption accelerated faster than the governance and evaluation practices around it, which is exactly the gap this report's experts are now flagging as the next challenge the industry has to solve before this growth outruns its own reliability.
Why speed created a new problem instead of solving the old one
Matt Kropp, chief AI officer at BCG X, put the core tension plainly: you can get synthetic data to be quite accurate, but it requires real effort to do that, and without a proper evaluation harness, you have no idea how it's performing in practice. That is a meaningful warning from someone advising enterprises on AI strategy, since it means the speed gains from generative synthetic data generation do not automatically come with the validation discipline needed to trust the output.
The risk this creates is largely invisible until it surfaces downstream. A test suite built on synthetic data that looks plausible but subtly misrepresents real-world data distributions can pass every test in a development pipeline while still producing a system that behaves incorrectly once it meets actual production data, a failure mode that is considerably harder to catch than a test that simply fails outright during development, and one that often only comes to light after the system is already live.
The reminder from Forrester that is easy to forget under deadline pressure
Forrester analyst Zeid Khater's observation, that at the end of the day, synthetic data is synthetic, sounds almost too simple to need saying, but it captures exactly the assumption that gets lost once synthetic data generation becomes fast and cheap. Teams under deadline pressure to stand up test environments quickly can start treating synthetic data as functionally equivalent to real production data, rather than as an approximation that carries its own distinct failure modes and blind spots.
Perforce CTO Ilker Taskaya's advice points toward the practical fix: you need to look at how the data is being used, and put the right controls in place, rather than treating synthetic data generation as a one-size-fits-all solution applied uniformly across every testing context. A single-feature developer test and an integrated test environment spanning hundreds of customers have genuinely different fidelity requirements, and conflating them is where synthetic data's limitations tend to surface.
The scale problem nobody budgets for upfront
Generating 50 synthetic records for a quick developer test is simple and cheap. Producing billions of rows that convincingly mirror a 60-terabyte production database for integration or performance testing is an entirely different proposition, one that drives up token costs and introduces efficiency tradeoffs most teams have not fully accounted for when they first scoped their synthetic data strategy. The gap between a small-scale pilot and a production-representative dataset is where budget estimates most commonly break down.
That gap also explains why adoption numbers alone can be misleading. An organization reporting that a quarter of its test data is synthetic may be running that generation almost entirely at the smaller, cheaper end of this spectrum, which tells you little about whether the organization has solved the harder, more expensive problem of generating synthetic data realistic enough to stand in for full-scale integration testing against a genuinely production-sized dataset.
What good evaluation practice actually looks like here
The experts cited in this research converge on a consistent recommendation: build an explicit evaluation harness that measures synthetic data against real production data on the specific dimensions that matter for each use case, rather than assuming visual or statistical plausibility is sufficient on its own. That means defining upfront what referential integrity, distribution shape, and edge-case coverage need to look like for a given test, and measuring the synthetic dataset against those criteria before trusting it in a pipeline.
It also means treating synthetic data governance as a distinct discipline from the data governance practices already in place for production data, rather than assuming existing data quality frameworks automatically transfer. Synthetic data introduces generation-specific risks, such as a model's own biases or blind spots propagating into the synthetic output, that a production-data-focused governance framework was never designed to catch, which is exactly the kind of gap a dedicated evaluation harness exists to close.
What this means for your own testing strategy
Before expanding your organization's synthetic data usage further, audit whether your current evaluation practice actually measures synthetic data quality against defined criteria, or whether adoption has outpaced that discipline the way this research suggests it has at many peer organizations. If you cannot point to a specific evaluation harness for your synthetic test data today, that gap is worth closing before scaling the practice further rather than after a production issue traces back to it.
Match your synthetic data investment to the actual fidelity each testing context requires, rather than applying the same generation approach uniformly across simple developer tests and full-scale integration environments. The organizations most likely to get genuine value from the speed synthetic data offers are the ones treating it as a technique requiring its own validation discipline, not as a free substitute for the real data it is meant to approximate.



