AI-enabled technologies have taken the world of software development by storm. More than 90% of professional developers have integrated AI into their core workflows, while the usage of agentic coding assistants nearly doubled over the past year. Software testing is one of the best-suited use cases for AI agents in development. Agents can write, update and run tests in a fraction of the time it’d take humans, and they can generate novel scenarios and examine a much larger number of edge cases.
But as AI agents and assistants take on more of that work, enterprises’ test data needs are changing quickly. Most vendors haven’t kept pace with this transformation. This market gap may explain what otherwise seems to be a contradiction among the findings of a recent survey of 518 enterprise technology leaders conducted earlier this year on behalf of Perforce Delphix. Two out of every three (66%) decision-makers surveyed said their organizations do not use synthetic data in software development and testing. Many of these leaders do not feel synthetic data meets their enterprise’s requirements for realism and referential integrity.
Agentic development requires more test data, faster, than manual development processes did. At enterprise scale, agentic development also demands that test data have cross-system referential integrity and maintain statistical realism. Until now, vendors have offered synthetic data solutions that had one property or the other. Delphix Synthetic Data is the first to deliver both.
Capable, efficient and data-hungry: AI’s role in software development
Assistance from AI can speed up software development by as much as 2x, but only if the agents are kept busy. With enterprise adoption of coding agents skyrocketing and vendors shifting toward consumption-based pricing, it is rapidly becoming cost-prohibitive for AI agents to sit idle. To avoid this situation, teams will often employ low-quality or stale data in testing, or, even worse, use production data that hasn’t been secured.
“We have long heard complaints about how long it took for test data to become available,” says Mayank Ahluwalia, senior product manager for Delphix Synthetic Data and data compliance for AI at Perforce Delphix. “Especially if it was being copied from production. But today, with the rise of agentic development, the costs are higher than ever. Enterprises simply can’t afford to have their agents wait for production data.”
Synthetic data meets this need because it can be delivered almost instantly. It can also be created at scale, which is important when workflows involve large numbers of agents working in parallel. AI-enabled development also requires test data that is highly varied.
“When agents are writing test cases, they’re not constrained by the limits of human imagination,” says Ahluwalia. “As humans, we can only test the scenarios we can think of, and our ideas are shaped by our biases. Agents use combinatorial logic to come up with a much wider variety of novel scenarios. Teams need test data that’s suitable for all of them.”
However, agentic development demands test data that accurately represents the real world. To do this, synthetic data must have two key characteristics:
- Referential integrity: Data has referential integrity when all of its components stay correctly connected across systems. In an e-commerce company, for instance, every record in the order management system should link to an account in the customer management system. A human tester might notice something wrong with an order when there’s no customer attached, but agents will keep testing and may even tag these failed test cases as successful.
- Statistical realism: When data is used for model training, it must have statistical properties that mirror those of production data; otherwise, the model will learn inaccurate information. Salary data, for example, should have the same spread, minimum and maximum as the real data.
To date, tools built for use in software development and testing have typically delivered data with some degree of referential integrity after extensive configuration, while those built for model training have focused on statistical realism. Among the survey participants who had evaluated existing synthetic data solutions, only 36% said the data had statistical realism. Even fewer (34%) said the synthetic data possessed referential integrity.
Born-in-AI solution overcomes the bottleneck
A newly launched solution, Perforce Delphix Synthetic Data, was built specifically to close this market gap. Applying multiple AI techniques, including natural language processing, machine learning, large language models and agent skills, this solution automatically identifies data structures, relationships and business context across data sources, without the manual configuration that legacy synthetic data tools require. Developers and AI agents can generate realistic, scenario-specific data in seconds, using a prompt-based interface.
“Our solution looks at the schema to understand what information each field holds and what that information means,” explains Ahluwalia. “It uses this understanding to generate data with referential integrity, and it samples the data to learn its statistical distribution. The included chat assistant is not just bolted on as it would be in a legacy solution. Instead, it can actively guide you by making recommendations about how to best generate data for the types of scenarios you need to test.”
The governance and compliance benefits of using synthetic data are undeniable, but reasons to employ synthetic data stretch far beyond them.
“Historically, people have considered synthetic data a compliance solution,” says Ahluwalia. “When you look at it that way, you leave a lot of the benefits on the table. If you can create synthetic data to expedite agentic development and testing, you’ll start reaping the full benefits of your investments in AI.”
Interested in learning more? Explore Delphix Synthetic Data or read the latest synthetic data research.
