Perforce Delphix Synthetic Data for AI software testing

Writing code faster does not solve the problem of software testing with useful data. A new feature may need customer records that do not exist yet, an unusual transaction history, or a combination of conditions that production has never produced. Someone still has to build those cases, and the records have to agree with one another when the application starts using them.

Perforce is addressing that work with Delphix Synthetic Data, a platform that uses AI to help generate realistic, scenario specific test datasets. It is designed for developers and AI agents working on new applications, features, and automated development workflows, particularly when production data is restricted, incomplete, or unavailable.

The part that interests me is the attention to relationships. Generating a table full of plausible names is straightforward. Generating customers, orders, payments, and related files that all point to the right records is a more useful test of the software. That consistency is what Perforce is trying to preserve across enterprise systems.

Test data has to behave like the real system

Referential integrity sounds like a database detail until a test fails because an order points to a customer who does not exist. A dataset can look convincing in isolation and still be useless once an application joins tables or moves a transaction between systems. Developers then spend time fixing the test setup instead of examining the code.

Perforce describes Delphix Synthetic Data as using AI to identify data structures, relationships, and business context, with statistical analysis helping reproduce the shape and distribution of the data. The aim is to reduce the manual configuration needed to produce related records that behave sensibly together.

Perforce Delphix Synthetic Data for AI software testing in practice

Users can describe the data they need through natural language prompts and adjust it for a particular scenario. That could mean creating data for an empty database or adding cases that are missing from an existing test environment. Perforce says this can reduce preparation from days or weeks to minutes, although the announcement does not provide a benchmark that would establish that result for every environment.

Consider a checkout feature that needs to handle a partially refunded order, an expired payment method, and a customer with several delivery addresses. Those are ordinary business conditions, but a convenient production sample may not contain the exact combination a developer needs. Building it deliberately gives the team a repeatable way to check the behavior.

This is also where synthetic data and masked production data serve different purposes. Masking can make an existing dataset suitable for broader use by protecting sensitive values. It cannot supply every situation that has never occurred. Perforce Delphix field CTO Ilker Taskaya makes that distinction in the announcement: historical records cover what happened, while testing also needs plausible cases that have not happened yet.

AI agents need dependable test environments too

Delphix Synthetic Data is part of the Delphix DevOps Data Platform, which combines data generation, masking, and delivery with centralized governance. Developers can work through the user interface or APIs, while MCP enabled workflows allow AI agents to request data as part of their development process.

That integration matters when an agent can write or revise code faster than a team can prepare its test environment. If each change waits for somebody to construct a dataset manually, the delay simply moves from coding to preparation. Providing data on demand could remove some of that waiting, provided the generated cases actually exercise the behavior the team wants to verify.

The platform supports bringing an organization’s own large language model so that processing can remain within its infrastructure. Perforce also says its AI discovery process examines metadata and schemas without exposing actual records or personally identifiable information to the model. Those are useful architectural details for teams deciding how the tool fits their existing data controls.

I would still want the people responsible for an application to define the business rules and expected results. A dataset can preserve every database relationship and still miss an important condition in a refund policy or account workflow. Generating the records is one job. Deciding whether they represent a worthwhile test is another part of the work.

Where developers should look for the benefit

A 2026 Perforce survey of 518 enterprise technology leaders gives some context for the launch. Among respondents who had evaluated existing synthetic data solutions, 34 percent said those tools provided referential integrity and 36 percent said they provided realism. Those findings come from a vendor survey, but they point to a recognizable problem: the value of generated data depends on whether an application can use it meaningfully.

IDC analyst Jim Mercer also emphasizes the need to produce realistic test data quickly, across complex environments, while meeting privacy requirements. It is a practical set of expectations. A tool that saves configuration time but creates hours of investigation into broken relationships has not improved the overall process.

For a development team, I would judge this by how easily it can reproduce a difficult case, keep related records consistent, and rebuild the same environment for the next test run. I would also look at how much manual correction remains after generation. Those details will say more about its usefulness than the AI label.

Delphix Synthetic Data addresses a real gap between producing software and having enough suitable data to test it. The opportunity is to make deliberate, repeatable test cases easier to create, especially for features and automated workflows that cannot depend on a convenient copy of production.