Synthesized creates test data and environments for enterprise AI agents

Synthesized builds test-data infrastructure for enterprise applications and AI agents. Its software generates, masks and subsets data so a team can exercise realistic business conditions without giving a test agent unrestricted access to live production records. The important distinction is between building the environment and judging the agent: Synthesized supplies the data and state on which a test runs, while the team’s evaluation system decides whether the resulting behavior is acceptable.

01 / ProductTest data becomes part of the environment an agent can act on

The Synthesized overview describes a platform spanning generation, masking and subsetting. Generation creates records for scenarios that may not exist in source data. Masking transforms selected values; subsetting produces a smaller coherent slice. Those methods address different needs, and a useful environment may combine them rather than treating every synthetic record as the output of one technique.

The Test Data Agent is explicitly labeled generally available. It maps a business scenario to records, relationships and system states, with provisioning through REST APIs or CI/CD. The page distinguishes the environment it supplies from the evaluation framework that runs and scores the agent. This is an important boundary: producing plausible data does not establish a passing release test.

The enterprise test-data offering connects that capability to existing application testing. Teams can refresh test environments as applications change. The value depends on preserving the parts of the business process that influence behavior, including relationships and failure conditions. A clean sample database is not equivalent to a realistic operating environment.

02 / AudienceA fit when linked enterprise state is the difficult part

Synthesized is relevant when an AI application must operate across systems whose records depend on one another. An inventory assistant may read stock, reservations and supplier status before proposing a replenishment action. A simple list of example questions cannot reproduce the state that makes a proposed action right or wrong. The test needs those dependencies to agree.

The product has particular emphasis on SAP and connected systems, but buyers should still validate their exact source types, versions and custom processes. A standard business process and a heavily customized installation can require very different preparation. The practical starting point is one process with a knowledgeable owner who can identify the records and exceptions that matter.

The Braintrust blueprint describes an adjacent evaluation layer for traces, datasets and scores. The BigID blueprint provides context for locating and governing sensitive information before it reaches development workflows. Neither task disappears when test data is generated: someone must define what data may be used and what behavior counts as success.

03 / WorkflowA proposed inventory-agent suite tests state, not just wording

Consider a proposed agent that recommends a transfer between two warehouses. It may prepare a transfer request but must not approve movement of reserved stock. Build a test environment containing products, locations, available quantities, reservations and pending transfers. The examples are synthetic and the agent’s actions terminate in test systems; this article does not describe a customer deployment.

Define the business outcomes first. A normal request should produce a feasible proposal. A stock balance that appears sufficient but is fully reserved should produce an explanation or escalation. A stale supplier record should not cause the agent to invent a delivery date. These outcomes identify the data dependencies the environment must represent before any model comparison is meaningful.

Use the platform’s generation, masking or subsetting approach according to each source. An existing application schema may be needed to preserve joins, while a rare reservation conflict can be generated deliberately. Keep the expected outcome in a separate test specification so it is not simply copied from the same generative process that created the input records.

Provision a fresh state for each repeatable run. If the first agent version leaves a pending transfer behind, the second version may see different stock and appear more cautious for the wrong reason. Reset or version the environment so the comparison isolates the application change. Record the data configuration alongside model, prompt and tool versions.

Run the agent through its usual orchestration and capture attempted tool operations. Judge both the proposed transfer and any mutation it attempted. An explanation that sounds careful cannot compensate for reserving stock in the wrong warehouse. The Test Data Agent supplies conditions for that test; the application team remains responsible for checking the operation against the business rule.

Add documents only under a confirmed scope. The Test Data Agent page says linked messages and documents are delivered engineer-led. The document-masking page likewise describes engineer-led PDF, DOCX and PPTX work, with one format per batch. Do not plan an unattended, self-service document pipeline on the strength of the structured-data release status.

Finish by rerunning a known failure after correcting the agent. Keep the failed and corrected traces with the exact environment version. If the result changes because the stock scenario was regenerated differently, it does not isolate the correction. A useful pilot demonstrates reproducible conditions and a defensible outcome, rather than simply producing a large quantity of test records.

04 / PricingPublished SAP packages are scoped annual commitments

The SAP pricing page lists four sales-priced packages sized by production data in scope and source count. Starter has a 12-month term; larger packages have 36-month terms billed annually. Annual billing is therefore not equivalent to a one-year commitment for every package. The table records published capacity and terms without inventing a currency amount.

That tariff is specifically for SAP test data. It should not be generalized to every Test Data Agent or document-masking engagement. Confirm those capabilities and their delivery work in the quote. The page also separates implementation, custom processes and extra sources from subscription pricing, and describes a fixed-fee validation after readiness rather than an unconditional free trial.

For the inventory pilot, define one source, one target and the business process before comparing costs. The size of the production estate may be much larger than the slice required for the pilot. Conversely, a small data volume can still require significant mapping across custom processes. Ask for a price and acceptance criteria that reflect both the data scope and the work needed to make the environment usable.

05 / DistinctionsBusiness-state preparation complements the evaluation stack

Synthesized’s useful distinction is its focus on the environment underneath an agent test. A model can answer a question plausibly yet make a poor decision once reservations, missing records or cross-system dependencies are introduced. Preparing those conditions helps the team evaluate behavior closer to the actual business process without giving the experiment live operating authority.

The private-deployment model also matters. The Test Data Agent page describes on-premises, private-cloud and hybrid operation under the customer’s controls. That can support organizations whose source data must remain within a chosen boundary. It does not remove the need to assess the generated dataset or configure the destination safely; deployment location and output suitability answer different questions.

PDF output needs written clarification. The document-masking page says in its limitations section that masked PDFs are images without searchable text, while its FAQ allows either image-based files or searchable files containing replacement values. Confirm the contracted format and inspect both rendered pages and any text layer. The page also says unsupported embedded Office content is refused. These distinctions affect applications that rely on OCR or searchable text.

06 / QuestionsConfirm the product boundary before building a release process around it

Which parts of the planned environment are standard capabilities and which require engineers to configure or deliver them? Request a demonstration with the actual source and document formats. A generally available structured-data agent does not make every related unstructured feature self-service. Record that boundary in the pilot so future teams do not assume it away.

How does the team verify masking coverage? The document page warns that a passed check does not prove nothing was missed and recommends labeled samples, especially for OCR. Test awkward layouts and unsupported embedded content. A readable archive or file is not evidence that every component was transformed successfully.

What makes repeated scenarios comparable? Determine how records, relationships and state are recreated and how failures are retained for regression testing. Also identify which business rules must be asserted outside the generator. This research establishes documented availability and commercial structure; it does not independently validate privacy, performance, SAP compatibility or the quality of generated data in a customer estate.

07 / DecisionBegin with a process whose correct outcome can be demonstrated

Synthesized is a worthwhile candidate when an AI test is limited by unrealistic enterprise data or unreproducible system state. Choose a business process with clear outcomes and a manageable set of dependencies. Prove that the environment can be rebuilt and that the evaluation distinguishes an allowed action from an invalid one.

For the warehouse example, a useful first result is the same reservation conflict producing a repeatable test across agent versions. Expand to additional systems and documents after the data relationships and delivery scope are understood. The investment makes sense when better environments expose meaningful failures that a prompt-only benchmark would miss.