Syntho combines masking, rules and AI for usable synthetic test data

Syntho provides a synthetic-data platform that combines masking, rule-based generation and AI-generated data. The goal is to give development, testing and analytics teams useful data without routinely handing them identifiable production records. The key decision is which properties the output must preserve: a valid database relationship, a statistical pattern and a rare business exception are different requirements, and they may call for different generation methods.

01 / ProductA single platform combines methods with different purposes

The Syntho platform groups its offer into synthetic data masking, rule-based generation and AI-generated synthetic data. Masking changes sensitive values, rules encode intended constraints, and AI-based synthesis aims to reproduce statistical relationships. Combining them can be useful because a dataset rarely has one uniform requirement across every column and table.

The AI-generated data page describes generating new datapoints from patterns in source data. That is distinct from simply copying a row and changing its name. The reader should still evaluate the resulting dataset for the intended use: realistic distributions do not automatically prove that every rare case is represented or that the output is appropriate for unrestricted sharing.

There is also an identity change to understand. Syntho’s June 2026 announcement says it acquired the MOSTLY AI brand, which continues as MOSTLY AI powered by Syntho. This article covers Syntho’s current platform. The announcement does not establish that every historical MOSTLY AI product, subscription or hosted service continues under the same terms.

02 / AudienceUseful when data access depends on preserving the right properties

Syntho fits development teams that need coherent test databases, data scientists preparing analytical sandboxes and organizations sharing representative data with a defined audience. These use cases differ. A demonstration might prioritize credible values and a smooth customer journey; model development needs relevant correlations; an integration test needs keys, constraints and edge cases to remain meaningful.

The initial buyer task is to identify the properties that matter. If a cancellation analysis depends on the relationship between tenure, plan and support history, preserving each column’s average separately is insufficient. If a billing test depends on an invalid date being rejected, a statistically typical dataset may never exercise the failure. The generation strategy should follow those needs.

The BigID blueprint offers context for discovering and classifying sensitive source material. The Immuta blueprint covers access policy for governed data. Synthetic generation complements those responsibilities but does not erase them: the organization still needs authority to use the source, a defined audience for the output and a process for deciding whether the transformed data meets its requirements.

03 / WorkflowA proposed analytics sandbox separates fidelity from edge-case coverage

Consider a proposed sandbox for a subscription business studying cancellations. It contains customers, subscriptions, invoices and support cases. The team wants analysts to explore patterns without giving them the original account identifiers or messages. Start with an approved source subset and define which analysis must remain meaningful, such as cancellation patterns by tenure and plan.

Map table relationships before generation. Syntho’s foreign-key management inherits defined relationships, suggests possible missing links and supports bulk import and export. Review suggested relationships with someone who understands the schema. Two columns with similar values are not necessarily the same business key, and an incorrect link can make a dataset internally coherent in the wrong way.

Choose treatment column by column. Replace direct identifiers with suitable synthetic values, apply explicit constraints where the application requires them and use AI-based synthesis where statistical relationships matter. Keep those choices documented. A generator that preserves a realistic distribution may still violate a domain rule, such as an invoice appearing before the subscription began, unless that relationship is considered.

Create a separate set of deliberately constructed cases for software testing. Include a cancellation after a partial refund, an invoice with no matching subscription and a customer with multiple concurrent plans. Label those cases as constructed rather than representative of actual prevalence. Mixing them into an analytical sample without that distinction could distort the very pattern the analysts are trying to study.

Review the quality-assurance report for the dimensions relevant to the project. Syntho describes distributions, correlations and multivariate comparisons, along with privacy-oriented metrics. Compare cancellation behavior within smaller groups as well as the overall population. A good aggregate match can conceal a poor representation of a rare but important subscription segment.

Run independent checks against the output. Database assertions should catch broken joins and invalid dates. Analytical checks should compare the measures the team actually intends to use. Privacy review should consider sensitive combinations and unusual records, not only whether the obvious identifiers changed. These are proposed acceptance activities; this research did not run them or establish a privacy threshold for a real dataset.

Publish the sandbox only to the agreed audience after those checks. Keep the generation configuration, source-scope description and review result with the output version. When a source schema or business rule changes, regenerate and reassess the affected properties. Reusing a previously approved configuration does not establish that a newly generated dataset has the same utility or exposure.

04 / PricingFeature-based licensing has a deployment-cost inconsistency to resolve

The pricing page lists Basic, Standard and Ultimate as quote-based plans and describes feature-based licensing without per-generation consumption charges. It also distinguishes database connections, connector types and report availability. The table focuses on those buying dimensions rather than converting a sales quote into an invented per-row tariff.

One consequential conflict appears within the same page. The headline benefits say unlimited deployments across environments, while the FAQ says the license includes one deployment and additional locations incur extra costs. The FAQ also describes usually starting with a one-year agreement. Obtain written clarification for the intended environments before relying on either deployment statement.

For the proposed sandbox, count the necessary source and destination connections and confirm whether the selected plan includes the quality reports the team will use for acceptance. The absence of a generation charge does not remove infrastructure expense. Compute, storage, database access and the time spent reviewing outputs remain part of the project’s operating cost, even when the software license is fixed.

05 / DistinctionsMethod choice and relationship management are practical strengths

The platform’s useful distinction is that it does not require every problem to be expressed as statistical synthesis. Some fields need a deterministic format, others need a business constraint and others need learned relationships. A team can evaluate whether that combination reduces the amount of custom data-preparation code it must maintain without abandoning explicit control over important fields.

Relationship management is similarly practical. Real schemas often omit foreign-key declarations even when applications rely on the relationship. Syntho’s scanner and bulk tools give teams a way to represent that knowledge. The essential review step remains human confirmation of what the relationship means; automatically finding a plausible join is not the same as proving it is correct.

The deployment guide describes Docker Compose or Kubernetes, air-gapped operation and CPU-based operation without a GPU requirement. It also states that Syntho collects no deployment telemetry or data. These are vendor-described deployment properties to verify in the selected version and configuration, not a claim that the customer has no infrastructure or operational responsibilities.

06 / QuestionsAsk whether the output supports the decision it will inform

Which quality dimensions are included under the proposed plan, and how are they calculated? A report should help explain the difference between source and output, not merely produce an attractive overall score. Ask for examples where fidelity is weaker and inspect whether the report makes those weaknesses visible at the group and relationship levels relevant to the task.

What would make a dataset unsafe to release to its intended audience? The quality page discusses privacy metrics and the need to evaluate external sharing. That does not establish a universal threshold for every business or source. Agree the relevant checks and review responsibility before generating a large output that someone later assumes can be published freely.

How will the deployment be licensed across development, testing and disaster-recovery locations? Resolve the pricing-page inconsistency and document the answer in the agreement. Separately confirm which support and update processes work in an isolated environment. This source review does not demonstrate actual runtime performance, statistical fidelity or privacy protection for a customer dataset.

07 / DecisionChoose one dataset and a decision that better access will unlock

Syntho is a useful candidate when a team needs realistic data and can state what realism means for its job. Begin with one linked dataset and a concrete analysis or application journey. Select generation methods according to those requirements, then judge the output through both product reports and independent checks.

For the subscription sandbox, the milestone is an approved dataset that preserves the relevant cancellation relationships and behaves coherently across customers, invoices and support cases. Additional volume is valuable only after those properties hold. Confirm the feature gates and deployment terms before broadening access, and keep the current Syntho offer distinct from assumptions about legacy MOSTLY AI services.