The enterprise data cul-de-sac without exposure

Enterprise projects that need to train or evaluate LLM agents have long faced an uncomfortable fork. On one side, using real database schemas guarantees adherence to rules but introduces legal and security risks that are hard to accept. On the other, statistical synthesizers require seed data that in many environments do not exist or are not available. SAP's Synthesis Through Simulation (STS) framework, now open source on GitHub, proposes a third path: not extracting data from the company, but simulating the business environment and letting an LLM agent generate data by operating against APIs that enforce policies. Data is born valid because the environment that produces it is the same one that defines validity.

This inversion matters more than it seems. Until today, synthetic generation for enterprise agents has mixed two distinct problems: modeling the distribution of data and applying operational constraints. STS separates them. Structural validity is guaranteed by construction, not by extraction from real schemas or statistical approximation. For those working on local pipelines, this shifts the critical point: there is no need to copy or expose the schema, but there is a need for a simulated environment capable of representing policies with fidelity.

The numbers reported by the source are stark: the Generalist Populator component, a domain-independent agent, achieved an average marginal fidelity of 0.88 and 100% constraint satisfaction on ten environments, without accessing database schemas. It is not a perfect metric, but it signals that the trade-off between validity and realism can be addressed in a new way. For self-hosted infrastructures, it means that material for fine-tuning can be generated without moving sensitive data to the cloud.

Caution remains necessary. The marginal fidelity of 0.88 indicates that some distributions remain approximate. Those evaluating these tools must ask whether the rare tails or edge cases of their domain are represented. But the architectural shift is structural: risk moves from real data to the quality of simulated policies.

Validity by construction shifts risk without eliminating it

The central mechanism of STS deserves attention. An LLM agent executes operations against APIs that enforce policies in simulated environments. Because the APIs accept only valid transactions, generated data inherits structural validity from the environment. This differs from generating records with a statistical model and then checking constraints: here the constraint is part of the generative process. The consequence is that constraint satisfaction can reach 100% without ever seeing a real database schema.

The separation between constraint enforcement and distribution modeling is the most relevant contribution. In traditional methods these two objectives travel together and often one penalizes the other. A statistical model can produce realistic distributions but violate business rules; a schema-based approach can respect structure but fail in variability. STS tries to avoid this forced coupling.

The cited results also show the limits of competing approaches. Statistical synthesizers are not applicable in seven of the ten environments because they require seed data. Agents with schema access fail 82% of trajectories in the airline environment because task composition is too fragile. This datum is a strong signal: access to the real schema is not only a legal risk, but can produce fragile pipelines.

For those designing local systems, the lesson is that operational validity must be tested as a property separate from the simple schema. The simulated environment becomes a testbed for agents before deployment. The limit remains: the simulation is only as good as the policies implemented; if the environment does not capture a real rule, the generated data may be valid only in appearance.

From simulation to fine-tuning: the local pipeline changes TCO

The message for on-premise deployments is direct. STS opens the possibility of building a local pipeline that starts from simulation and arrives at fine-tuning, maintaining complete control over data. There is no longer a need to move sensitive datasets to the cloud to train specialized agents. This reduces dependence on external providers and shifts attention to the quality of the simulated environment, which becomes the new critical point.

The TCO calculus changes. In the past, much of the cost and time for fine-tuning enterprise agents derived from data preparation: legal approvals, anonymization, and often the use of cloud services to avoid maintaining local data infrastructures. With policy-based synthetic generation, those costs can shift toward designing and maintaining the simulated environment. For those already managing self-hosted servers for inference, the additional step of data generation can integrate without introducing an external dependency.

The hardware aspect should not be overlooked. Synthetic generation via LLM agents still requires compute capacity, which in local contexts adds to VRAM requirements for inference and fine-tuning. But the load is not necessarily comparable to training on massive real datasets. The bottleneck shifts from data bandwidth to the ability to model business policies.

The most concrete benefit is for teams managing self-hosted infrastructures: they can iterate on synthetic datasets without waiting for legal approval to use real data, accelerating development cycles. This does not eliminate compliance, but changes when it enters the process. Compliance focuses on simulated policies, not on consent to use personal data.

Fragile trajectories: why the real schema is not a shortcut

The 82% failure of trajectories in the airline environment with agents that have schema access is one of the most instructive data points in the source. It shows that having the database structure does not guarantee the ability to compose realistic operations. Task composition becomes too fragile when the operational context defining valid sequences is missing.

This overturns a widespread belief: that the real schema is the gold standard for enterprise data generation. The schema describes structure, not process rules. An agent can generate formally coherent but operationally impossible records. Policy-based simulation, by contrast, constrains the transactions themselves, not just the fields.

For on-premise systems, the fragility of trajectories has a practical consequence. If an organization uses generated data for fine-tuning LLM agents, structural validity is not enough: operational validity is needed. STS aims at this by construction. However, the marginal fidelity of 0.88 leaves open the question of how much simulated distributions represent less frequent cases.

In regulated sectors, where exceptions can be costly, attention must be paid to policy coverage. 100% constraint satisfaction says nothing about distributional fidelity. Those adopting these frameworks should verify both validity and statistical representativeness, using the simulated environment as a laboratory, not as a photocopy of the company.

Data sovereignty and incentives: who gains and who loses

Data sovereignty is the theme that makes STS particularly relevant for AI-RADAR. The ability to generate synthetic data coherent with business constraints without exposing or copying the real schema changes the compliance calculation. European companies, for example, can avoid transferring datasets outside the local perimeter to train specialized agents. The local pipeline becomes a concrete option when compliance requires it.

Incentives shift. Teams managing self-hosted infrastructures can iterate faster without waiting for legal approvals. Providers of cloud data generation services must confront the possibility that companies prefer local solutions when compliance requires it, especially in financial, healthcare and public sectors.

This is not just market competition. STS signals a paradigm shift in AI infrastructure: the simulated environment becomes a strategic asset. Companies may invest in modeling their own business policies rather than accumulating historical data. This moves value from data availability to the quality of representation.

Clear limits remain. The marginal fidelity of 0.88 indicates that some distributions remain approximate. The opening of the code and of the ten environments allows the community to verify and adapt the framework to its own domains, but adoption will depend on the ability to build faithful simulations. Code transparency is an advantage for on-premise environments, where direct inspection reduces the risk of opaque dependencies.

What to watch: signals for the coming months

For those observing the enterprise AI market, there are precise signals to monitor. The first is the adoption of policy simulation frameworks in regulated sectors. If banks, insurers and healthcare organizations begin to generate synthetic data locally for fine-tuning, the STS paradigm will have moved beyond the laboratory phase.

The second signal concerns fidelity measurement. The average marginal metric is not enough: it is necessary to understand how conditional distributions, temporal sequences and rare exceptions are represented. A simulated environment can have high average fidelity but fail precisely in the cases that matter most for operational risk.

The third signal is related to hardware. If local synthetic generation becomes a widespread practice, demand will grow for self-hosted systems capable of alternating generation, fine-tuning and inference without excessive bottlenecks. It is not only a question of VRAM, but of orchestrating end-to-end pipelines.

Finally, the evolution of the role of policies should be observed. The quality of the simulation depends on the fidelity of the implemented policies. If standards emerge for describing and verifying business policies, the simulated environment could become a reusable component of AI infrastructure. The STS framework is a starting point, not a finish line. The coming months will show whether the market prefers this path or continues to seek shortcuts based on real data.