Generating a Consistent Enterprise: Synthesis and Reference-Free Evaluation of Multi-System Business Data

A deterministic synthetic-data generator can produce a full fictional multi-system enterprise, consistent across 66 business-product formats, whose realism is scored reference-free and which drove a fixed-yardstick realism score from 60.3 to 99.1 with an adversarial detector's flagged records dropping from 55.2% to zero over 23 generated companies.

The authors tuned a generative system for fictional enterprises (workforce, customers, deals, tickets, calls, chats, documents) across 21 subsystems, evaluating with a five-axis, 28-check scorecard read from a fixed table of reference targets, an adversarial detector hunting marks of synthetic generation, and soundness checks including a boosted-tree two-sample test against a shuffled copy of the data. Scores were tracked across 23 generated companies during development, with a held-out regression seed never used in tuning.

Mean realism rose from 60.3 to 99.1 and worst-company from 41.1 to 94.9 under the fixed scorecard; detector flags fell from 55.2% of records to zero; the held-out seed scores 91 with no flags. Scale: a 20,000-employee company with about 250,000 records generates in about nine minutes at roughly 25 MB. Scope: all of this is measured against the authors' own reference targets and their own instruments, with no external or human validation, and the reference table itself contains three unsourced 'estimate' targets.

The authors concede three targets are estimates without sources, that realism is judged against the same targets used to generate the data (generator cannot drift from what the scorecard checks by construction), and that the whole setup is self-referential. Unstated limitations: the adversarial detector only hunts known marks of synthetic generation, the realism score is a self-defined yardstick rather than evidence the data fools humans or real downstream models, only 23 companies on 10 industries and 6 size tiers, the companion agent benchmark results are not in this paper, and no external replication or code-level independent audit is reported.

A skeptic would call this a closed loop: the generator, the scorecard, the detector, and the targets all come from the same team, so the 60.3-to-99.1 trajectory measures agreement with the authors' own rubric, not realism in any external sense. 'Reference-free' here really means 'reference is a hand-written table we also generated from.' The adversarial detector finding nothing after tuning may just mean the team taught the detector what to stop looking for. No comparison against hand-built environments like CRMArena or TheAgentCompany on downstream agent-task outcomes is included in this paper.

Little or nothing yet for the AI economy in the next 12 to 24 months as a direct read; this is infrastructure plumbing, not an economic measurement. The indirect relevance is that credible enterprise-agent evaluation may require exactly this kind of consistent multi-system synthetic environment, so if the companion benchmark validates, it could improve the trustworthiness of vendor claims that agents work in enterprise settings. On its own, no revenue, labor, or capex signal.

Technical Capability: Teams building and benchmarking enterprise agents get a way to exercise multi-system agents on consistent, dirty, realistic data without touching customer production records, addressing a real bottleneck in agentic evaluation.

Market & Valuation: If synthetic enterprise environments become the standard test bed, vendors selling agent benchmarks and evaluation infrastructure (this is a commercial hosted service) gain a defensible product line, and buyers gain a way to compare agent claims.