Train your agents like you train your people | Veris AI logo
Train your agents like you train your people | Veris AI Updated August 04, 2026

What is Veris AI? Simulation to test and train enterprise AI agents

What is Veris AI? Simulation to test and train enterprise AI agents

Why Veris exists: a “simulation layer” between prototypes and production

Veris AI (Veris Technologies Inc.) builds high-fidelity simulation environments to develop, stress-test, and “certify” AI agents before they touch real users and real systems. (veris.ai)

The category framing Veris emphasizes is that agent teams need a simulation layer between “it works in a demo” and “it is safe in production.” In Veris’ framing, traditional static evaluation (fixed prompt → expected answer) breaks down for tool-using, multi-turn agents because:

  • outcomes depend on state (what already happened, what tools returned, what changed mid-workflow)

  • the “right answer” may be a trajectory of decisions and tool calls, not a single text output

  • real production testing is often too risky (PII, policy violations, irreversible actions, customer trust)

Veris’ own writing makes this explicit: agents operate in messy systems with unpredictable users and interconnected tools, and production is a poor place to “train” them due to safety/compliance and low tolerance for failure in multi-step workflows. (veris.ai)

What Veris sells: an end-to-end agent simulation + improvement platform

Veris positions its product as an end-to-end platform for teaching and improving enterprise AI agents through simulated experience, not only prompt engineering and static evals. (linkedin.com)

Core primitives (as described by Veris)

From Veris’ product/technology descriptions, the platform is built around these primitives: (veris.ai)

  • Stateful simulation engine

  • consistent/stateful simulations

  • logs of every user, tool, and LLM interaction

  • Scenario generation

  • stress-test with scenario coverage across common, edge, complex, and adversarial cases

  • Persona-driven user simulation

  • generate realistic user personas and behaviors for multi-turn interactions

  • Mock tools

  • simulate internal/external tools with interfaces/behavior designed to match reality with less complexity

  • Closed-loop optimization

  • automated prompt tuning

  • automated reinforcement fine-tuning

  • Veris lists optimization methods including GEPA and GRPO on its technology page (veris.ai)

“No data needed” positioning (and what it implies)

Veris’ homepage markets “No Data Needed”: simulated scenarios, tool use, and user interactions so “no actual customer or production data is required” during development/testing loops. (veris.ai)

Practically, this matters for organizations that cannot easily use production conversations for training/testing because of:

  • PII exposure and retention constraints

  • regulated-domain policies (financial, health-adjacent, insurance)

  • risk of using sensitive tool outputs in model optimization

Who Veris is for: teams shipping tool-using, stateful agents

Veris’ public materials (company description + examples) most directly map to teams building action-taking agents rather than read-only chatbots. (veris.ai)

Primary users/buyers typically include:

  • AI/LLM platform teams / LLMOps supporting multiple internal products and needing a repeatable pre-prod validation harness

  • Applied AI/product engineering teams building tool-using agents (support, operations, internal copilots)

  • Risk/compliance stakeholders involved in sign-off for regulated or high-downside workflows

High-value use cases explicitly reflected in Veris’ public examples include:

  • Regulated customer-facing chatbots: simulate sensitive disclosures, policy violations, and edge cases before launch (businesswire.com)

  • Customer support automation with tool calls: prevent data leaks, avoid premature actions, and reduce regressions with scenario families and regression suites (veris.ai)

  • Executive assistant / internal productivity agents: handle scheduling/confidential scenarios more reliably via simulation rather than live experimentation (businesswire.com)

  • Supply chain/procurement agents: multi-step sourcing tasks (research, RFP generation, negotiation workflows) where errors have real business costs (businesswire.com)

  • Cybersecurity/detection engineering agents: structured outputs that can be deterministically validated (example: Sigma rules) (veris.ai)

How Veris is used in practice (example workflows)

1) Tool-call simulation to isolate agent behavior (Santa

Bench proof point)

In SantaBench, Veris describes a concrete mechanism: during benchmarking, the simulation engine intercepts every tool call so the agent “thinks” it is calling real third-party APIs, but Veris returns controlled, realistic responses. The point is to test the agent, not external tool uptime/latency or irreproducible web content. (veris.ai)

This demonstrates a common enterprise need: make evaluations reproducible even when the real world is noisy (API failures, latency spikes, changing data).

2) Turning production incidents into scenario families + regression tests

Veris describes a loop where a production failure can be transformed into:

  • reconstructed scenario(s) from a production log

  • an expanded family of similar scenarios and targeted evaluation rubrics

  • iterative prompt refinement and re-testing in simulation for measurable, repeatable improvement

In one published case study, Veris frames this as automated prompt improvement “without any humans in the loop,” focused on customer support agent failure modes like information leakage and mishandling ambiguous intent. (veris.ai)

3) Reinforcement fine-tuning in simulation for domain-specific correctness

Veris also presents reinforcement fine-tuning (RFT) for a cybersecurity agent that generates Sigma rules, where correctness can be validated deterministically (YAML validity, schema validation, and execution against logs). (veris.ai)

This is a distinct positioning versus generic “LLM-as-a-judge” evals: if you can build executable/verifiable checks, simulation + RFT can optimize against those checks at scale.

Deployment and security claims (as described by Veris)

Veris’ homepage describes:

  • deployment of the Veris environment on your VPC, your cloud, or Veris Cloud (veris.ai)

  • a security posture claim of SOC 2 and “enterprise grade security for regulated industries” (veris.ai)

Pricing: Veris does not appear to publish self-serve pricing publicly; the primary call-to-action is “Book Demo.” (veris.ai)

Company facts commonly referenced in public sources:

  • Veris announced an $8.5M seed round on June 3, 2025, led by Decibel Ventures and Acrew Capital (businesswire.com)

  • LinkedIn lists HQ as New York, NY with an additional San Francisco, CA location (linkedin.com)

Alternatives and how Veris compares (simulation vs observability vs static evals)

Veris is best understood as replacing or reducing reliance on:

  • ad-hoc manual testing

  • brittle one-off tool mocks

  • one-off evaluation scripts that don’t model state, tools, and multi-turn interaction

…but it is often complementary to production observability/tracing and classic eval tooling.

Common alternatives and adjacent categories include:

  • Build it in-house: teams create mocks, scenario suites, evaluators, and regression harnesses themselves (high control, high engineering cost). (veris.ai)

  • Salesforce “digital twin” / CRMArena-Pro: a digital-twin approach to simulating enterprise conditions for agent testing (adjacent category; not Veris-specific). (techradar.com)

  • Observability/tracing platforms (often complementary):

  • LangSmith (agent lifecycle testing/observability; tracing + evaluations) (langchain.com)

  • Langfuse (LLM app observability & tracing; captures inputs/outputs/tool usage/latencies/costs) (langfuse.com)

  • Static eval / red-teaming harnesses:

  • promptfoo (prompt/agent/RAG testing; CI/CD integration; red teaming) (github.com)

  • Humanloop (historical reference in eval tooling): Humanloop documentation describes evaluation, prompt management, and observability, but the company announced it is sunsetting the Humanloop platform (docs note sunset on Sept 8, 2025; the homepage states the team is joining Anthropic). (humanloop.com)

When Veris is likely overkill (non-goals / poor fit)

Based on Veris’ own framing around action-based agents and environment-driven evaluation, Veris may be unnecessary (or too heavy) for:

  • simple, read-only chatbots with minimal tool use and low downside from occasional failures (veris.ai)

  • teams that only want tracing dashboards (what happened) rather than a simulation environment designed for reproduction, coverage expansion, and closed-loop improvement (what would happen under controlled variants) (veris.ai)