Veridect launches AI verification layer for enterprise agents

Becker Transactions announced Veridect, a new enterprise AI control layer that verifies answers and blocks or routes autonomous agent actions before they execute. The system is already in production at Fortune 100 companies and is now open for public testing as businesses face growing risk in regulated workflows.

Why it matters: - Veridect is designed for regulated, high-consequence workflows where a wrong AI answer or unauthorized agent action can create legal, financial, or patient-safety risk. - The product aims to give enterprises a way to verify AI outputs, govern autonomous actions, and preserve an audit trail that can be defended to regulators, boards, courts, or customers. - Becker Transactions positions Veridect as a control layer for banks, insurers, healthcare organizations, and systems integrators serving those sectors.

What happened: - Becker Transactions announced Veridect, which the company describes as the “verdict layer” for enterprise AI. - Veridect runs four leading AI models in parallel to verify answers and gates agent actions before they execute. - The system is already in production inside Fortune 100 enterprises. - Veridect is live and open to test at the Veridect demo site.

The details: - Veridect’s Quad-AI Consensus Engine uses Claude Opus 4.5, GPT-5.1, Gemini 2.5 Pro and Sonar Pro from four independent providers. - A verifier mesh cross-examines the model outputs, applies weighted-consensus voting and returns one answer with a calibrated confidence score. - The confidence output is designed to flag whether uncertainty comes from a reasoning gap, stale knowledge, hallucination risk or domain mismatch. - On an independently reproducible MMLU-Pro benchmark with 100 questions, consensus scored 85.0% versus 82.0% for the best single model. - That produced a 3.0 percentage-point lift, which the company says came from the models cross-examining one another. - The same engine also works as a pre-action gate for autonomous agents. - Before an agent moves money, sends data or changes a record, Veridect checks the action’s verb, target system and parameters, not just the text. - The gate returns one of three outcomes in about three seconds: greenlight, escalate to a human or block. - In-scope, high-confidence actions clear automatically. - Uncertain, high-risk or out-of-policy actions go to a person with the reasoning attached. - Actions that exceed an agent’s authority are stopped before execution. - The company cites examples including a $99,000 payment that clears, a $101,000 payment that escalates at a policy threshold, a write to a personal record that escalates and access to protected health data that routes to the highest-risk tier. - In May 2026, a credentialed Fortune 100 model-risk reviewer ran seven adversarial scenarios live on Veridect’s public demos across legal, finance, HR, supply chain, multi-jurisdiction and healthcare. - The gate matched its stated behavior in all seven cases, including holding boundaries, escalating at the right thresholds and decomposing failure modes correctly. - Veridect records each answer and agent decision in a tamper-evident bundle that is SHA-256-chained to the prior record. - Each bundle includes raw provider responses, routing weights, a heatmap of model agreement and divergence, failure-mode decomposition, weighted synthesis and the prior hash link. - Structured JSON export is available now for governance-stack ingestion. - Formatted PDF and CSV adapters are on the near-term roadmap. - The audit trail is intended to support model-governance and AI-risk review under frameworks such as the Federal Reserve’s SR 11-7. - Veridect also bundles four AI vendors behind one endpoint with circuit breakers, automatic failover and caching. - The company estimates that approach can reduce integration work by 60% to 80% versus building a custom consensus layer. - Veridect is provider-agnostic and cloud-agnostic, and can run on major clouds or inside a company’s own data center.

Between the lines: - The product is framed less as a model competitor and more as an independent control plane that sits beside existing AI systems. - Veridect’s public red-team results show an emphasis on transparency, including cases where consensus underperforms a single model. - On MedQA with 50 questions, consensus fell 2.0 percentage points below the best single model, and the company says tuning for medical reasoning is underway. - The product pitch centers on accountability, not replacement of the human expert who owns the decision. - Lisa Russell, CEO of Veridect, said the main barrier to enterprise AI is accountability, not capability. - Dean Becker, principal of Becker Transactions, said the product addresses the problem standing between enterprises and trustworthy AI.

What's next: - Veridect says the consensus engine, pre-action gate and red-team validation are all open for public testing now. - The company plans to add formatted PDF and CSV adapters soon. - Veridect will likely compete for adoption in regulated industries that need auditable AI controls before scaling agents into production.

The bottom line: - Veridect is trying to make enterprise AI defensible by verifying outputs, gating actions and preserving replayable audit records before autonomous agents can cause damage.