Block Convey

The product exists only because models are in production and it operates on them directly, capturing every language model call, tool invocation and reasoning step, scoring output quality, generating its own analysis of each trace with a risk level and recommendations, and running evaluations that become readiness reports.

Its own diagnosis frames the need precisely: agents in production make thousands of decisions daily, and without observability an institution has no visibility into what they are saying, whether they follow instructions, or whether they are exposing sensitive data.

The product is oversight infrastructure and its workflow keeps people in the loop by design, moving from detection where alerts flag a quality drop, through investigation of specific traces to identify root cause, to a fix applied as a versioned prompt change, a new guardrail rule or a parameter adjustment, then verification. Guardrails act automatically at ingestion, which is appropriate for data protection. Held at B because no escalation path, severity threshold or human approval requirement is described for the automatic blocking decisions.

The evidence chain is complete and each link is specified. Traces are tamper-proof and immutable from creation, which is what makes them admissible rather than merely informative; regulator-facing evidence packs export in under sixty seconds; agent trajectories and model audits reconstruct how a decision was reached; and prompts are versioned through a library, treating them as controlled model artefacts rather than editable configuration, which is a control most institutions have not yet implemented themselves. Compliance teams review session-level transcripts rather than raw interface logs, which is the difference between a record and an auditable one.

No customer, institution or deployment is named, and no adoption count, trace volume or outcome figure is published. Verifiable presence amounts to a place in an industry sandbox programme and participation in conferences. For a governance product whose value proposition is surviving examination, the absence of a single institution willing to be named is the gap most worth closing.

The boundary this product enforces runs outward, keeping customer data from reaching external models, and that is a real control the buyer acquires. What is missing is the same boundary drawn around the vendor itself: the platform captures every model call, tool invocation and reasoning step across its customers, which is an exceptionally rich corpus of how regulated institutions actually use AI, and nothing states whether that material informs its own models, is isolated per customer, or is retained after an engagement ends.

Privacy is engineered and described mechanically rather than asserted. Guardrails strip the eighteen safe harbour identifiers at ingestion, which is the recognised health de-identification standard named precisely, and social security numbers, card numbers, dates of birth and credentials are scrubbed before reaching storage rather than after. The endpoint layer prevents personal and health data leaving for an external model at all. Held at B because no data processing agreement, retention schedule or subprocessor register was located for the trace corpus that remains after scrubbing.

No attestation, certification or trust centre was located, which is conspicuous for a company whose product is compliance evidence and which maps its releases against payment card and operational resilience standards on its customers' behalf. Technical controls are strong and described in detail, including scrubbing before storage and immutability from creation, and none of it is independently attested.

This is the most complete regulatory mapping in the index, and unusually it is obligation-by-obligation rather than a list of logos. United States banking model risk guidance is treated correctly, with the revised interagency guidance superseding both the 2011 supervisory letter and the 2021 statement, the three-pillar discipline carried forward and scaled to each bank's model risk profile, and evidence produced at every tier.

State cybersecurity rules are mapped to AI use by covered entities. Fair lending is handled specifically, noting that the consumer regulator has confirmed equal credit opportunity requirements apply to AI-driven decisions and that adverse action notices need specific reasons, which agent trajectories and model audits generate.

Insurance model bulletins and state patchwork are covered, alongside European AI, operational resilience and data protection regimes, financial reporting controls, payment card standards, health privacy and the national AI risk framework.

Fairness is a stated purpose of the product, which helps customers demonstrate their models are fair, compliant and examination-ready, and the fair lending mapping is substantive because producing specific adverse action reasons from recorded agent reasoning is the operational form that fairness obligation takes. Insurance model bulletin coverage carries related expectations. Held at B because no bias testing methodology, disparity metric or fairness evaluation approach of its own is described, so what the readiness report actually measures on fairness is unclear.

No guarantee, indemnity or correction process was located. The product improves the position of the affected consumer indirectly and genuinely, since specific adverse action reasons generated from recorded reasoning are exactly what a declined applicant is owed, and nothing states what the vendor stands behind if an evidence pack proves incomplete under examination or a guardrail fails to catch regulated data before it reaches an external model.

This is the most complete supply chain disclosure in the index, naming twelve components across three layers. Five model families are identified with specific versions where relevant, alongside support for custom and fine-tuned models; four cloud AI platforms are named; and three agent frameworks are listed plus custom implementations.

Because the product's purpose is observing what those models do, naming them is not incidental but structural, and an institution can therefore see exactly which external dependencies sit inside the governance layer meant to document its other dependencies.

Integration targets the AI stack rather than banking systems, and within that scope it is thorough, connecting to any system making model calls in any language or framework, with setup stated at under fifteen minutes. The endpoint product deploys through the three major enterprise device management platforms, which is the practical detail that determines whether a bank can actually roll it out. Held at B because no core banking, origination or case management system appears, so evidence produced here still has to reach the institution's own control environment by unstated means.

Four cloud AI platforms are named as supported environments, so an institution already running models in its own cloud tenancy knows the tooling meets it there, and the endpoint layer operates on the device before data leaves. Held at B because the vendor's own hosting, region options and residency commitments are not stated, which matters for a platform that stores immutable traces of regulated decision-making.

No pricing, packaging or basis of charge was located for either product. Charging model matters unusually here because one product captures every model call and the other is deployed per endpoint, so cost could scale with either usage volume or headcount, and neither is indicated.

Coverage spans banks under model risk guidance, insurers under model bulletins and state rules, lenders under fair lending requirements, asset managers, and fintechs and insurtechs needing to demonstrate readiness to their own institutional buyers. The obligation mapping extends into health data and payment card handling, which broadens applicability beyond financial services. Held at B because breadth of addressable segment is not the same as presence in it, and no deployment in any segment is evidenced.