Fiddler AI

Fiddler ships its own inference as product, which a pure governance system of record does not. Named evaluator model families score hallucination, toxicity and personal data leakage, and explanations are computed using Shapley values. The platform underneath is observability infrastructure, though. Without the evaluator models, drift statistics, performance monitoring, alerting and audit capture remain, which is a working product and historically was the whole product. The closer a product sits to oversight of other people's models, the less of its value comes from its own inference.

Guardrails sit in the execution path rather than beside it, and the mechanism is well described. They moderate, intervene in and block agent conversations in real time, detect personal data leakage before exposure, and stop agents deviating from approved scripts. Human approval is enforced on high value loans. The gap is in the rules themselves. Enforcement is described as blocking regulatory violations and policy breaches without naming one rule. The human approval threshold is left to the customer, with no default and no floor.

Model risk disclosure is dense and specific for what the product does. Explanation runs by Shapley values and a proprietary variant at local and global level, with what if analysis on prediction outcomes. Drift root cause analysis goes down to contributing features, and segment level analysis covers feature impact, correlation and distribution. Generative metrics are named individually as faithfulness, context relevance, coherence and consistency.

Model risk and governance reporting workflows are supported explicitly. Every one of those mechanisms applies to the customer's models, though. Nothing documents Fiddler's own evaluator models, their validation, their error rates or their drift policy. A third party directory reports documented alignment to United States supervisory model risk guidance, with reports generated for periodic examination. Fiddler's own material does not confirm it.

The one named customer, with a named executive quoted on production monitoring, is outside financial services. The financial services evidence is real but anonymized. It includes a Fortune 100 institution across wealth, brokerage and asset management, and an investment grade institution deploying evaluator models inside its own virtual private cloud.

A consumer lending platform reports a large annual hours saving, but that figure comes from a third party directory rather than Fiddler's own material. The named reference and the quantified outcomes never join.

The pooled data question is answered by architecture rather than by promise. Evaluation runs on Fiddler's own model families inside the customer's estate, with no external model interface call, so customer prompts and outputs are not sent to a third party for scoring. What is not published is an explicit statement that customer data never influences another customer's results. There is also no stated model hosting policy, supplier retention posture or concrete lifecycle terms.

No privacy posture, processing terms, retention schedule or subprocessor list is published. Detection of personal and health data leakage ships as a product feature, which is a capability sold rather than a posture disclosed. The in environment architecture is a deployment property, not a privacy disclosure.

Fiddler's own site shows no certification, attestation or trust portal. For a platform that sells into a Fortune 100 financial institution and deploys inside customer cloud estates, that gap likely says more about disclosure than about controls. Such a platform will have passed enterprise security review repeatedly and is likely to hold attestations. For a developer facing product, this material usually sits on a trust or security page or in the documentation portal.

There is no license, no supervised test of the product by a financial regulator and no program enrollment. Producing reports intended for supervisory examination is a product feature bought by a supervised institution. It is not regulatory standing held by the vendor.

Fiddler sells fairness metrics to its customers and publishes none about its own models. Four named metrics ship out of the box, disparate impact, group benefit, equal opportunity and demographic parity, aimed explicitly at credit card approval and lending decisions. Nothing is published about fairness testing of the evaluator models. Those models decide what counts as toxic, what counts as a policy breach and which agent conversations get blocked. These are consequential judgments applied to customer communications. A model that moderates language has well documented potential for uneven performance across dialect and demographic group.

Nothing is published on responsibility when a guardrail fails to block, when an evaluator wrongly flags, or when a governed agent causes harm the platform did not surface. There is no route for an affected individual, who is two steps removed from the vendor. The exposure is real rather than theoretical. A product whose job is to block and escalate fails silently by construction, since a violation that was not detected leaves no artifact.

This is a deployment disclosure that reads like a supply chain disclosure. Fiddler names its own evaluator model families and says customers may instead use custom evaluators or external judging models, which tells a buyer where inference happens and who operates it. It names no base model, provider, version, architecture or training provenance for any of them. Saying the models run in your environment answers where, and says nothing at all about what.

Agents are governed through the gateway the enterprise already runs, which is the layer that matters for this kind of product, and no parallel path is required. A published interface, model pipeline integration and reporting feed existing model risk and governance workflows. Coverage extends to first party, third party and coding agents, so it reaches software the customer did not build, the harder half. No core banking, loan origination, decisioning or model operations platform is named as a certified connector. Integration is described by architectural position rather than by counterparty.

The evaluator models run inside the customer's virtual private cloud. Generative agents can be governed with no data sharing and no call to an external model interface, and this is stated as an architectural property rather than a configuration option. That answers the two questions a regulated buyer actually asks: where the data goes, and whether it leaves the estate. No region list, tenancy model, residency guarantee or subprocessor detail is published.

No price, tier, unit or pricing basis is published. The one adjacent commercial claim is that Fiddler's in environment evaluator models remove external model interface costs and fees. That is a cost argument about a competing architecture, not a disclosure of its own pricing.

Within financial services the span is wide and specifically addressed. It reaches banks, credit card organizations, consumer lending platforms and fintechs, plus a Fortune 100 institution running wealth management, brokerage and asset management divisions. Use cases reach credit, underwriting and automated lending, as well as fraud and payment default. Collections, robo advisory and trading are covered too. Two limits matter.

Financial services is one vertical of a horizontal platform that also sells to technology, healthcare and government. The institutions evidencing the breadth are described by size and type rather than named.