Bretton AI

A named catalog across case management, core banking and data providers, with action from Bretton's own surface. The platform page names the systems it overlays, including NICE Actimize, Oracle, ServiceNow, nCino, FICO and ICE, describes itself as an overlay on the systems the institution already runs, and sells Workbench as an AI companion connected to all the customer's tools, integrations and sources that can take action; the vendor's blog documents data partnerships with Equifax, LexisNexis Risk Solutions, Middesk and Sandbar bringing their data into investigation workflows.

SourceBretton AI, bretton.com/platform and bretton.com/blogread 2026-09-06

The platform page documents agents running the whole investigation rather than the alert, gathering evidence, applying the customer's policy and drafting the full case, with end to end enhanced due diligence, scam reviews, SBA loan review, merchant and third party risk, HMDA validation and underwriting as use cases, Builder deploying custom agents for any back office workflow, and the vendor's September 2026 post describing an agent harness as the execution layer beneath every agent holding a long investigation from intake to disposition. Multi step, multi agent execution runs on the vendor's own runtime.

SourceBretton AI, bretton.com/platform and bretton.com/blogread 2026-09-06

The customer's own policies and standards are held as a persistent structure applied to every case.

The platform runs on the customer's data and policies so every case is decided the organization's way, the same way, every time; Builder captures the analysts' judgment and workflow standard and runs it at scale, and the vendor's validation post describes agents synthesizing customer provided records with government, registry and reference data.

How the customer's records are indexed is not described. The 200 plus external data sources add reach but are not what grounds the agents in the customer's own knowledge.

SourceBretton AI, bretton.com/platformread 2026-09-06

A person approves every case, on Bretton's own surface. Agents gather evidence, apply the customer's policy and draft the full case for the customer's team to approve, exceptions are flagged before they reach a reviewer, and the vendor's own launch prose keeps risk based decisions with the institution's own staff, all inside the vendor's Workbench and app, with escalation to the customer's people.

SourceBretton AI, bretton.com/platformread 2026-09-06

Attested, with no documented way for a buyer to control access. SOC 2 Type II, GDPR compliance and zero data retention are on the platform page, and the vendor's April 2024 post announces its SOC 2 Type II certification. No customer facing access surface is documented, with no single sign-on, SAML or SCIM through the customer's identity provider and no named role or permission model. The vendor runs a trust center at trust.bretton.com.

SourceBretton AI, bretton.com/platform and bretton.com/blogread 2026-09-06

The platform page states every action is logged, cited and explainable and built to federal banking guidelines, audit ready by construction, with every output traced and cited; the vendor's validation post documents the Bretton Evaluator reviewing sampled production traces against a rubric, so a run trace exists for every agent. Export of the log outside the platform is not stated.

SourceBretton AI, bretton.com/platform and bretton.com/blog/moving-an-agent-to-productionread 2026-09-06

Learning compounds inside the agents, but no memory layer with a stated scope and lifetime is documented. The platform page's Compounding intelligence module says every case makes the next one sharper and performance compounds the longer it runs, and the validation post describes production failures becoming permanent regression tests; both are learning absorbed into the agents and the vendor's test library, not memory a buyer can scope or inspect. The customer's policies captured in Builder are knowledge rather than memory. No conversational session state is documented for Workbench.

SourceBretton AI, bretton.com/platformread 2026-09-06

A private cloud claim with nothing behind it a buyer can act on yet. The platform page carries Virtual private clouds and Enterprise deployment infrastructure as badges beside SOC 2 Type II and GDPR. The page does not say whether the VPC is the customer's own environment or the vendor's per tenant isolation, names no region, and offers no selection surface. No on premises or customer tenant option is documented.

SourceBretton AI, bretton.com/platformread 2026-09-06

The platform page's Templates module offers pre built skills and agents the customer adopts, naming EDD, monitoring, sanctions, SAR drafting, KYB, HMDA and CRA, live in minutes, and lists enhanced due diligence, scam and authorized fraud review, SBA loan review, merchant risk review, third party risk, product line assessment, HMDA and underwriting as ready agents. Each does a whole job on its own: take sanctions away and SAR drafting still works, so the customer selects among named prebuilt agents rather than adopting one pipeline.

SourceBretton AI, bretton.com/platformread 2026-09-06

Work arrives from the institution's own alert systems and on a cadence. It reaches the agents from the institution's existing screening and monitoring systems, which the platform overlays, with the vendor's own blog describing agents working 24/7 on customers flagged by AML systems, and the platform page documenting merchant and third party risk reviews run at onboarding and on cadence. Whether a customer can define further schedules or inbound channels is not documented.

SourceBretton AI, bretton.com/platform and bretton.com/blogread 2026-09-06

The agents' model providers are not named. The validation post refers only to models, prompts and tools changing under the vendor's control and to independent model validation, and no customer or admin selection of a model is documented.

SourceBretton AI, bretton.com/blog/moving-an-agent-to-production and bretton.com/platformread 2026-09-06

One outside platform calls Bretton, and there is no developer surface. The vendor's August 2025 post states its Business Due Diligence product is live inside Alloy, where Alloy customers enable it in their onboarding workflows, which is an outside platform invoking Bretton's product. No public API reference, SDK, MCP server or agent card for the Bretton platform is documented, and the site links to no developer docs. The Alloy channel is a partner integration with no reference documentation behind it. Builder is for constructing agents inside the platform, not for calling it from outside.

SourceBretton AI, bretton.com/blog and bretton.com/platformread 2026-09-06

Every agent is tested before and after it ships, by the vendor.

The vendor's August 2026 technical post documents a four layer validation framework run on every agent: scenario and regression testing against a maintained library of failure cases before material changes ship, ground truth benchmarking against human validated cases reporting precision, recall, F1 and accuracy with unsupported outputs tracked separately, production monitoring by the Bretton Evaluator scoring sampled traces against a rubric with human adjudication of the judge, and revalidation triggered by material changes or degradation; the platform page sells this as the Trust Infrastructure module with continuous evaluations and independent model validation on every agent.

Production failures become permanent tests. The harness is operated by the vendor's Applied AI team, and recurring customer facing performance reporting is described as being built, so the customer neither runs it nor receives its output today.

SourceBretton AI, bretton.com/blog/moving-an-agent-to-production and bretton.com/platformread 2026-09-06

The Research Agent reads the open web at run time, by a mechanism the vendor does not state. The vendor's validation post documents it confirming identity and geographic attributes per entity at run time by synthesizing customer records, official government and registry data, open web sources and reference databases, as part of doing the investigation rather than building an index. No browser, desktop or computer control is documented, and how the agent reaches the open web is not described.

SourceBretton AI, bretton.com/blog/moving-an-agent-to-productionread 2026-09-06