Quantexa

Entity resolution is the product. Quantexa matches messy, inconsistent records into one view of each person and company, and graph analytics builds the network around them. Q Assist and Agent Gateway add generative and agent tools on top. Take the models away and matching falls back to rules, the legacy approach the platform replaces.

Analysts get one resolved view with the network already assembled, so the work shifts from gathering data to judging it. Q Assist drafts research and reports and shows which data each answer used. Agent Gateway lists routing, guardrails, approvals and immutable audit trails for agent actions, but does not describe how an approval works or when one is required. In the Insolvency Service deployment, investigators make every enforcement decision.

The matching models are described as transparent and tunable, train without pre resolved entity lists and are tuned country by country, which tells a validator where their behavior comes from. Quantexa cites matching accuracy of up to 99%, including 99% in an independent test with Dun & Bradstreet, with no method published. Its trust center lists an AI model risk management document. Error rates for resolution itself and support for supervisory validation are not public.

HSBC's case study reports a 60% cut in case volumes and £4 million of potential savings from replacing an earlier system, with Jennifer Shasky Calvery, its group head of financial crime risk and compliance, quoted. Neither figure states a measurement period. Belgium's tax authority reports blocking more than 100 false refunds. A commissioned economic study claiming a 228% return was paid for by Quantexa.

Customers running their own instance keep data and matching models inside it, which settles the cross client question for them. For the hosted product, Quantexa states no boundary on training across customers. Q Assist works with outside models such as OpenAI, Gemini and Mistral, and its product page says nothing about what those providers keep.

The privacy notice, updated 25 June 2026, lists processing personal data that appears in client datasets as a purpose. That matters because a resolved view profiles counterparties and beneficial owners with no relationship to the bank. The notice sets retention for client contacts and screening records and names UK and EU transfer safeguards. It names no hosting subprocessors and says nothing about training models on client data.

The trust center lists ISO 27001:2022, ISO 27017 and Cyber Essentials Plus certificates, with IRAP and NIST 800-53 entries alongside. Two 2026 SOC 2 Type 1 reports cover the Decision Intelligence Platform, and Cloud AML with Unify. Type 1 checks control design at one point in time, not whether controls worked over a period. Business continuity carries an outside attestation from Gallagher, aligned with ISO 22301 rather than certified to it.

Quantexa presents itself as an independent software vendor, so the bank, insurer or agency running the platform carries the regulatory obligations. It says Q Assist gives traceability in line with the EU AI Act, without stating a risk classification. Supervisors, tax authorities and law enforcement are customers too, which it discloses openly. No license is claimed, as expected for a technology supplier.

A false merge can attach one person's sanctions exposure to another. Name matching error can also differ by national origin through transliteration, script and patronymics. Quantexa trains and tunes its matching models for each country, but publishes no error rates by population. The responsible AI principles listed in its trust center have no evaluation behind them.

A wrong merge can be inspected and adjusted inside the customer's own instance, so the error is traceable. A person wrongly pulled into a network has no way to learn of it or seek correction, and no guarantee stands behind false merge rates.

Input categories are stated, from internal systems to corporate registries, and self hosted customers keep the chain inside their own environment. Q Assist names OpenAI, Gemini and Mistral as models it works with, but not which one runs, so those are examples rather than a disclosure. No subprocessor list appears in the public part of the trust center.

Internal systems, third party feeds and corporate registries are reconciled despite poor data quality, and Unify is sold separately for data unification. Agent Gateway is pitched as working with any language model or agent framework, shown alongside Azure, Google Cloud, AWS and Databricks. A claims accelerator for Guidewire ClaimCenter went on the Guidewire Marketplace in January 2026. Core banking, case management and screening connectors are not named.

Cloud, hybrid and on premises deployment are all offered, along with a managed application on Azure Marketplace, and Google Cloud and AWS are named partners. Quantexa sells sovereign deployments to governments, including HMRC and the Insolvency Service. Hosting regions and tenant separation for its hosted product are not stated.

Quantexa publishes no rates, tiers or minimums, and pricing starts with a sales conversation. It offers a managed application on Azure Marketplace, and Cloud AML is a packaged product for US credit unions and mid size and community banks. Quantexa's own pages give no price or billing basis for either.

Global banks such as HSBC, Standard Chartered, BNY and ING sit at one end of the customer range. Cloud AML is built for US credit unions and mid size and community banks at the other. Insurers, telecommunications firms, tax authorities and law enforcement are named markets, and HMRC signed a £175 million, ten year contract in May 2026.