SAP Fioneer

The Clearwater precedent settled by a single word the vendor chose itself: the artificial intelligence agent is generally available as an add on to the existing banking, insurance and finance products. An add on is by definition not the product. Strip it and core banking, commercial lending, policy management, pensions administration and the financial products subledger all continue to run for roughly twelve hundred institutions.

The learned line is real, dated and generally available rather than announced, which carries the build, and the lending specific agents for onboarding, underwriting and portfolio monitoring are described in the company's own material as capability it is building rather than as shipped, which is honest scoping and is recorded as such.

No gate described, and the published material pulls in two directions without reconciling them. The shipped product page states that the agent interprets natural language prompts to trigger actions, surface insights or automate tasks, and triggering an action is the agent acting rather than advising, with nothing stating a threshold, an approval step, a confidence band or what it may not do.

Separately, forward looking lending material describes an underwriting agent that aggregates data, checks covenants, builds a risk profile and delivers a structured decision package to the underwriter, which is a properly described handoff, but that appears in material about capability being built rather than in a product specification. Auditability is asserted as a standard the product complies with and is not described as a mechanism.

One genuine mitigation is recorded rather than credited: the interaction model is prompt driven, so a person initiates each task, which is a real property of the design and is nowhere published as an oversight commitment.

Transparency is claimed as a property of the output and no method sits behind it. The agent is described as delivering contextual, transparent and actionable results and as trained on finance grade logic and models, and neither phrase identifies what was trained, on what, or how it is evaluated. Absent entirely: any accuracy or error rate, any validation methodology, any backtesting, any drift monitoring, any versioning and any external assessment.

The gap is sharpened by the bring your own model architecture, because a customer supplying its own model needs to know how the vendor's surrounding logic behaves across different models and nothing addresses that at all.

One of the deepest named customer sets on this roster and not one number attached to any of them. The company's own case study library carries at least seven named institutions each with a named executive on record by title, including a head of financial and regulatory reporting at a global reinsurer, a head of information technology at a German mortgage bank, an executive vice president at the largest bank in Honduras, a president and chief executive together with a chief information officer at a Canadian lender, and a senior vice president for technology at a Japanese insurer.

Scale is stated at around eight hundred banks and more than twelve hundred institutions overall, and independent assessment includes a functionality standout placement from Celent in a review of thirty one policy administration platforms.

Held at B because the quantified result published for the lending product, roughly twenty percent efficiency improvement across the whole loan process, belongs to an unnamed European specialist bank, and the named executives speak to transformation journeys rather than to measured outcomes.

The strongest architectural position on this roster and still not a commitment. The company states institutions can automate processes, gain insights and make decisions without needing to share data externally, and the bring your own model option means a customer can run inference against a model it controls. Together those materially reduce the exposure this axis measures.

Graded C because neither says whether customer data trains or tunes anything, whether anything is pooled across the twelve hundred institutions on the platform, what an agent retains from a session, or what is excluded from a training corpus. The standing rule applies and applies where it costs a vendor rather than where it saves one: an architecture that makes a bad outcome unlikely is not the same as a commitment that it will not happen.

Full compliance with data privacy and auditability standards is asserted, and an assertion of compliance is not a disclosure of practice. No retention schedule, no deletion terms, no subprocessor list and no tenant separation statement. The statement that data need not be shared externally is the most substantive thing published and it addresses where data goes rather than what is kept, for how long, or by whom. The exposure is meaningful because the subledger product holds instrument level accounting data and the lending product holds borrower documentation.

No certification, attestation report, audit period or trust portal was located in the material reviewed, so the grade follows the standing rule that a credential must be found and named rather than inferred from the parent company's stature.

This is unproven absence rather than evidenced absence and the queued check is cheap: a company delivering on a major enterprise vendor's cloud infrastructure will inherit and hold attestations, and the likely reason none surfaced is that they sit with the parent rather than on the product pages. Worth recording that inheriting a parent's security posture is not the same as publishing one, and a buyer researching this platform does not meet a credential in the product journey.

A software vendor with no licence, registration or supervised standing of its own. The bank service provider question resolved elsewhere in this sweep does not carry across on the evidence available, since the statutory examination authority attaches to performing services for a depository institution rather than to licensing software an institution runs, and this company's delivery model is extensions of an enterprise suite deployed by the customer.

Regulatory language in the published material concerns the buyer's obligations, including regulatory reporting and compliance with country specific requirements, which the company frames as the fragmentation its architecture exists to solve.

Nothing. Development is stated to be aligned with industry standards and no standard is named, no framework is identified, no fairness testing is described, no protected characteristic is addressed and no position is taken on any artificial intelligence regulation, which is a notable silence for a German headquartered vendor selling into European institutions.

The exposure is currently lower than for peers shipping credit decisioning agents, because the shipped capability is natural language interaction with data rather than a decision about a customer, and that is recorded as mitigation rather than as disclosure. The underwriting agents described as being built would change that, and nothing published indicates a governance position arriving with them.

No recourse position published, and the bring your own model architecture creates a division of responsibility this index has not seen before. When a customer supplies its own model, runs it against the vendor's agent logic over the vendor's product data, and the output is wrong, responsibility divides between the model the customer chose, the retrieval and orchestration logic the vendor wrote, and the institution that acted on the result. Nothing addresses that split. The conventional exposure is present too, since the shipped agent can trigger actions in banking and insurance processes and nothing states how a wrong action is detected, reversed or contested.

A genuinely different and buyer favourable answer to this axis. Rather than naming a model it has chosen, the company publishes the architecture of the model dependency itself: the agent supports bring your own large language model strategies, it can alternatively use models from the parent platform's own artificial intelligence service, and it interoperates with the parent's assistant and with a named third party agent product.

That tells a buyer three things most vendors here withhold, namely that the model layer is swappable, that no undisclosed provider is locked in beneath the product, and that the institution can place the dependency with a supplier it has already assessed. Held at B and not A because no specific model or version is named for the default path, no model inventory is published, and nothing states which capabilities work with which models or how behaviour varies between them. Recorded as a new shape for the catalogue: giving the customer the choice is a legitimate answer to the supply chain question rather than an evasion of it.

Graded A on the same basis as the other core suppliers on this roster, with one distinction that is worth stating because it is a different kind of system of record: alongside core banking and policy management, the company sells a financial products subledger, which is the accounting book for financial instruments and sits between the operational systems and the general ledger. A vendor holding that position is inside the number that reaches the audited financial statements.

Integration is described as application programming interface first with a native event streaming foundation, and named interoperability extends to the parent platform's technology stack, its data architecture, its cloud infrastructure, its own assistant and third party agents including a named Microsoft product. Delivery is as extensions of the parent's enterprise suite rather than as a standalone stack.

Cloud delivery rests on the parent company's global cloud infrastructure and the products are delivered as extensions of its enterprise suite, and one claim reaches toward this axis without landing on it: the company states institutions can use the agent without needing to share data externally. That is a statement about data movement rather than data location.

Nothing published states region availability, residency commitments, or how data is kept inside a jurisdiction, for a customer base spanning Germany, Romania, Honduras, Canada, Japan and the Asia Pacific region. The bring your own model option means part of the processing location is the customer's choice, which is a genuine architectural mitigation and is credited on the supply chain axis rather than counted twice here.

No pricing published. The artificial intelligence capability is sold as an add on to existing products, which is a commercial structure stated without a commercial figure, and nothing indicates whether it is licensed per user, per agent, by consumption or with the underlying product. That question is sharper than usual here because the agent supports customers bringing their own model, which means part of the cost sits with a third party the vendor does not price.

More than twelve hundred banks and insurers, with around eight hundred banks specifically on the banking platform, spanning two separate industries with genuinely different products rather than one industry described broadly. Named customers demonstrate the range rather than merely asserting it: a global reinsurer, a German mortgage bank, a Romanian commercial bank, the largest bank in Honduras, a Canadian lender serving new immigrants and the self employed, a Japanese insurer moving its global core, a pensions and insurance administrator, a renewable energy investment firm and a German state promotional bank. Buyer types therefore include reinsurance, mortgage lending, commercial banking, specialist lending, pensions, promotional banking and insurance policy administration.