Insights

Frontier models, measured in operating numbers

How UST deployed an Anthropic Claude-powered agentic knowledge platform for a Fortune 500 payments leader, cutting answer times from hours to seconds.

Adnan Masood, PhD, Chief AI Architect

There is a category of enterprise AI announcement that describes a pilot, a proof of concept, or an intention. This is not one of those. Over the past several months, UST's Alpha AI practice built and deployed an agentic knowledge platform for one of the world's largest payments technology companies, running on Anthropic's Claude Sonnet 4.6 through Amazon Bedrock. It is in production, it is measured, and the numbers are worth explaining because they say something about what the UST-Anthropic partnership actually produces when it meets a hard problem.

DIVIDER

The problem: A million documents and a thin bench of experts

Payments is a business built on interpretation. Every servicing decision, every adjustment, every compliance response traces back to a clause in a policy, a contract, or a standard operating procedure. Our client's operations teams worked against a corpus exceeding one million documents, with individual policy documents running past 3,000 pages. Finding the governing language for a specific question meant either reading or asking someone who had already read.

In practice, it meant asking. Routine interpretations were routed to subject matter experts over Slack and email, and those experts became the constraint on the entire operation. Wait times ran from hours to days. Two analysts asking the same question could receive two different answers, with no record of which source supported either one. Across a hundred users, roughly 6,000 expert hours a month were being consumed answering questions the documentation already covered. The knowledge existed. Access to it did not.

The client had tried the conventional remedies. Keyword search, portals, and first-generation retrieval tools all surfaced documents. What operations needed was answers, with the evidence attached.

DIVIDER

The solution: An agentic copilot with governance underneath it

The platform UST built has two planes and a control layer.

The ingestion plane converts raw enterprise content into governed knowledge. Documents arrive with a manifest, are hashed and compared against the current state so that only genuine changes get reprocessed, then parsed and chunked in a way that preserves document hierarchy. Amazon Titan Text Embeddings V2 generates the vectors, and everything lands in PostgreSQL with pgvector, with one detail that matters more than any other: the document's access permissions are stored alongside its vectors. Entitlements travel with the knowledge itself.

The inference plane turns a question into a cited answer in about ten seconds. A query passes through single sign-on validation and guardrail screening, checks a semantic cache so repeat questions cost nothing, then runs a permission-filtered vector search followed by re-ranking. Claude Sonnet 4.6 reads the retrieved evidence, reconciles conflicts between policy versions, and composes a grounded answer with citations back to the approved source. Guardrails screen the output on the way back, and every interaction is logged for audit reconstruction. A LangGraph-based agent runtime orchestrates the whole path.

The control layer is why this passed enterprise review in a regulated payments environment. No data leaves the client's cloud. Users cannot receive answers built on documents they are not cleared to read. Every response resolves to a source a reviewer can open and verify. The system was designed to survive an audit, and the design shows.

DIVIDER

Where the frontier partnership earns its name

UST is an Anthropic Global Premier Partner, and it is fair to ask what that phrase means beyond a logo on a slide. Here is the practical version.

Our engineers work with direct access to Anthropic's applied guidance on agent design, guardrails, and evaluation. When we made architectural decisions about retrieval depth, context assembly, and abstention behavior, we made them informed by how the model was built to be used, rather than by trial and error against an API. Frontier capability arrived at the client with the deployment discipline already attached: entitlement-aware retrieval, evaluation harnesses, observability, and responsible AI governance from design through to audit evidence.

That combination is the point of the partnership. A frontier model without production engineering is a demo. Production engineering without a frontier model is a well-built system that gives mediocre answers. The client received both at once, and the elapsed time from first design session to measurable operating results was a single quarter.

DIVIDER

Why Claude specifically

The architecture around the model is solid engineering, and most of it is portable. The results are not portable because they depend on behaviors specific to the model doing the reasoning.

The first is fidelity to evidence. Claude stays inside the retrieved material, and when the corpus does not answer the question, it says so rather than filling the gap. In a regulated function, that single behavior separates a deployable system from a liability. Every hallucinated answer would have landed on the compliance queue, and the tolerance for that in payments is zero.

The second is long-context synthesis. Answering a real policy question means reasoning across dozens of passages pulled from documents thousands of pages long, including superseded versions that contradict current ones. Claude reconciles them and explains which governs.

The third is citation discipline. Attribution happens at the passage level as a property of how the model generates, without a bolt-on attribution layer trying to reconstruct provenance after the fact.

The fourth is consistency. The same question asked by different analysts returns the same interpretation. That is what converts an AI feature into an operational control, and it is the property compliance leaders care about most.

DIVIDER

The results

Measured across a hundred users in payments operations following deployment:

Time to an authoritative, cited answer moved from hours or days to roughly ten seconds. Manual escalations to subject matter experts fell 80 percent, returning about 6,000 expert hours to the business every month. Compliance incidents fell 90 percent, driven by consistent interpretation and audit coverage above 95 percent on answered questions. Business throughput rose more than 90 percent with the same headcount. The reclaimed expert capacity alone is worth roughly $960,000 a year, a figure that deliberately excludes avoided compliance cost and the revenue effect of faster decisions.

The structural change matters more than any single number. Policy knowledge stopped being something the organization held in a small number of heads and became something the organization can query.

DIVIDER

What comes next

The expensive work is done once. The same ingestion pipeline, entitlement model, and guardrails now extend to additional corpora and additional functions, from contact center assist to servicing and contract repositories, at incremental cost rather than as a new build. That is the compounding logic of the platform, and of the partnership behind it: every subsequent deployment starts from production-proven components and a model roadmap we can see coming.

Enterprises evaluating agentic AI tend to ask whether the technology is ready. After this deployment, our answer is that readiness was never really about the model. Claude has been ready. The question is whether the engineering around it treats governance, entitlements, and evidence as first-class requirements. When it does, the results arrive quickly, and they arrive in the operating numbers where a CFO can find them.

AI value is measured in operating results, not experiments. Discover how UST Alpha AI helps enterprises translate frontier AI capabilities into production-ready solutions that improve productivity, governance, and business performance.

Explore UST Alpha AI

UST is an Anthropic Global Premier Partner. The Alpha AI practice designs and deploys frontier model solutions for Global 2000 enterprises across financial services, healthcare, and technology.