Architecting Trust: From Agentic AI to Autonomous Risk Intelligence
The FIS and Anthropic announcement is more than another “AI in banking” headline. It marks agentic AI entering one of the most sensitive and high-consequence corners of financia...
Surendra Reddy15 min read
FIS and Anthropic are building a Financial Crimes AI Agent meant to support AML investigations. It assembles evidence across bank systems, checks activity against known laundering typologies, surfaces higher-risk cases for investigator review, and compresses investigation timelines from hours or days down to minutes. BMO and Amalgamated Bank are among the early institutions involved, with broader availability planned for the second half of 2026.
For me this is not just a market announcement. It is personal.
I founded Quantiply in 2014 because I believed financial-crime compliance was approaching a breaking point. Banks were drowning in alerts. Investigators were buried under fragmented data, disconnected systems, manual evidence gathering, inconsistent case narratives, and an overwhelming volume of false positives. The industry had plenty of dashboards. What it lacked was a new intelligence architecture.
At Quantiply we were building toward what I called autonomous risk intelligence: an AI-driven AML and KYC platform that could fuse data, detect complex patterns, cut false positives, explain risk, support human investigators, and leave a regulator-ready decision trail.
In May 2025 I wrote a piece titled “Quantiply Was Early, But It Was Not Wrong.” It was a reflection on the deeper lesson of the company. Quantiply was not only an AML software company. It was an early attempt to build a sensemaking layer for regulated enterprises. The lesson was simple.
The FIS and Anthropic announcement gives that reflection a new context. What read as a founder’s retrospective now looks more like category validation.
It would be easy to read this as rivalry, or as old product against new, or as a failed company watching a better-funded pair arrive at the same idea. I want to set that frame aside on purpose, because it is the wrong one. Quantiply saw the architecture before the market was ready. FIS and Anthropic show that the market is now arriving. That is the honest reading, and it is the more useful one.
By January 2019, Quantiply had shipped Predicts Cognitive Intelligence Machine, Release 3.7.0, as part of the broader Sensemaker platform. Predicts was a machine-learning platform built for financial institutions with AML specifically in mind. It handled feature management, model development, model versioning, model deployment, prediction monitoring, explainable AI, and decision augmentation. The architecture was built around transparent, auditable, explainable AI, so the platform could be audited end to end while giving case analysts intuitive explanations to work from.
The technology was promising. The market’s trust model was not ready.
Today the language has changed. We call it agentic AI. We talk about human-in-the-loop workflows, traceable reasoning, source-linked conclusions, model governance, and accountable automation. The underlying question has not moved. How do we use AI to build accountable, explainable, human-governed decision infrastructure for financial-crime compliance, rather than just a faster way to generate more alerts? That is the real story.
What I saw early at Quantiply
When I founded Quantiply, I did not think financial-crime compliance was fundamentally a rules problem. Rules, scenarios, thresholds, and typologies all mattered. They were not enough. The deeper problem was sensemaking.
A transaction is rarely suspicious on its own. A customer is rarely high-risk because of a single attribute. A pattern is rarely meaningful without context. The real AML question is always what a given signal means in relation to everything else the institution already knows.
That demanded a different architecture. It needed data fusion, contextual enrichment, graph-like reasoning, behavioral fingerprints, explainable scoring, human-in-the-loop learning, and the ability to adapt as new patterns appeared. Above all it needed trust. That was the heart of the Quantiply thesis. We were building an intelligence layer that could help regulated institutions make faster, better, more defensible decisions, not another alerting tool.
Looking back, I would put it in one line. Financial-crime compliance is, at its core, an enterprise sensemaking problem. That is why the current agentic AML conversation feels so familiar. The industry is rediscovering the same problem, now with better models, better language, and a more mature appetite for AI-assisted decisioning.
What Predicts and CIM already contained in 2019
The 2019 Predicts and CIM architecture matters because it shows Quantiply had moved past the conceptual stage. Many of the components now treated as essential for trustworthy AI in regulated settings were already built.
| Predicts and CIM, 2019 | Today's name | What it secures |
|---|---|---|
| Feature management | Feature intelligence | Meaningful risk signals from messy, fragmented activity |
| Scenario-based feature generation | Typology-aware features | An agent that understands how each alert was generated |
| Titan | Regulated MLOps, model governance | A lifecycle that evolves without losing accountability |
| Model versioning | Forensic traceability | Reconstructing which model and data produced a decision |
| A.S.Q. explainability framework | Explainable AI as a core layer | Narratives for analysts, compliance, regulators, and the board |
| Decisioning layer | Decision augmentation | Investigator judgment strengthened, not replaced |
Feature intelligence from raw activity
Predicts had a feature-management layer that pulled features out of raw customer, account, transaction, and watchlist data. Those features could classify transaction data, segment customers by socio-economic profile and account and transaction activity, and help unsupervised models learn the difference between suspicious and unsuspicious actors. That sounds routine. It is foundational. AML is not about feeding transactions into a model. It is about turning messy, fragmented activity into meaningful risk signals, and a model is only as good as the representation of behavior it receives. We were turning raw financial activity into feature intelligence, which is exactly what a modern agentic AML system has to do before any agent can reason well.
Scenario intelligence and typology-aware features
We also built scenario-based feature generation. Predicts mapped raw data into suspicious scenarios with look-back periods, value and volume thresholds, repetition logic, and binary feature representations. It could take something like frequent ownership change, test whether it crossed a threshold over a defined window, and convert that into a signal for the model. This matters because an AML agent cannot just sit on a pile of alerts and summarize them. It has to understand how those alerts were generated: the typologies, thresholds, behaviors, look-back periods, and scenario logic behind each signal. Without that, it is a fluent interface over an opaque machine. We were already building that typology-aware feature architecture, and that was the real intelligence layer.
Automated model development and lifecycle
Titan was our framework for model development and management. It automated hyperparameter sampling, training, validation, testing, model ranking, and deployment, supported supervised or unsupervised approaches depending on label availability, and helped users find and deploy the best-performing models. Today we would call that regulated MLOps, AI lifecycle management, or model governance. The idea was the same then. A financial-crime AI system cannot be a one-off model. It needs a lifecycle that includes experimentation, validation, test discipline, performance monitoring, and controlled deployment, and it has to evolve without losing accountability. That was already in our thinking.
Model versioning and forensic traceability
Predicts included model versioning so an institution could go back and determine which models produced predictions during a given period and attribute prediction quality to the right model. This is one of the most important lessons for the agentic era. In regulated financial services, the value is not only whether AI can produce a recommendation. It is whether the institution can reconstruct the decision later: which model ran, what data it saw, which scenario contributed to the signal, what feature changed, what threshold applied, what version was live, what the investigator accepted or overrode, and whether the bank can explain all of it to a regulator months later. That is why I have come to believe that in regulated AI, auditability is not a compliance feature. Auditability is the product.
Explainable AI as a core layer
Predicts did not treat explainability as an afterthought. It carried explainable AI as a core component, designed to explain the model’s decision process and give analysts simple, intuitive narratives. We built our own XAI framework, A.S.Q., because in AML every prediction had to be explainable in a regulated context. This was one of the hardest parts of the market at the time. Many institutions liked the promise of AI but were not yet comfortable trusting it in a workflow where the consequences were regulatory, reputational, operational, and human. They were right to be cautious. An AML system that cannot explain itself creates more risk than value. It may speed investigations and accelerate opacity at the same time. The analyst needed the explanation, and so did the compliance officer, the model-risk team, the regulator, and eventually the board. That conviction is more relevant now, not less.
Decision augmentation and case prioritization
Predicts had a decisioning layer that automated augmentation and prioritization of alerts and cases while giving the user a full snapshot of suspicious activity. That is close to what agentic AML platforms now promise: evidence assembly, case prioritization, investigator support, narrative generation, and traceable decision support. The important distinction is that we built Predicts to strengthen the investigator’s judgment, not to replace it. In AML the goal cannot be “computer says suspicious.” The goal is a system that helps the investigator understand why a pattern matters, what evidence supports it, what uncertainty remains, and what action can be defended. That is the line between automation and augmentation, and it matters deeply.
Why timing mattered
The hard truth is that Quantiply was early. When we were building autonomous AI for AML and KYC, the industry did not yet have the trust vocabulary for it. Calling something “autonomous AI” in AML sounded dangerous. Regulators were not ready, banks were cautious, model-risk teams were uncomfortable, explainability standards were still forming, and data integration was painful. The operating model for AI-assisted investigation was immature, and the phrase “agentic AI” did not exist as a category yet.
This is one of the harder lessons a founder can learn. Sometimes the future is visible before the adoption architecture exists. A buyer can agree with your problem statement and still not be ready to change. The market can need your solution and still lack the trust infrastructure to absorb it. A technology can point in the right direction and still arrive before the category exists to receive it. That was Quantiply. We were not wrong about the direction. We were early on the trust curve.
Then COVID compressed the market further. Capital got harder, enterprise adoption slowed, and strategic investors had their own pressures. GPV had invested in Quantiply, but the timing of autonomous AI for AML and KYC was still difficult, and COVID made it harder. I made the practical founder’s decision to preserve the company’s work rather than let it die. Selling Quantiply preserved the thesis rather than ending it. Now, with FIS and Anthropic bringing agentic AI into AML investigations, I see that thesis reappearing in a new language.
What FIS and Anthropic validate
The announcement confirms several things I believed at Quantiply.
AML investigation is close to an ideal agentic workflow. It requires multi-step reasoning across transactions, counterparties, entities, behavioral patterns, KYC records, adverse media, prior cases, typologies, thresholds, analyst notes, and regulatory expectations. That is precisely the kind of work where an agent that can gather and connect evidence earns its place.
The investigator stays central. The reporting around the FIS and Anthropic effort stresses that the agent supports investigators rather than replacing them, and that humans keep decision-making authority while the system gathers and evaluates evidence more efficiently. That matches the augmentation principle Quantiply was built on.
Evidence assembly is one of the biggest bottlenecks. FIS describes the agent helping banks assemble and evaluate information across systems, with the potential to compress case review from days to minutes. Evidence gathering was always one of the core pain points in AML. Investigators spent too much time collecting and not enough time exercising judgment.
AML is becoming a platform problem rather than a point solution. FIS is not bolting a chatbot onto AML. It is embedding agentic AI into banking infrastructure, compliance workflows, data systems, and investigator processes, while Anthropic continues to expand its broader financial-services agent strategy across finance, operations, research, and client-coverage work.
And agentic AML has to be governed from the start. The more AI participates in investigations, the more source-linking, explainability, audit logs, model versioning, data lineage, human approval, and continuous monitoring matter.
That is the deeper validation of Quantiply. The point is not that FIS and Anthropic built what we built. They did not. The models, interface, infrastructure, and market are all different. The architectural problem is the same one we faced: how do you turn AI into trusted action inside a regulated institution? That was Quantiply’s question. It remains the enterprise question.
Jim Kaskade’s concern is valid but incomplete
Jim Kaskade is right to warn that AML agents could become automated suspicion at scale. That is a real risk. AML already suffers from over-flagging, defensive compliance, and de-risking, where banks optimize to avoid regulatory liability in ways that harm legitimate customers. An AI agent could improve that or make it worse. Used well, it can cut false positives, speed investigations, improve case quality, and let investigators focus on meaningful risk. Used badly, it can industrialize suspicion, amplify bias, bury its reasoning, and make it harder for customers to challenge harmful outcomes.
I would sharpen the question, though. The issue is not only whether customers can understand the model’s reasoning. The deeper issue is whether AML agents are designed as accountable decision infrastructure or as automated enforcement engines. That distinction is everything. An agent rewarded for generating more flags is the wrong design. The right measures are precision, explainability, evidence quality, investigative usefulness, SAR quality, reductions in false positives and false negatives, consistency of reasoning, human override patterns, model drift, threshold changes, demographic impact, and regulatory defensibility. The real opportunity is the shift from alert factories to governed financial-crime sensemaking systems, not a higher volume of investigations. That is the lesson I carry from Quantiply.
Agentic AI and autonomous risk intelligence
Here is the distinction I would draw now. Agentic AI is the current interface and orchestration layer. It gathers evidence, reasons across systems, drafts narratives, checks typologies, and assists investigators. Autonomous risk intelligence is the larger operating model. It governs how risk signals are detected, interpreted, explained, challenged, escalated, learned from, versioned, monitored, and audited over time.
Quantiply was aiming at the second layer. Predicts and CIM were not just a model. They were an architecture for learning, prediction, explanation, and governance. Predicts presented its inferential entities, including relations, attributes, learners, transformers, and predictors, as web resources reachable through a common interface. The goal was to improve interoperability, hide implementation detail, and support federated machine-learning solutions composed from independent services. In today’s language that reads as composable intelligence, interoperable services, federated learning, prediction services, model lifecycle governance, explainable decisioning, and human-in-the-loop augmentation.
That is why this moment matters. The market is excited about agents. The durable category is bigger than agents. It is governed decision infrastructure for regulated enterprises.
The due-process question in AML
Kaskade also raises a real question. If an AI agent helps decide who gets investigated for financial crime, what rights should the customer have to understand and challenge the model’s reasoning? In AML the answer is complicated. Banks cannot always disclose investigative logic, thresholds, typologies, or SAR-related reasoning to a customer without creating tipping-off risk. Customer explainability in AML cannot work the way consumer-credit explainability does.
The better framing is to ask what due-process architecture should exist when AI materially influences an AML investigation, an account restriction, a customer offboarding, or enhanced due diligence. I think that architecture needs at least four layers.
Four layers of due process
Internal
Investigators and compliance leaders should see the evidence, the rationale, the source links, the confidence and uncertainty, the scenarios, the model version, and the alternative interpretations.
Regulators
Examiners should be able to review why the system behaved as it did, how models were tested, how drift was monitored, how thresholds changed, and how humans governed the final decisions.
Customers
Anyone affected by an account closure, restriction, or enhanced review should have a meaningful path to correct data, challenge an outcome, and request human review, without exposing sensitive AML logic.
Governance
Boards and risk committees should see the aggregate picture: false positives, false negatives, demographic impact, escalation rates, override rates, model drift, investigator disagreement patterns, and decision quality over time.
That is the difference between transparency and accountability. AML cannot always be fully transparent to the subject of an investigation. It can always be accountable.
How the thesis has evolved
My May 2025 article was reflective. It asked what Quantiply taught me. This moment is evidentiary. It asks what FIS and Anthropic now validate, and the answer is coming into focus. The lesson that high-consequence AI adoption rises or falls on trust architecture is being borne out as agentic AI moves into exactly those workflows. The claim that AML is a sensemaking problem is showing up in how the market now centers evidence assembly, typology reasoning, and investigator augmentation. The insistence that explainability and auditability are not optional has become a baseline adoption requirement through source-linked reasoning and investigator control. And the conviction that autonomous risk intelligence was always bigger than rules, dashboards, and static alerts is visible in the move toward governed, AI-assisted investigation. The market is doing more than discovering AI for AML. It is rediscovering the architecture Quantiply was reaching for.
My take
This is a major validation moment for Quantiply. The market is now saying out loud what I believed when I founded the company in 2014. What AML needs is a trusted reasoning layer that can assemble evidence, interpret behavior, support investigators, cut false positives, produce regulator-ready narratives, preserve human accountability, and govern the whole decision lifecycle.
The irony is that Quantiply looked too early when “autonomous AI for AML and KYC” sounded risky. In 2026 the same idea is being repackaged as agentic AI, and the market is finally ready to understand it. The lesson is that Quantiply was operating ahead of the trust curve, not that it failed. Predicts and CIM show we were already working in terms of scenario intelligence, feature vectors, model lifecycle automation, explainable AI, decision augmentation, model versioning, and end-to-end auditability. Those are foundational capabilities for trustworthy agentic AI in regulated environments, not peripheral ones.
That carries a direct implication for 451 Ventures and the SuperCIO thesis. The next wave of AI value in regulated enterprises will not come from agents alone. It will come from governed agents embedded inside accountable decision infrastructure, and the SuperCIO is the leader who understands that distinction. The CIO role has kept changing through each platform shift. An earlier generation managed systems. The next modernized platforms. The cloud era optimized infrastructure. The data era built analytics capability. The AI-era SuperCIO has to architect trusted decision infrastructure, which means connecting data, AI, workflows, governance, security, talent, operating models, and business value into one coherent enterprise capability.
Quantiply gave me that conviction before the market had fully arrived. Now the market is arriving. The future belongs to enterprises that learn to make sense, decide, act, govern, and reinvent continuously, not to the ones that merely adopt AI. That was the Quantiply vision, and it is what FIS and Anthropic now validate. The journey from Predicts and CIM to agentic AML is not a postmortem. It is the next chapter in the long arc toward the AI-native enterprise.