Alexis Peter Francis has spent more than two decades building data, risk and regulatory platforms in financial services, and he currently leads product for customer and counterparty data at one of the world’s largest financial institutions. A lot of his recent work comes down to a stubborn question banks still answer mostly by hand: who owns whom? Francis led the design of a multi-agent system in which separate agents search SEC filings, corporate registries and company disclosures, then an orchestration layer weighs where the evidence agrees and where it conflicts. The approach cut manual review effort by about half. A human reviewer still attests to every hierarchy.
In our conversation, Francis discussed where agentic AI performs well in regulated settings, the operational work behind meaningful human oversight, and why he brings compliance and legal teams in before the architecture is set.
You’ve deployed agentic AI to automate evidence collection from SEC filings and regulatory disclosures at one of the world’s largest financial institutions. What surprised you most about where the technology performed well versus where human judgment remained non-negotiable?
We already knew AI could summarize regulatory filings, but the surpise was how effective an agentic approach was once we broke the problem into specialized agents and allowed them to investigate the same question independently across different sources.
One problem I worked on involved understanding relationships between legal entities and identifying the ultimate parent based on factors such as majority ownership and voting rights, which is where evidence from SEC filings and regulatory disclosures was used. We could systematically construct these corporate hierarchies using trusted external ownership information matched against internal business data, but certain relationships still required an independent review to validate that the hierarchy was accurate.
That review was traditionally very manual. An analyst might spend hours searching SEC filings, company disclosures, government registries, news articles and other trusted sources to verify an ownership structure.
Agentic AI was a natural fit. Instead of asking one model to search broadly and produce an answer, I led the design of specialized agents around different evidence sources. One could focus on SEC and EDGAR filings, another on company disclosures, and others on corporate registries and jurisdiction-specific government sources. Because this was a global problem, the sources also had to vary by market.
An orchestration layer then brought the evidence together, identified where sources agreed or conflicted, and provided a confidence assessment around whether the systematically constructed hierarchy appeared correct. That approach cut manual review effort by roughly half, and every hierarchy was still attested by a human reviewer.
Where human judgment remained essential was ambiguity. Ownership structures can involve holding companies, joint ventures, minority investments with enhanced voting rights, cross-border entities and restructurings. Two credible sources may present ownership differently because of timing or disclosure requirements. AI can surface those differences extremely well, but determining their significance still requires human judgment.
That’s where I think agentic AI delivers the most value. Let AI perform much of the investigative work (finding, comparing and organizing evidence) and let people focus on the cases where their expertise actually matters.
“Human-in-the-loop” gets used as a reassurance phrase across the industry right now. What does it actually require operationally to make that meaningful in an audit-intensive environment?
I think “human-in-the-loop” is one of the most overused phrases in enterprise AI today because it can sound like putting an Approve button at the end of a process automatically makes the solution controlled.
In reality, making it meaningful requires designing the workflow around human accountability from the beginning.
For a recommendation generated by AI, the reviewer shouldn’t just see the final answer. They should be able to understand which sources were searched, what evidence was collected, why that evidence was considered relevant, where sources agreed or conflicted, and the confidence behind the recommendation.
That transparency changes the conversation. You’re not asking someone to trust the AI. You’re giving them enough evidence to independently validate its reasoning.
It also requires traceability. In an audit-intensive environment, someone may come back months later and ask why a particular conclusion was reached. You need to be able to reconstruct the evidence available at that point in time, the recommendation the system produced and the decision the human ultimately made. Human overrides and the rationale for them are important parts of that audit trail as well.
The other operational reequirement is knowing when not to automate.
You need defined thresholds and escalation criteria where the AI stops rather than continuing to make assumptions. If evidence is incomplete, credible sources conflict or confidence falls below an acceptable threshold, the case should be routed to someone with the appropriate expertise.
Sometimes the best decision an AI system can make is recognizing that it doesn’t have enough information to make the decision.
To me, that’s what meaningful human-in-the-loop really means: not humans supervising every task, but humans exercising judgment at the points where judgment genuinely adds value.
Financial services has always had to balance speed with control. How does deploying agentic AI change that tension, and where do you think most organizations underestimate the governance work involved?
Agentic AI amplifies the tension between speed and control.
Traditional automation generally executes predefined rules. Agentic systems can make decisions about what information to retrieve, which tools or sources to use and how to complete a task. That additional autonomy creates tremendous efficiency, but it also expands the governance challenge.
If I’m using AI to validate a legal-entity relationship, an agent may decide which sources to investigate, retrieve information from several of them, compare the evidence and reach a recommendation. If I’m using it for Financial Crimes, an agent could potentially identify that something material has changed about a business and determine that additional due diligence may be appropriate.
Model accuracy is one part governance. You have to govern the entire decision pipeline: which data sources are trusted, which external sources can an agent access, hat if two sources disagree, how are changes to prompts or workflows controlled, and how do you keep up with changes to regulations?
Those questions can become larger operational challenges than the AI model itself.
Financial institutions already have strong control frameworks, but agentic AI requires extending those controls to autonomous workflows. The important principle for me is that the decision path should remain observable, explainable and reproducible. Good governance is what gives an organization the confidence to scale it.
A lot of AI initiatives stall on data quality before they ever reach deployment. What does it take to build a data foundation that can actually support AI decision-making at enterprise scale?
I’ve spent much of my career working on customer data, identity, master data and enterprise relationship platforms, and one lesson has consistently held true: AI amplifies whatever quality of data you give it.
Building the right foundation starts with trusted data rather than simply more data.
That means clear ownership of critical data elements, consistent definitions, strong lineage and reliable identifiers across systems. If different parts of an organization represent the same customer or company differently, an AI system inherits that ambiguity regardless of how sophisticated the model is.
Legal-entity resolution is a good example. Before AI can reason about the relationship between two companies, you first need confidence that the company you’ve identified externally is actually the same company represented internally. That can involve legal names, trade names, registered addresses, tax identifiers and other business identifiers. Those attributes aren’t always represented consistently across systems or external sources.
A company might appear as “ABC Incorporated” in one system and “ABC Inc.” in another, or one system may contain its legal name while another contains a DBA or brand name. Real-world corporate changes make this even more complex.
For example: JPMorgan Chase & Co., JPMorgan Chase Bank, N.A., J.P. Morgan and Chase—names that are closely related but can represent different legal entities, businesses or brands. Corporate changes add another layer: Facebook became Meta Platforms, Square became Block, and Google reorganized under Alphabet.
And it isn’t just names. Consider a company like Walmart, with thousands of stores and operating locations. One relationship with a financial institution might have been established using its corporate headquarters, another using a regional office, and another using the address of an individual operating location. Add changes in legal names, tax identifiers, addresses or ownership over time, and records for a connected business can look very different across systems.
If you can’t reliably connect those records, you don’t have a comprehensive view of the legal entity or all of its relationships with the institution. That can flow downstream into KYC and risk assessments, aggregation of credit exposure, deposit-insurance calculations and regulatory reporting. In areas such as Single-Counterparty Credit Limits or FDIC Part 370, incomplete entity and relationship data can ultimately contribute to inaccurate risk or regulatory outcomes.
The second piece is context.Data without relationships has limited value. Understanding how customers, legal entities, subsidiaries, counterparties and products relate to one another gives AI the context necessary to reason rather than simply retrieve information.
Finally, data quality isn’t a one-time cleanup exercise. Enterprise data changes constantly. Companies change names and addresses, corporate structures evolve, customers update information and upstream systems change. You need continuous monitoring, validation and feedback loops.
The organizations that are best positioned for AI are the ones that treat data as an ongoing product with ownership, quality standards and governance, not something you clean up once before starting an AI project.
Where do you see financial services firms making avoidable mistakes in their AI implementations, particularly in areas touching credit risk or regulatory reporting?
I see two avoidable mistakes fairly often.
The first is starting with, “Where can we use AI?” instead of, “What problem are we trying to solve?”
The most valuable use cases I’ve worked on actually started with fairly mundane business problems. Analysts spend too much time collecting and validating evidence. Customers are repeatedly asked for information they may have already provided. Legal-entity information becomes stale over time. Data-quality teams spend significant effort tracing discrepancies across systems.
Those are business problems first. AI happens to give us new ways of solving them.
The second mistake is assuming that if AI produces accurate answers during testing, it’s ready for production in an autonomous workflow.
In areas like credit risk, financial crimes and regulatory reporting, accuracy is only part of the equation. You also need consistency, explainability and very clear controls around what happens after the AI produces an answer.
KYC and KYB are a good example, where AI can reuse information from documents customers have previously provided. If AI extracts an attribute with high confidence, using it to prefill a field for the customer to confirm carries far less risk than automatically updating a critical customer or business attribute. The output is the same in both cases.
The same applies to adverse-media monitoring, where AI can help identify potentially relevant information about customers or legal entities across large volumes of external information. High model accuracy doesn’t eliminate false positives, entity-matching issues or differences in source credibility.
This matters more because customers and businesses aren’t static. A business can change its legal structure, address or ownership. It can go through a merger, acquisition or spin-off. The information and risk profile you established when the relationship began may not represent that entity several years later.
AI gives us an opportunity to move from primarily periodic controls toward more event-driven monitoring, where meaningful changes can become signals for additional review. But that only works if those signals are connected to well-designed decisioning and governance.
Otherwise, all AI creates a bigger queue for humans to investigate.
What’s the conversation you wish more product and technology leaders were having with their compliance and legal teams before they get to deployment?
I’d like those conversations to start with a shared objective instead of Product talking about innovation and Compliance talking about risk.
Too often, Compliance and Legal are brought in after important product or architectural decisions have already been made. At that point, they are being asked to approve a design they had no part in shaping.
I think the conversation should start much earlier around defining acceptable autonomy and risk together.
That means agreeing on which decisions AI can make independently and which require human approval, which sources are acceptable, what level of explainability is required, what evidence needs to be retained, what should cause the system to stop and escalate, and how success will be measured beyond productivity.
The answers directly influence the architecture and the customer or employee experience.
I’ve found that involving Compliance and Legal early doesn’t necessarily slow innovation down. It can actually accelerate delivery because you’re designing the controls into the product instead of trying to retrofit them later.
For example, when I was working on identifying relationships between legal entities and determining their ultimate parent based on ownership and voting rights, one of the key design questions was when the AI had enough evidence to support a relationship and when it needed to escalate for human review. Early input from our control partners helped us define both confidence thresholds and a hierarchy of trusted sources.
If two credible sources: for example, an SEC filing and an official corporate registry—showed conflicting ownership information, the case would be escalated because the difference could reflect a recent acquisition, restructuring or timing difference between disclosures. But not every conflict required escalation. If a current regulatory filing or official company disclosure showed one ownership structure while an older news article or an unverified third-party source showed something different, the workflow could give greater weight to the authoritative, current source rather than automatically sending the case for manual review.
So, compliance wasn’t simply reviewing the solution at the end—their input helped define which sources we trusted, what constituted a meaningful conflict, and exactly where the AI had to stop and hand the decision to a person.
The organizations that succeed with enterprise AI will be the ones that build trust into the system from the beginning.
