How Alloy’s AI Assistant is transforming KYB workflows

A Q&A with Alloy's Engineering Manager for Agentic AI, Mike Luby

Know Your Business (KYB) has always been one of the most operationally heavy parts of onboarding. It’s messy, fragmented, and often requires human judgment across multiple data sources. That’s where Alloy’s AI Assistant comes in.

We sat down with Mike Luby, Engineering Manager for Agentic AI at Alloy, to talk about how our AI Assistant works under the hood and what it unlocks for compliance and operations teams.

Q: From an engineering perspective, what does the AI Assistant “see” that a human analyst or third‑party AI tool doesn’t when it’s making a KYB recommendation?

The biggest difference is context. Because the AI Assistant was natively built within the Alloy platform, it has access to our full workflow graph: not just the data, but everything that’s happened to that data. This includes the original inputs, rule results, evaluation history, case notes, and final outcomes. It also sees raw outputs from third-party KYB data solutions, along with how those outputs were interpreted by your rules. So instead of trying to piece things together after the fact, it already understands how your system is treating each signal. And since the AI Assistant is customizable to your specific policy, it’s grounded in how your organization defines things like risk levels or escalation criteria, ensuring the AI Assistant is reasoning the same way your team would, just faster and more scalable.

Q: How do you make sure the Assistant’s KYB recommendations are explainable and audit-ready, especially for regulators or internal model risk teams?

Every recommendation the AI Assistant makes can be traced back to exactly which sources were used, what fields were compared, where discrepancies were found, and how it arrived at its conclusion. All of this is logged directly in Alloy, creating a complete audit trail of inputs, outputs, and actions, just like any other decision step in our workflow, which makes it easy to review decisions internally or demonstrate compliance to regulators. An important callout is that the AI Assistant doesn’t ever introduce or assume new logic. It operates within your existing policy framework and aligns with rules your organization has already approved, encouraging more confident adoption by your model risk and compliance teams.

Q: What guardrails are in place so that an institution can dial in the right level of human-in-the-loop vs. full automation for KYB decisions?

There’s no one-size-fits-all approach, so the AI Assistant is designed to be fully configurable to a company’s standard operating procedure (SOP). Many teams start with the AI Assistant handling the prep work, such as gathering information and making recommendations, while a human makes the final call. From there, they can automate specific segments that are lower-risk or well understood. You can also define where human review is always required, like certain jurisdictions or industries, and manage access through role-based permissions. If something doesn’t feel right, there are straightforward ways to turn off automation or fall back to manual workflows, with full visibility into what changed.

Q: KYB research often relies on messy, fragmented data. How does the AI Assistant decide which third-party sources and open-web signals to trust when verifying a business and its UBOs?

There’s a best practice hierarchy for this. The AI Assistant starts with authoritative sources like corporate registries, tax IDs, and verified KYB vendors/sources, and then layers in open-web data as supporting evidence. It also cross-checks key fields like business name, address, and ownership across multiple sources, flagging anything that doesn’t line up. Built directly into our orchestration engine, the AI Assistant already understands where each piece of data came from and how your policies treat that source. That context helps it weigh signals more appropriately. When the AI Assistant presents a conclusion, it links back to the underlying source URLs, allowing analysts to quickly verify without redoing the work.

Q: What have you done technically to reduce hallucinations or bad inferences when the Assistant is doing open-ended business research?

A lot of it comes down to limiting where the model pulls information from and how it uses it. Instead of letting it search broadly, the AI Assistant uses constrained retrieval, meaning queries are routed through curated sources, and only relevant news snippets are passed in. It also anchors recommendations on structured data first. Internal data and vendor outputs are treated as the source of truth, while web data is used more as a secondary layer for confirmation or enrichment. The prompting is also intentionally conservative; if there isn’t enough evidence or sources conflict, the AI Assistant will surface that instead of guessing. 

Q: What lessons learned or best practices would you share with teams trying to roll out KYB automation for the first time?

Teams tend to see the best results when they start with a clearly defined standard operating procedure, or SOP. When policies are documented and consistent, the AI Assistant has a strong framework to follow, which leads to more reliable outcomes. It also helps to begin with simpler, well-understood segments and expand from there, rather than attempting to automate the most complex cases first. Finally, it’s not necessarily a best practice, but having clear success metrics upfront, such as shorter review times, backlog reduction, or higher straight-through processing (STP) rates, makes it easier for our teams to optimize the AI Assistant and measure impact. 

Q: What will a fully ‘agentic’ KYB and business onboarding stack look like in the next six months at Alloy? 

We’re building toward a world where the AI Assistant will evolve from supporting a single review step to acting as a true copilot across the entire business entity lifecycle, from application submission to periodic review, including integration with Alloy’s pKYB solution (coming soon!). 

Alloy workflows will become more dynamic and adjust in real time based on what the AI Assistant uncovers, automatically requesting the right documents, triggering enhanced due diligence, or refining risk assessments as new information comes in. Additionally, with deeper integrations into partner systems, the AI Assistant will be able to run in the background while teams stay in their existing tools, for less impact on their day-to-day work.

And importantly, the AI Assistant will become smarter over time. Analyst decisions, overrides, and regulatory outcomes feed directly back into the AI Assistant, continuously improving how it prioritizes signals, flags risk, and recommends actions. The end state is a KYB process that’s not just faster, but more adaptive, reliable, and scalable.