In a live demo, Actian's Phil Wall showed a compliance lead using plain-language questions to find low-risk bank customers breaching SAR thresholds. He explained how governed metrics, glossaries and semantic SQL keep the AI analyst from making up answers.
Key takeaways
- One question replaced a report request. The demo analyst planned, queried, charted and wrote a compliance report in about one to two minutes.
- It surfaced patterns, not just numbers. The analyst flagged possible structuring, uniform activity and missing enhanced due diligence.
- Business context is the trust layer. Governed metrics, glossaries and models ground the answers, Wall said.
- Saying 'I can't answer' is a feature. Asked a question its data couldn't support, the analyst offered leads instead of guessing.
- Semantic SQL goes beyond text-to-SQL. An LLM interprets the question, then SemQL translates the plan for each database.
When a business user has a question no existing report answers, the usual options are a report request, a service ticket or a do-it-yourself data download. In the Data Science Connect webinar "Your Data, Your Questions: Self-Service Customer Insights Without the Bottleneck," Phil Wall, Senior Sales Engineer at Actian, demonstrated a different route: asking a governed AI analyst in plain language.
The session, presented by Actian, walked through Actian AI Analyst and the Actian Data Intelligence Platform using a compliance scenario built on know-your-customer (KYC) and anti-money laundering (AML) data.
How can compliance teams use AI for AML and KYC investigations?
Wall played the head of financial crime compliance at a retail bank, with a compliance committee review the next day. A team member had flagged customers rated low risk whose activity sat close to or above suspicious activity reporting (SAR) thresholds. Without a prebuilt report, he said, that kind of research "can take hours, days, possibly weeks."
He typed one question: did the bank have low-risk customers who exceeded the SAR threshold in the last 30 days? The analyst built a plan, pulled data from connected systems, broke results down by the bank's defined risk factors, and produced charts and a full compliance report. Wall said the run takes one to two minutes.
In the demo data, the analyst found about 93,000 customers with risk factors between 1 and 11 who had breached the $10,000 SAR threshold over a rolling 30-day window. It also flagged transaction amounts clustering just under the threshold, a possible sign of structuring, unusually uniform patterns that Wall said real customers don't show, and no enhanced due diligence coverage for the entire flagged group. It closed with recommended actions, such as reclassifying customers, triggering enhanced due diligence and filing SARs.
How does conversational analytics go beyond dashboards?
Wall then drilled in. Asked for the five customers with the biggest mismatch between their KYC classification and their transaction behavior, the analyst created its own scoring formula, explained it, and ranked the customers. A follow-up asked which transaction channels drove the exposure. The answer pointed to channels such as foreign wires, ATM cash deposits and Zelle peer-to-peer payments, each mapped to a typical laundering pattern.
Traditional dashboards from tools such as Tableau, Power BI and MicroStrategy "really only address the questions that they were built to address," Wall said. Analysis is rarely linear, he added, and a conversational interface lets users follow the trail. "Notice I've created no report requests," he said.
How does business context reduce AI hallucination?
Wall said the answers can be trusted because they are grounded in governed business context: defined metrics, glossary terms, risk factor definitions and data models. That context, he said, keeps the AI from "hallucinating off of something it read on Reddit," querying the wrong data or applying inconsistent definitions.
He showed what happens when the analyst can't answer. Asked why the customers were originally classified as low risk, it said it couldn't answer the question directly. Instead it listed leads to investigate, including an unusually flat onboarding cadence and customers concentrated in eight states. Other AI systems "are always wanting to help us," Wall said, while this one can propose information "but it cannot make up information."
During Q&A, he said guardrails based on answer confidence can invalidate an analysis and force a different approach, backed by scoring and challenger methods.
What is the difference between text-to-SQL and semantic SQL?
Asked whether the product is mostly text-to-SQL, Wall said no. Text-to-SQL takes "show me sales" and finds something that looks like sales, he said. Actian's analyst first uses an LLM to interpret what the user is asking, then queries through SemQL, a semantic SQL layer that translates the plan for each underlying system, since Snowflake, Databricks and SQL Server each work differently.
How do you set up an AI data analyst?
In the analyst studio, administrators define each analyst's instructions, data models, metrics, glossary and access rules. Separate analysts can serve sales, compliance or support. Metrics matter, Wall said, because organizations calculate figures such as net sales differently.
A data steward agent helps build that context. It connects to data sources and catalogs, proposes dimensions and measures, and prepopulates instructions and glossary terms. Wall said his recorded setup took about an hour because he double-checked his work along the way. The steward also monitors usage, emails administrators when performance drops, and suggests turning frequently used terms into glossary entries. All activity is audited and logged, which Wall said matters when compliance teams must prove how they reached an answer.
He said the product is a SaaS offering that connects to most major warehouses, including Snowflake, BigQuery, Databricks, Redshift and SQL Server.
See more live AI demos
Data Science Connect hosts 24 free live webinars on its 2027 schedule, including the AI Demo Days and Plug & Play series, where vendors show their tools working on real data rather than slides. See the webinar schedule for upcoming sessions.
Frequently asked questions
How can AI help with AML and KYC investigations?
In Actian's demo, an AI analyst answered plain-language questions about customer risk ratings and transactions, flagged possible structuring and recommended compliance actions in minutes.
What is conversational analytics?
Asking questions of business data in natural language and getting analysis back in conversation, with follow-up questions building on earlier answers, as Wall described it.
How is an AI analyst different from text-to-SQL?
Text-to-SQL translates a request into a query. Wall said Actian's analyst first interprets the question with an LLM using governed business context, then queries through a semantic SQL layer.
How do you keep an AI data analyst from hallucinating?
Ground it in governed metrics, glossary terms and data models, and use confidence-based guardrails that invalidate analyses that go beyond them, according to Wall.
Watch: Your Data, Your Questions: Self-Service Customer Insights with Actian AI Analyst | Webinar Replay on YouTube
Event: Webinars · More coverage: Webinars insights
