Alekh Jindal: Co-Founder & CEO of Tursio

1. Core banking platforms have been in production at many credit unions for decades. The data structures they created weren't designed for modern querying, and replacing them isn't a realistic option for most institutions. What does that mean for any analytics tool trying to sit on top of that infrastructure?

Given that core platforms are sticky and any migration takes years, a modern analytics tool needs to work with where the data already is: the core databases, the on-prem data warehouses, or even the cloud data warehouses. Furthermore, core data is often connected to other data sources that track member journeys, e.g., CRMs, CDPs, web analytics, and so on. This requires linking all these sources together, interpreting the complex schemas and data structures of each, performing in-situ data transformations to make them queryable, understanding the domain-specific semantics needed to serve business users, and making the relevant pieces of data accessible to stakeholders securely and in natural language — all of which is fortunately possible today with AI.

2.  The argument for specialization is the industry context which has to be built in, not bolted on. How much of Tursio's value sits in that embedded context versus the underlying technical architecture and how long did it take to build that knowledge layer?

Tursio's primary value sits in the embedded context, not the underlying technical architecture; the architecture is what makes that context fast to build and keep current, but the context itself is the differentiator. In fact, industry context is critical for any AI to work. Unfortunately, there is no one-size-fits-all: each organization has its own nuances, even within the same industry, and all of them must be incorporated for the AI to behave just like that organization. Tursio’s architecture assists that by ingesting database schemas, BI reports, and query histories to bootstrap a context graph that mimics the current organizational understanding — a well-defined process that completes within a few weeks. In addition, many of the nuances remain undocumented and constitute the bespoke knowledge in people’s heads. Tursio records feedback from all interactions, from users and agents, and attaches it as annotations to relevant portions of the graph. This part of the process is ongoing, and it treats context as a living data structure that continually evolves over time.

3. Trust is the word that comes up in every regulated industry conversation. What does it mean for an institution to trust an AI system that is making queries against their most sensitive data and how does the deployment model address that rather than just the contractual terms?

Trusting an AI system is a leap of faith, and it is our job as a vendor to facilitate that. Tursio does this by deploying on-premises and connecting to data sources within the credit union environment, so that all responses come from their data and are not hallucinated from external knowledge. On-premises deployment further keeps the access boundaries within the credit union, protecting member data from any accidental leakage. Also, Tursio uses LLMs from credit unions’ own Azure subscription, wherein all prompts are fully private and guaranteed not to be used for any model training anywhere. By securing data sources, data access, and LLM prompts, Tursio puts the credit union fully in control of their AI deployment and helps build trust from the get-go.

4.  The "your subscription, your tokens" model you've described changes the vendor incentive structure. When the vendor's revenue doesn't scale with token consumption, why does that matter specifically in an AI/BI context where query efficiency has a direct cost implication? 

Many AI applications are LLM wrappers that meter usage with tokens; essentially, they resell tokens and rely on growing token usage over time. This is a problem for AI/BI since the number of queries can grow very quickly and vendors have little incentive to optimize them. Tursio flips this model by allowing credit unions to plug in their own Azure subscription with their choice of Azure-compliant models, while Tursio's revenue grows with the size of the context (data sources and organizational nuances) that it helped the credit union build. This means Tursio is focused on optimizing token usage, rather than bloating it, since that will lead to more adoption and hence more context to be built. At the same time, the credit unions leverage the enterprise-grade governance and control provided by Azure across their entire AI estate.

5. The goal of natural language database search is that a non-technical executive can ask a question and get an answer without knowing how the data is structured. How close is that to working reliably in production and where does it still require human expertise in the loop? 

The reliability of natural language database search depends on the quality of the context. For Tursio, everything captured by currently documented semantics (schemas, query histories, Power BI reports) can already be answered reliably in production. However, human expertise is still needed to help capture the bespoke knowledge that is not yet documented or that keeps evolving over time. Tursio provides the tools and processes to continue capturing and vetting that additional context with expert assistance. In terms of analysis types, Tursio’s data search can be layered with agents like Copilot, Claude, or others on top, via its Model Context Protocol (MCP) server, for rich analysis ranging from descriptive all the way to prescriptive.

6. The data governance requirements that protect sensitive information also tend to restrict who can access it and how. How do you design an analytics system that is open to non-technical users while maintaining the access controls and audit trails that a regulator would require?  

Managing data access controls is one of the first questions when opening sensitive data to querying by non-technical users. Given the shared context graph that Tursio operates, it is uniquely positioned to enforce access controls centrally, in one place. Specifically, Tursio integrates with existing enterprise authentication (e.g., Microsoft Entra ID) to enforce access control at a fine-grained level. Admins can create one or more logical datasets that define the querying scope (tables, columns, etc.) from each database and decide which users (Entra IDs) or user groups (Entra groups) have access to which datasets. They can further apply table-, row-, or column-level access controls (TLS, RLS, CLS) on a given dataset, again via their Entra ID. Likewise, Tursio’s MCP server gatekeeps all access via users’ Entra IDs, so they only get to search what they have access to. In addition to Entra ID, the Tursio search portal lets admins organize users into roles such as admins, owners, users, and viewers, and personalize the search experience and visibility for each role — e.g., viewers cannot download any data artifacts. For enterprises managing data governance via Microsoft Purview, Tursio can connect and sync definitions and access controls across all data products. All data controls, as well as data access, are logged along with the identifying Entra IDs and roles for audit purposes. These logs are further compressed in-situ as they grow, or users can connect them to other log-management systems, such as Splunk or Grafana, for observability.

7. You've written about natural language becoming the interface layer for data not as a feature on top of existing BI tools, but as a replacement for the query-and-dashboard model. How far away is that as the production reality for organizations, and what has to change at the infrastructure level before it gets there?

The query-and-dashboard model has one fundamental problem: the questions come from business users, but they must wait through long dashboarding cycles before getting the answers. Since these questions are in natural language anyway, LLMs can now connect business users directly to enterprise data, cutting down layers of inefficiency and unleashing new levels of productivity. Still, the gap remains in the context layer that is not yet built, which is actually an infrastructure problem. Building and managing context that can be harnessed to serve business users is the change organizations must go through. Tursio is one such effort in this direction, but the trend of opening highly valuable business data to business users and agents is likely to amplify in the near future. That also means rewiring the data and BI teams into context teams that build reliable context with proper security and access controls built in to make it ready for agents. Most organizations are already thinking in that direction, but it is still early days, with efforts being contained and siloed to avoid any big disruption. I expect organizations to build more confidence and move far more quickly as the context infra matures over the next 6-12 months.

8. As CEO, you're making product bets today on where data access is in three to five years. What are you building toward that the market isn't asking for yet and what does the current state of the technology tell you about where the problems still are?   

Today, data is fragmented across systems and processes; it is dirty, with incomplete or even incorrect values; and it is ambiguous, with lots of loaded semantics needed to interpret it — in short, it is very hard to use. Making AI work on such a data estate is highly challenging yet sorely needed — data migration and integration are tediously long, data cleaning is painfully hard, and data semantics are excruciatingly slow to capture. At Tursio, we are building a disambiguation layer that abstracts clean, unambiguous data — usable by AI — from the organization-wide messy data underneath. Automatically inferring this disambiguation layer, refining it over time, enabling sophisticated data access over it, and keeping a human in the loop throughout are some of the challenges we are working on. As LLMs become increasingly more capable, we believe all of these challenges are going to be addressable.