From experimentation to production: Building the Lumin Solaire AI intelligence layer

The conversations about AI in financial services have changed in the last twelve months. In early 2025, most of us were asking “what could AI eventually do in digital banking?” By the end of 2025, the question had shifted to “what is it actually doing right now?” In 2026, the question that matters most is more specific and more pointed: which kinds of AI features actually move the needle for banking users, which kinds don’t, and what should financial institutions and their platform partners be building next?

Lumin has spent time learning answers to those questions. That meant primary research with end users, as well as ongoing conversations with our clients about where their teams need the most assistance. It involved running a number of proof-of-concept experiments and being candid with ourselves about which ones earned the right to graduate to production. What follows is a synthesis of what we learned, what surprised us, and what we are building in response.

What 1k+ users told us

We ran two waves of user research over the past twelve months, averaging more than 500 users per wave, consisting of US-based adults, panel-recruited to reflect consensus demographic distribution. The key finding is straightforward in one direction and more interesting in another. Familiarity and hands-on usage with AI is up. But consumer comfort with AI in their banking app depends on its accuracy, visibility, and intent. On familiarity and usage, the trend lines are clean. The share of respondents who say they “understand AI well” rose 8.7 points, from 22.6% to 31.3%. Combined “basic understanding” plus “understand well” rose 10.4 points. The share of respondents who had never heard of AI or didn’t know what it meant dropped 9.3 points. Hands-on experience with AI chat tools including ChatGPT, Gemini, and Claude climbed 12.2 points in regular or occasional use over the same window.

But comfort within the banking app tells a more nuanced story. When we asked how comfortable users would be with their bank using AI for behind-the-scenes security features, such as fraud detection and unusual transaction monitoring, comfort held essentially steady. Nearly 60% of respondents said they were very or somewhat comfortable, statistically identical to the prior wave. The small movement that did happen was a softening of intensity: “very comfortable” slid 2.1 points toward “somewhat comfortable,” with a tiny 0.4-point uptick in “not comfortable.” Attitudes are settling into cautious acceptance rather than shifting toward skepticism.

The story changed materially when we asked how comfortable would you be receiving financial recommendations or suggestions from AI in your banking app?” 48.7% reported being uncomfortable or unsure. “Very comfortable” dropped 4.1 points, the single steepest shift in the entire survey, from 19.9% to 15.8%. The 3.5-point rise in “somewhat comfortable” suggests some of that enthusiasm is migrating to cautious openness rather than disappearing entirely. But the headline is that consumers are growing more reserved about AI giving them financial advice, even as they become more comfortable with AI generally.

When we asked what specifically concerns them about AI in their banking experience, the concerns that grew were “data could be misused” (up 5.9 points) and “AI might make errors” (up 4.7 points). The only concern that fell was “I don’t understand how AI works,” which makes sense given the usage trend.

This is the most important insight we took from the research, and it has shaped the way we think about building AI features. Understanding the intent of AI use in digital banking is critical. Users accept the use of AI if it is accurate and used with their best interests in mind. They want visibility into the reasons behind recommendations, as well as the opportunity to validate those recommendations at their discretion before taking any action. When we asked which AI features users were most interested in seeing in their digital banking experience, unusual transaction alerts saw the biggest gain, up 8.8 points to roughly 40%—the most desired AI feature in the survey. Personalized tips followed at +6.6. Users want AI to help them stay safe and stay aware.

What clients are telling us

In parallel with consumer research, we have spent a lot of the last year listening to what financial institutions are asking for. Three priorities surface consistently across nearly every conversation.

Increasing self-service feature adoption

Banks and credit unions have invested heavily in self-service capabilities, but adoption is uneven and the journey from “I have a question” to “I solved it myself in the app” is full of small frictions that turn into support requests. Natural language is the most promising tool here. If a user can describe what they need in plain English, the gap between intent and action collapses.

Managing support volumes

Support teams are absorbing rising volume from increasingly complex user needs, and the unstructured data sitting within in-app messages, calls, and audit logs holds most of the answers. The job for AI here is to turn that unstructured data into structured assistance for the support staff.

Reducing manual tasks for employees

A lot of digital banking administration, such as building targeting rules, drafting forms, writing FAQ content, and configuring campaigns, is work that benefits enormously from a natural-language first pass. The point is not to replace the admin’s judgment. It is to make the FI employees’ jobs faster, more efficient, and easier.

The opportunities are concentrated in the spaces between teams: support to user, admin to system, and system to user, where natural language and contextual understanding can lower the activation energy of work that already needs to happen.

The taxonomy that emerged

When we synthesized the research and the client feedback against the work we had been running internally, three distinct areas of focus emerged. We have organized our Lumin Solaire intelligence layer along those three lines.

Solaire Assist is the category of AI features that enhance existing product workflows. These are targeted, non-agentic uses of AI that make a process less tedious for the end-users on our platform as well as the FI employees who manage it. For example, our Target Manager assistant generates campaign targeting rules from a plain-English business objective — describe the users you want to target, and the assistant builds the rule logic for review. The Secure Form assistant drafts secure forms from natural-language descriptions and can convert uploaded PDF forms into digital ones. Our FAQ assistant proposes new FAQ content based on patterns in knowledge base documents and high-volume support topics. Each of these is a “first pass” feature: the admin still reviews and approves, but the high level of effort of starting from scratch is removed. Without AI, every extra step costs time, energy, and effort, on both the user and the admin side. Reducing the steps to produce a good first draft changes the economics of what gets built and shipped.

Solaire Connect is the category of AI features that bridge self-service to support. For example, our Search Assistant lets users ask questions in conversational natural language and returns contextual self-service guidance with relevant insights (“What are my recurring transfers?” returns not just a list but a small summary, a link to scheduled transfers, and the relevant transaction preview). Message Suggestions analyzes messages live as users write them, suggests self-serve solutions, and flags missing information the support team will need. Message Assistant, within our Admin Portal and intended for FI employees, summarizes long message threads for support staff and proposes template responses they can adjust for tone. The Insights Dashboard tracks the performance of these features so admins can understand how support volumes are being impacted and how AI features are being used by employees.

Solaire Agents contain features that can leverage available tools to gather information and perform tasks, sometimes asynchronously. These sit in the territory the user research was most cautious about, so they have been designed with the trust gap explicitly in mind. The Response Assistant is a planned agent that supports administrators responding to secure messages and live chats, with knowledge base documents, user audit log activity, and user profile context informing its suggestions. Agentic Chat is a user-facing natural-language agent that performs asynchronous tasks the user requests: modeling the impact of a different monthly transfer amount on a savings goal, reviewing recent subscription cost changes, or setting up a new budget category and an alert when spending exceeds a threshold.

What ties all of the features together is an intentional, core principle to deploying AI in digital banking: human-in-the-loop approval for every action. The agent can recommend, model, propose, and prepare, but the human approves before anything actually happens. That design constraint addresses the trust concerns we heard in the research. It also forces us as a product team to design agents that genuinely earn the user’s approval, rather than agents that act first and apologize later.

Why the architecture matters

The Target Manager Assistant, Response Assistant, and Agentic Chat are all genuinely useful in isolation, but all become more powerful when connected to a common ecosystem of data and tools. As that ecosystem grows, so does the utility of its individual parts. New agents provide capabilities directly to the workflows for which they’re designed, but also lend new options to existing agents. 

This is the architectural difference between bolt-on AI and embedded AI. A bolt-on AI feature lives in a single application surface and reasons over whatever data is local to that surface. It can do useful work, but each feature is essentially siloed. Embedded AI lives across the platform and reasons over a unified data and intelligence layer,  which means each new AI feature compounds with the AI features already in place, sharing context and signals, and getting more useful as more of the platform is connected to it.

On a platform assembled from acquired parts, AI tends to ship as bolt-ons. Each integration seam is a place where context fails to travel; the AI in one module can’t easily see what the AI in another module knows. You can ship bespoke AI features that way, but you cannot ship AI that gets smarter with the user over time, because the platform itself does not get smarter with the user over time.

This is the architectural reason we have built Lumin Solaire as an intelligence layer embedded across our platform rather than as a set of point AI features: it is designed not just to run AI, but to get smarter about each user every time it does.

What we’re building, and why

We started with the question that matters most in 2026: which AI features actually move the needle, which don’t, and what should we be building next? The research gave us the trust map. The client conversations gave us the priority order. The architecture gave us the compounding mechanism. Solaire Assist, Solaire Connect, and Solaire Agents are the answer we arrived at—not all at once, and not without honest conversations about what didn’t earn the right to ship. The features are visible. The compounding happens underneath. That is the point, and how we’ll continue to build AI going forward.

Byron Tatman
Principal Product Manager, Lumin Digital

Tyler Anticevich
Senior Product Manager, Lumin Digital