As AI experimentation becomes mainstream adoption, many financial services firms are still searching for ROI. Too often, data and AI investments fall short—because most AI is not designed to meet the highly specific operating needs of our industry. This series, called “Real ROI: Is Your Data & AI Strategy Delivering Tangible Value?”, examines the challenges that stand in the way, and offers practical solutions for how to achieve real ROI. Read on for Chapter 1.
According to McKinsey, 80% of financial institutions have adopted AI in some capacity. But a recent PwC survey reveals that 77% say they have seen no meaningful ROI from their AI investments. The gap between adoption and impact is wide, and understanding why is the first step toward closing it.
There is certainly no shortage of new AI solutions to try. Every day, new tools promise to transform your daily workflows. Advisor solutions ease meeting prep and CRM, chatbot assistants query internal systems and documents, and bundled packages assemble numerous point solutions like these together. But while such solutions can offer efficiency gains, especially for simpler, lower-stakes tasks, their value is inherently limited. Truly transformative ROI remains elusive, especially as tasks grow in stakes and complexity.
Most AI tools fall short of delivering real ROI for a common reason. Look closely, and a pattern emerges: most off-the-shelf solutions for our industry stack a conversational layer on top of existing infrastructure. The problem with this approach is that the AI is bolted-on instead of built-in. The distinction is more important than ever. Bolted-on AI can help people find answers faster; built-in AI is designed to deliver reliable results that people can trust and act on.
Let’s unpack the differences between bolted-on and built-in AI on these five key topics:
- Data access
- Accuracy
- Security
- User experience
- Human oversight
Data Access: Visiting Your Data vs. Living with It
Commercial AI solutions are consistently built to reach your data through external connectors, such as APIs and MCPs (Model Context Protocols), rather than as native components. Those are legitimate, useful standards. But a connector merely opens the door to your data; it can’t supply the necessary governance, context, and domain knowledge that make an answer trustworthy enough to act on.
We’re not building a conversational tool; we operate the books-and-records systems where accuracy isn’t optional
BetaNXT takes a fundamentally different path, because we’re solving a different problem. We’re not building a conversational tool; we operate the books-and-records systems where accuracy isn’t optional, so our AI has to earn the same trust as the data it works on. That’s why we developed DataXChange as a single, unified ecosystem where data and AI work as one. DataXChange’s foundational data layer is a governed, trusted source of truth for our AI layer, InsightX, to access, analyze, and act upon. Our natively embedded AI inherits the robust governance and rich metadata, schema, and mapping already built into our data layer.
Accuracy: Best-Guess Responses vs. Reliable Outcomes
Accuracy is a follow-on issue from the data source issue. The direct consequence of AI not living with your data is that every response is the model’s best guess. A bolted-on tool typically works via in-the-moment retrieval: it fetches a document, a record, or a snippet of data, hands it to a general-purpose model as text, and asks the model to answer a question about it. This works fine when the answer lives in one place and doesn't depend on relationships between things. This approach breaks down the moment an answer depends on how several things relate to each other—which, in our industry, is most of the time. In that case, you can ask the same question twice and a general-purpose model may reason its way to two different answers, because nothing anchors it to how your firm’s data actually behaves.
For example, a question such as "Which accounts are affected by tomorrow's mandatory reorg, and what do they need to do?" isn't answerable from one retrieved document. It requires knowing which positions are held where, which election deadlines apply to which account types, and which of those are still outstanding. A native system like DataXChange holds all of that information as structured, trusted data, and understands the meaning drawn across multiple data sources. A connector has to reconstruct those relationships on the fly, from whatever text it can retrieve—and reconstruction under time pressure is exactly how confident, plausible, wrong answers emerge.
Our InsightX AI also operates with greater domain-specific intelligence. Rather than build one broad, do-everything agent, we design more specialized models and agents trained for particular data domains and business outcomes. As a result of this Domain Intelligence, the outputs are more predictable and trustworthy, so that our AI can move beyond answering questions to carrying out tasks with human approval. Greater accuracy is how we can feel confident moving from answers to action with AI.
What we’re building will enable you to walk into the office in the morning and have much of your routine work already done—every day, with reliable accuracy that gets even better over time.
Security: New Points of Failure vs. Zero-Trust Protection
As regulatory scrutiny of AI grows, data protection and risk management will only become more critical. A bolted-on approach can introduce serious data security and operational risks. Non-native AI tools often send your data outward to someone else’s hosted model—increasing possible exposure—and every added connection is one more place where access controls must be kept in sync.
When AI reasoning happens within a private, protected environment, your sensitive content and proprietary IP stay under your control. Our DataXChange platform’s Zero-Trust security framework safeguards each client’s raw data within its Client Data Vault, along with all the analysis and insights derived from it. AI learning stays inside that private vault rather than being shared across environments.
Today, when AI agents are acting autonomously alongside human users, strong security means controls over two kinds of access: identity access as well as content access. In other words, the first question is who is asking for access and the second question is what they are allowed to see and do. Without both levels of control, failures are bound to happen.
A verified identity doesn’t mean every piece of content is appropriate for it
A verified identity doesn’t mean every piece of content is appropriate for it: an advisor’s agent, for instance, shouldn’t be able to summarize another advisor’s clients. And fine-grained data permissions mean little if an agent acts under a broad service account.
In our architecture, our access and orchestration layer calls upon the underlying governance and client-specific context to verify both identity and content access, delivering a more secure experience.
User Experience: Context-Switching Friction vs. Smoother User Experience
When AI is added onto existing workflows, it can introduce new friction to user experience. A bolted-on tool almost always lives in its own window, its own interface—separate from whatever system the person was already using. Every use means stopping, switching contexts, asking the question, and switching back. In the workflow of someone handling dozens of accounts a day, a tool that requires switching windows is used when there's time to do so thoughtfully, and cast aside under deadline pressure—which is precisely when a correct answer matters most.
When AI is natively embedded in your operating model and existing workflows, insights surface organically, within the familiar interfaces and systems of record your teams already use daily. User entitlements and permissions carry over, since the AI runs on the same platform that manages your systems of record. This intuitive integration leads to greater user adoption.
Human Oversight: Last-Mile Checks vs. Human-in-the-Loop Throughout
While most AI solutions providers talk about human-in-the-loop oversight, it’s important to examine the position and depth of that human aspect. When AI reasoning happens separate from your data infrastructure and ingrained business rules, it can feel like black-box intelligence that offers human users only a chance to approve or reject responses at the end of the line.
By contrast, we infuse human-in-the-loop oversight throughout our AI capabilities—from the build to the resulting business outcomes. At the foundation of InsightX, our subject matter experts translate their collective decades of institutional experience into proprietary Domain Intelligence Packs (DIPs). Each DIP packages human understanding of operational workflows, requirements, and realities for a particular function or process into a form that our AI can ingest and activate. Our SMEs are continuously maintaining and improving upon InsightX’s Domain Intelligence, so it can keep pace as the industry evolves. (More on this topic in Chapter 2.)
We infuse human-in-the-loop oversight throughout our AI capabilities—from the build to the resulting business outcomes
When it comes to taking action, firms and users may differ in how much autonomy they want to give AI. In trade settlement fails, for example, InsightX identifies the fails, checks them against encoded clearing rules, and recommends next steps triaged by root cause and priority. An operator can review every suggested action before signing off, or choose to automate resolutions once the tool has earned its trust through a consistent track record.
Questions to Ask When You Consider AI Solutions
If you want to accelerate your data and AI strategy to achieve true business value, the first consideration is whether your AI solutions are built-in or bolted-on. Here are a few helpful thought-starters for what to ask potential vendors.
- Is this AI a connector to a general-purpose model, or built on my governed data directly?
- Where does my data go once submitted—is it ever used to train the underlying model?
- Does this AI know my regulatory environment, or is it just "AI for finance" in broad terms?
- Does the security model distinguish between identity access and content access?
- Does this add new friction to my user experience, or fit seamlessly into workflows?
- Where and how is human-in-the-loop oversight integrated?
Next up
Even built-in AI is only as good as what it knows. In Chapter 2 of the Real ROI series, we'll look at why even the most advanced AI can go wrong—not because the models aren't smart enough, but because they're missing Domain Intelligence. We’ll dig into the layers of context and the human expertise that form the proprietary Domain Intelligence inside DataXChange.
