Why Enterprises Stumble on AI for BI

By Marius Moscovici
Founder & CEO at Metric Insights

On the face of it, AI for BI shouldn't be that hard. AI models have gotten dramatically more capable over the past year. Most of us now use AI every day to improve emails, draft documents, accelerate office work, and ask questions about publicly available data. So, deploying AI to help people understand their own business data should be a walk in the park. Right?

Instead, for those who have tried, the journey feels more like climbing the Huayna Picchu "Stairs of Death" in Peru. The steps are narrow, uneven, and slippery from the cloud forest mist. On one side is a cold stone wall. On the other, a straight thousand-foot drop into the Urubamba River valley. You know the destination is worth reaching, but every step demands your full attention, and the consequences of a misstep are severe.

That metaphor isn't dramatic hyperbole. Very few large organizations have successfully deployed AI for BI at scale, despite AI's tremendous potential to improve how people understand and engage with their data. Gartner predicted that 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and the early data suggests the reality may be worse. I've spent the past year talking to dozens of enterprise BI leaders about the challenges they've encountered, and a handful of themes repeat in nearly every conversation. The stumbles happen at predictable points in the journey, and understanding where and why they occur is the first step toward avoiding them.

Here's what I keep hearing.

This is, almost universally, the first obstacle. Simply picking an AI-for-BI technology stack has become a daunting task, not because there aren't enough options. If anything, there are too many.

AI is a fast-moving space. The landscape of tools, models, and approaches shifts every quarter. Every BI tool vendor now claims AI capabilities. Cloud database providers are launching their own AI agents. A wave of AI-native BI startups has entered the market. The noise is deafening, and it's extraordinarily difficult for a BI leader to separate genuine capability from marketing.

Compounding the noise is a lack of clarity around use cases. Most BI teams I talk to know they're supposed to "do something with AI," but they don't know where to start. Should they start with natural language querying? Automated anomaly detection? AI-generated summaries of dashboards? Without a sharp answer to that question, teams either try to do everything at once or pick the shiniest demo and hope it sticks.

And then there's the expectations gap. Users have spent the past couple of years interacting with ChatGPT, Claude, and other consumer AI tools. They expect the same polished, conversational experience when they ask questions about company data. Most enterprise AI-for-BI tools fall far short of that bar, and the gap between what users expect and what the tools deliver creates early disillusionment that's hard to recover from.

Many BI leaders I speak with have already attempted at least one AI-for-BI initiative. They've turned on Copilot in Power BI, or Pulse in Tableau, or experimented with a database-layer AI agent. And the consistent refrain is: underwhelming.

One head of analytics at a large retailer told me his team enabled the AI capabilities in their existing BI platform, ran a handful of test queries, and quickly realized the tool couldn't handle anything beyond the most basic "what were our sales last month" questions. Those are the questions users can already answer themselves with a well-designed dashboard. The value isn't there.

A dashboard's semantic model might contain table definitions and basic joins, but it doesn't know how your organization defines gross margin, why Q3 numbers need to be adjusted for a product recall, or what abbreviation the sales team uses for a key customer segment. Without that context, the AI gives plausible-sounding answers that are subtly wrong, which is worse than giving no answer at all.

Then there's the "Why" problem. A user asks, "How is gross margin trending?" and the tool answers correctly: it's down 10% from last year. Fine, but a dashboard could have shown them that. The user follows up: "What's driving the change?" Now the tool struggles, maybe surfacing that cost of goods increased by 20%. Then comes the question that actually matters: "Why did that happen?" That's where the user needs to learn that tariff costs spiked 40% on a specific product line, and that's where almost every tool I've seen falls apart. I've watched countless demos that handle the first question confidently and collapse the moment a user pushes deeper. But those deeper questions are exactly where AI should deliver value that dashboards can't.

The final piece of this stumble is verification. Even when an AI tool produces a plausible answer, there's usually no way for a business user to validate it. Some tools will show you the SQL query they generated, but that's not helpful to the VP of Sales, who doesn't know the database schema and hasn't written a line of SQL in her life. The only way most business users can build trust in an AI-generated answer is by comparing it to what they already see in their certified reports and dashboards, the numbers they know are right. If the AI experience doesn't make that cross-reference easy, trust never develops. And without trust, adoption doesn't happen.

This is perhaps the most frustrating stumble, because the team has already invested significant effort and has something that works. Sort of. They've built a proof-of-concept that handles a narrow set of questions within a specific data domain. The demo looks great. Leadership is excited. And then reality sets in, and the project stalls. If that sounds familiar, you're in good company. According to S&P Global, the average organization scraps nearly half of its AI proof-of-concepts before they ever reach production.

The first problem is that the experience is disjointed. The AI capabilities live in their own interfaces: Data Agents in Microsoft Fabric, Genie in Databricks, and Cortex in Snowflake. But the business users who are supposed to benefit from these tools don't work in those environments. They work in Tableau dashboards, Power BI reports, and Excel spreadsheets. The AI agents are available, but they're in the wrong places and disconnected from the tools users actually use throughout the day.

This fragmentation makes root cause analysis, the thing AI should be best at, nearly impossible. Real investigations span data domains and tools. If the AI can answer questions about sales data but not supply chain data, and the user has to leave the AI interface to check a separate dashboard for inventory numbers, the experience falls apart. The user ends up doing the same manual assembly work they were doing before, just with an extra tool in the mix.

The second problem is the gap between a narrow prototype and a full-scale solution. A proof-of-concept that answers questions about one data domain with 30 tables is very different from a production system that covers the dozen domains an enterprise actually needs. Questions quickly extend beyond the boundaries of whatever limited semantic model was built for the prototype. Users don't think in terms of "which data domain am I querying." They just ask questions. When the AI can't handle something that feels like it should be in scope, confidence erodes fast.

Underneath all of these specific stumbles is a more fundamental tension: the gap between what users expect an AI-for-BI experience to look like and what current tools actually deliver.

Users expect an experience that feels like having a knowledgeable analyst on call. But that's not what they get.

Most AI-for-BI tools today are essentially chatbots. They sit and wait for someone to type a question. But that's not how a good analyst works. A good analyst notices that revenue in the Northeast region dropped 15% last week and pings you about it before you even think to ask. A good analyst doesn't stop at the boundary of a single dashboard; they pull together whatever information is needed to reach the answer. Business users expect that same proactive behavior from AI: reach out to me when something important changes, flag the anomaly, tell me what's worth paying attention to today. Instead, what they get is a text box that answers one narrow question at a time within the scope of a single data model. That gap between the proactive, analyst-like experience users expect and the passive chatbot experience they get is the deepest source of disappointment.

And there's a compounding problem: because most organizations have multiple BI tools, databases, and AI agents layered on top, the user has to figure out which tool to use, where the data lives, and how to piece together a complete picture. They were already frustrated by having too many dashboards. Adding a set of disconnected AI agents doesn't solve the problem. It makes it worse.

None of these obstacles is insurmountable. But they do require a more deliberate approach than simply turning on AI features in your existing BI stack and hoping for the best.

The organizations I've seen make real progress share a few common traits: they start with a specific, well-defined user and a specific set of questions that AI can answer better than existing tools. They invest in the business context and the semantic layer before they invest in the AI interface. They treat security and governance as foundational requirements, not Phase 2 add-ons. And they design the AI experience to work alongside their existing dashboards and reports, not replace them.

Watch our webinar on enterprise-ready AI for BI here: Enterprise-ready AI for BI.

These are the challenges that have shaped what we've been building at Metric Insights. In this webinar, I walk through how we're addressing each of these stumbling points—from bridging the gap between AI agents and the BI tools users actually work in, to making AI answers verifiable against trusted reports, to delivering the proactive, analyst-like experience that business users expect.