Enterprises turn to DIY AI tools to cut BI licensing costs, but the custom builds cost more, create security gaps, and break as user counts grow.
Humera Madni, BI Architect at Ganit, explains that AI delivers real value when layered on top of established BI platforms, not when it replaces them.
She advises teams to validate AI outputs, prioritize change management before deploying new tools, and treat mature BI infrastructure as the foundation AI builds on.
Artificial intelligence obliterated the barrier to entry for standard data tools. The result is a surge of organizations bypassing established BI platforms entirely, convinced that AI can replicate in hours what those tools took years to build. Chasing speed and cheaper licenses, organizations slap bespoke AI wrappers onto their data stacks. What those shortcuts skip is exactly what production demands: authentication, governance, and scalability.
Humera Madni is a BI Architect at Ganit, a full-stack data and AI consultancy that serves enterprise clients. A certified Microsoft Fabric Analytics Engineer with experience across financial, healthcare, and public sector data environments, she builds solutions spanning Tableau, Power BI, and Apache Superset. Having vetted custom AI builds against established platforms firsthand, she has seen what happens when the architecture underneath those decisions isn't ready.
"Clients think building their own AI apps will be cheaper than traditional BI tools. But in reality, they end up paying more, and the solutions break as soon as they try to scale," says Madni. It is a miscalculation she encounters repeatedly.
Across her client base, Madni watches the same instinct take hold: the conviction that building a custom AI application will be cheaper and faster than paying for an established platform.
Shiny object syndrome: "It's sometimes like children getting excited with new tools. Everyone thinks, 'Let's go to AI. I can just take away my licensing cost and be the cool guy in the team,' says Madni. "But it's not going to happen, because we're going to break it very soon." The excitement fades quickly. In Madni's experience, clients routinely end up paying more than they would have for the platforms they walked away from.
The cost miscalculation is only part of the shift. As AI handles foundational coding tasks, the measure of a strong BI developer is shifting. Data professionals are moving away from syntax and toward optimization, performance, and architecture. With tools like Claude connected through an MCP server handling the logic, developers now act as supervisors auditing outputs rather than hand-writing every measure.
DAX in a dash: Writing DAX formulas, the calculated measures that drive Power BI reports, is among the most time-intensive tasks in BI development. "It took around two days for her to write some of the DAX measures in Power BI, which are really complicated. And Claude could do it in around ten to fifteen minutes," notes Madni. "It's a breakthrough." The speed gains are real, but they come with a significant caveat. MCP server connections give AI direct access to the data, but the data alone does not tell the full story of how a business actually operates.
Fast but flawed: "You will not be able to set a context just by giving a prompt. Claude can misinterpret how the relationships are in the real processes," she explains. Madni's approach has been to define clear criteria for her team on when to apply AI assistance and when the complexity of the underlying business logic requires a human to work through it directly.
The gap between a fast-tracked proof of concept and a live environment forces teams to confront operational hurdles they ignored at the prototype stage: the security and governance requirements that custom builds routinely skip, and the human realities of the users those builds are supposed to serve. When Madni recently vetted a client's custom platform against Tableau, she found hundreds of issues. The tool was simply not designed as a scalable application, meaning it would inevitably buckle as users scaled from 10 to 100. The clients building it had never asked users what they actually needed. In some cases, they were asking to switch off the native AI features inside mature platforms like Copilot or Tableau Pulse, convinced a cheaper custom build would serve them better, without having done the work to find out.
Prompting isn't protection: The security gaps are often invisible until something goes wrong. Some teams treat public models as a black box, feeding in sensitive data without establishing proper authentication or role-based access. "Tools like Claude or ChatGPT do not really come with a disclaimer that they are going to protect the data that we are going to upload," says Madni. The exposure problem compounds over time. Custom applications also struggle with reusability. As business logic evolves, bespoke wrappers are slower to adapt than platforms built for iteration, which offer versioning and flexibility that are genuinely difficult to replicate from scratch. For true governance and security, Madni sees experienced practitioners gravitating toward enterprise-grade platforms like Microsoft Fabric, where data is secured at the most granular level and AI can be layered on top safely.
Spreadsheet safe space: Leaders pushing for broad AI adoption need to understand what their users actually need before deciding which tools will serve them. If a team builds a custom dashboard without understanding the stakeholder's actual pain points, the technology fails on arrival. "I have business users who are still on Excel sheets," she explains. In many organizations, AI's first job should be supporting change management and adoption, not replacing interfaces that still work.
When every developer can generate identical code from identical prompts, human insight becomes the only premium product left. Professionals who suppress their own creativity and thoughts in favor of unvalidated AI outputs don't just risk quality, they lose the ability to differentiate themselves from each other. Keeping a human in the loop is no longer just a governance principle. It is how practitioners prove their value.
"I see a lot of plagiarism there, and I see misquoting stuff. We should really be honest that this output has been generated from AI," she notes. "That transparency should be there." The tools can jump-start work and compress timelines, but they cannot supply the judgment, context, or accountability that make outputs trustworthy. What separates a successful deployment from a broken prototype is how clearly practitioners communicate where the machine ends and the human begins.
