The Trust Gap: Why Agentic AI Has Failed Data Workers

While software engineers have enjoyed a transformative leap in productivity thanks to agentic AI, the broader population of business and financial analysts remains largely sidelined. Despite these tools being technically capable, they have failed to move the needle for the non-technical workforce. As Raj Bains, CEO of Prophecy, points out, the difference lies in the ability to verify output—a luxury engineers have long mastered but business users are still fighting for.

The Productivity Paradox

In the modern enterprise, software developers are outnumbered by business and financial operations professionals by a ratio of roughly six to one. Yet, while developers have embraced AI agents to generate, debug, and test code, the average business analyst has seen only marginal gains.

The reason is structural. When an engineer uses an AI agent, the output is code—a language they speak fluently. They can review the "diff," run unit tests, and perform a line-by-line validation. Trust is earned through verification. Conversely, when a business user asks an AI to analyze revenue trends or risk exposure, the agent returns a final figure or a polished chart. The "how" remains a black box. For a financial analyst, a number without a clear, auditable provenance is not a productivity gain; it is a liability.

The Three Barriers to Adoption

The failure of agentic AI to permeate the business sector is not a result of poor model performance, but rather three specific, solvable systemic issues.

1. Data Inaccessibility and "Shadow" Truths

Business data is currently trapped in a fragmented landscape: spreadsheets on local drives, SharePoint files, Salesforce instances, and cloud-native platforms like Databricks or Snowflake. When a business user needs data, they are often forced to file a ticket with an overburdened data platform team.

Because these teams are chronically understaffed, the "workaround" becomes the standard. Analysts often resort to using stale CSVs or "shadow" versions of datasets—copies that have been handed off and modified so many times that their lineage is impossible to trace. When an AI agent is pointed at this compromised data, it produces results based on a "shadow version of the truth." In highly regulated environments like banking or insurance, this lack of data integrity is not just an inefficiency—it is a potential career-ending event.

2. The Absence of a Data "Fact-Check"

In the engineering world, code is the artifact. In the business world, the artifact is the decision-making process. Current AI tools for business users often rely on "text-to-SQL" interfaces that assume data is clean, pre-processed, and ready for query.

However, real-world data is messy. It arrives in inconsistent vendor formats, is stored in disparate architectures, and requires significant preparation before it can be queried. When AI tools skip the preparation phase and move straight to output, they provide no context. A business user cannot simply ask an AI to "fact-check" its own logic if they cannot see the steps the AI took to arrive at a conclusion. The pressure to deliver insights quickly often forces users to accept these black-box results, creating a dangerous dependency that falls apart under the scrutiny of an auditor.

3. The "Prep" Gap

Most current enterprise AI tools are designed for the "last mile"—the query. They assume the data is already in a pristine state, neatly organized in a relational database. This ignores the reality of the business analyst’s workflow, which begins with the arduous task of gathering and cleaning disparate files. By neglecting the data preparation layer, these tools demo beautifully but fail in practice, leading to the low adoption rates we see today.

A Chronology of the Trust Deficit

To understand how we reached this impasse, we must look at the evolution of "self-service" data tools over the last decade.

2015–2018 (The Desktop Era): Companies flooded the market with desktop-based BI tools. These allowed business users to build their own workflows, but they created massive data silos. Every department developed its own version of "revenue," leading to a decade of technical debt and inconsistent reporting.
2020–2024 (The Cloud Migration): Organizations moved to centralized cloud data platforms (Snowflake, Databricks). While this solved the storage issue, it made data harder to access for the average user, as it required sophisticated SQL knowledge and strict governance protocols.
2025–2026 (The Agentic Hype): The arrival of LLM-powered agents promised to bridge the gap between business users and cloud data. However, the focus remained on natural language queries rather than the transparency of the data supply chain.
Late 2026 (The Current Turning Point): Industry leaders are now shifting the focus from "generative speed" to "verifiable intelligence." The consensus is moving toward platforms that prioritize data lineage and auditability over simple conversational outputs.

Implications for the Enterprise

The failure to address these issues has profound implications for how corporations will function in the coming years.

The Rise of "Inspectable AI"

The solution to the black-box problem is not to stop using AI, but to mandate that agents "show their work." For business users, this means moving toward visual workflows. An agent should not just return a number; it should generate a transparent map of how it read the data, what transformations it applied, and what business logic it followed. If an agent misinterprets a business rule, the user should be able to intervene, edit the specific step, and re-run the process. This "human-in-the-loop" approach turns the AI from an opaque oracle into a transparent co-pilot.

The Governance Mandate

Governance and speed are often viewed as opposing forces, but they don’t have to be. By running AI agents within a governed platform—where data access is controlled and processes are monitored—organizations can actually increase the speed of business. When the "governed path" is the fastest path, "shadow spreadsheets" naturally disappear.

The Productivity Opportunity

The "analytics backlog" that plagues most large enterprises is a direct result of the inefficiency of the current data-to-insight pipeline. By connecting business users to real, governed data via inspectable AI, companies can effectively democratize data science. This is not about building more model benchmarks; it is about building better plumbing.

Looking Forward: Beyond the Hype

The next wave of enterprise AI will not be defined by the model’s ability to summarize text or write code. It will be defined by its ability to integrate into the existing, messy reality of enterprise data.

For the six-to-one majority of business professionals, the promise of AI has yet to be realized. But by focusing on the three pillars of accessibility, fact-checking, and thorough data preparation, the industry is poised to move past the black-box era. We are entering an era where AI doesn’t just do the work—it explains the work. For the data-driven enterprise, that distinction is the difference between a tool that is used and a tool that is trusted.

Ultimately, the goal is to prevent the mistakes of the previous self-service wave. We don’t want another decade of technical debt built on top of "AI-generated" silos. By prioritizing transparency and governance today, organizations can ensure that the next phase of agentic AI drives meaningful, reliable, and scalable business value. The path forward is clear: connect users to real data, show the work, and keep the human firmly in the loop.