Why Data Teams Lag Behind Developers in Agentic AI Adoption
While software engineers quickly benefited from agentic AI, data workers face major hurdles with unverified outputs, siloed data, and a lack of inspectable workflows.
Software engineers were among the earliest professionals to realize tangible productivity leaps from agentic AI. The primary driver was testability: when a model writes code, an engineer can review the diff, execute automated test suites, and read the logic before pushing it into production. The output matches an established format engineers already understand, reducing the need for blind trust.
For data analysts and business operations teams, the situation is completely different. In the United States, workers in business and financial operations outnumber software developers by roughly six to one. Yet according to Raj Bains, CEO and co-founder of AI data prep platform Prophecy, these professionals have captured only a small fraction of the productivity gains experienced across software engineering. When an analyst asks an agent to conduct an analysis, the output typically arrives as an isolated chart, a final number, or several hundred lines of raw code. None of these formats make it clear whether the agent gathered the correct data points, omitted crucial information, or applied the proper formulas.
The bottlenecks slowing enterprise data agents
According to Bains, three practical barriers prevent data workers from trusting and deploying autonomous agents effectively:
- Data access and governance silos: Crucial business data remains distributed across spreadsheets, SharePoint folders, Salesforce instances, and centralized platforms such as Snowflake, Databricks, or BigQuery. Accessing this data requires requests to data platform teams that are frequently constrained by limited headcount. As a workaround, business users often end up working with stale CSV exports or spreadsheets that have been modified multiple times without tracking. When an AI agent processes these unmanaged shadow copies, data lineage is absent, making it impossible to trace where metrics originated.
- A lack of fact-checking mechanisms: When an AI returns a result, analysts rarely receive the contextual breakdown required to confirm its validity. Providing raw generated code is not an effective solution, as business users are not expected to read it. When auditors or executives question the calculations behind a metric, relying on an unchecked answer from an agent is unacceptable.
- Tools that bypass data preparation: Many contemporary data tools focus on text-to-SQL capabilities, operating under the assumption that incoming data is already cleaned, formatted, and structured in uniform tables. In reality, daily data workflows involve handling multi-format vendor files and fragmented storage systems. Skipping the core preparation stage leads to tools that perform well in demonstrations but fail in practical usage.
Restoring trust with visible data workflows
To unlock agentic adoption for business teams, Bains argues that AI must clearly show its work through artifacts that analysts can comprehend. Rather than operating as an uninspectable black box, agents should output visual, step-by-step transformation workflows showing how data was ingested and altered at each stage.
This visibility allows domain experts to trace key metrics, inspect date range adjustments, and adjust errors where the agent misinterpreted specific business logic. Implementing these automated workflows inside a governed environment—complete with monitoring and standard access permissions—protects data lineage and reduces reliance on uncontrolled shadow spreadsheets without slowing down analysis queues.
What it means for developers
For developers designing agentic applications and internal data infrastructure, these dynamics highlight a critical UX and architectural shift. Building effective tools for non-technical or semi-technical domain experts requires moving beyond simplistic text-to-SQL prompts and raw code generation.
Developers must prioritize auditability and lineage tracking in their agent architectures. Systems should generate structured, step-by-step visual pipelines that non-programmers can audit and adjust directly. As teams test and implement various language models for agentic reasoning and data transformation pipelines, developers can evaluate top AI models cheaply through one API at https://apixoai.online.
Solving the analytics bottleneck will not depend on new model benchmarks alone. Software engineers already have tools tailored to their verification habits. Extending those gains to the broader operational workforce requires building data platforms where agents make every transformation step verifiable, governed, and easy to inspect.
Source: Why data work is one step behind on AI — InfoWorld AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
