Image by Author 

The promise of AI agents was felt everywhere in 2025. Consultants pitched them as digital coworkers. Vendors rebranded chatbots as agents. LinkedIn overflowed with takes on how autonomous systems would transform data work overnight.

But away from the noise, something more interesting happened. In narrow, well-defined workflows, agentic systems started delivering measurable value. Not by replacing analysts but by handling the repetitive, multi-step tasks that consumed their time.

This article examines three concrete workflows where agentic approaches moved from hype to reality in 2025. For each, we will contrast three approaches: pure manual work, single-prompt assistance from a language model, and multi-agent orchestration. The differences reveal where the technology genuinely helped and where it remained more promise than product.

IBM Research documented this disconnect between expectations and outcomes. Their experts noted that while nearly all developers are exploring agentic systems, the gap between experimentation and production deployment remains substantial. 

Understanding the Three Approaches

Before diving into specific workflows, let’s define our terms. The distinctions matter because vendors often blur them.

Manual Approach: A human performs each step sequentially. They write queries, inspect results, make decisions, document findings, and move to the next task. Every judgment call requires active attention. This approach offers complete control but scales poorly.

Single-Prompt LLM: A data professional asks a language model for help with one specific task. They might request code to clean a dataset or ask for interpretation of statistical output. The LLM responds once, the human reviews it, and they move on. This speeds individual tasks but still requires orchestration by the human.

Agentic System: Multiple specialized agents collaborate on a workflow. One agent queries data, another validates results, a third generates documentation. A coordinator oversees the process, passing context between agents and handling errors. The human sets objectives and reviews outputs rather than managing each step.

The key difference lies in orchestration. Single-prompt interactions remain stateless (each query starts fresh). Agentic systems maintain memory across steps, learn from intermediate results, and adapt their approach.

Workflow 1: Data Cleaning

Data scientists report spending 50 to 80 percent of their time preparing data rather than building models. This reality persisted through 2025, but the nature of that preparation began shifting for teams using agentic approaches.

The Manual Reality

Consider a retail analyst cleaning customer transaction data with 200,000 rows. The dataset has issues scattered throughout: missing purchase dates, inconsistent product categories, duplicate customer IDs, and outliers in transaction amounts.

The analyst scans column summaries, identifies missing values, decides on an imputation strategy, writes code to handle nulls, spots duplicates, writes deduplication logic, validates results, discovers new issues, and repeats. A typical cleaning session stretches across three to four hours.

Each decision requires judgment. Should missing purchase dates use the median date or be dropped? Are those high-value transactions legitimate or errors? The analyst context-switches between analysis, coding, validation, and documentation.

Single-Prompt Assistance

With a language model, the analyst describes their data issues and receives cleaning code suggestions. They might ask: “I have missing values in my purchase date column and some transaction amounts over $10,000 that seem suspicious. How should I handle these?”

The model suggests approaches: forward-fill dates based on customer history, cap extreme values at the 99th percentile, or flag suspicious transactions for review. The analyst reviews the suggestion, adapts the code, runs it, and evaluates results.

This cuts coding time but preserves the same sequential structure. The analyst still orchestrates every step and makes all judgment calls.

Agentic Approach

An agentic data cleaning system operates differently. Multiple specialized agents collaborate under a coordinator:

A profiling agent scans the dataset structure, identifies column types, and flags anomalies. A validation agent checks data against business rules and highlights constraint violations. A cleaning agent applies context-appropriate transformations based on patterns it learned from previous datasets. A documentation agent tracks every transformation and maintains audit trails.

The coordinator orchestrates these agents. When the profiling agent identifies missing purchase dates, it triggers the validation agent to check if those records have other anomalies. If they do, the cleaning agent might recommend removal. If they represent a systematic issue, it flags the pattern for human review.

The human analyst reviews proposed changes, approves the overall strategy, and examines edge cases the system couldn’t resolve confidently.

Real-World Impact

PureML, developed during a 2025 hackathon, demonstrated this approach with automotive data. Rather than simply filling missing values with averages, the system used retrieval-augmented generation to find contextually appropriate values. When manufacturer names appeared inconsistently, it consolidated synonyms without manual intervention. The team reported cleaning a 50,000-row dataset in under 10 minutes compared to the typical 2-3 hour manual process.

But the results remained mixed. Agentic cleaning worked best with structured data following known patterns. Novel data quality issues still required human expertise. The systems excelled at automating repetitive checks but struggled with domain-specific judgment calls that lacked clear precedents.

Enterprise implementations revealed similar patterns. Data teams found value in automating validation rules and standardizing cleaning procedures across projects, but edge cases consistently required human review.

Workflow 2: Feature Engineering

Feature engineering transforms raw data into predictive signals. The process requires domain knowledge, statistical intuition, and iterative experimentation. In 2025, agentic systems began automating parts of this workflow, but with notable limitations.

Manual Feature Engineering

A data scientist building a customer churn model starts with raw transaction data. They aggregate transaction counts by time period, calculate recency and frequency metrics, create ratio features comparing recent to historical activity, engineer time-based features, test interaction terms, and evaluate feature importance through iterative modeling.

This requires creativity and domain knowledge. Which time windows matter most? Should they capture spending velocity or absolute amounts? The scientist tests dozens of feature combinations, many proving uninformative. The cycle can extend across days or weeks.

Single-Prompt Assistance

With language model help, the data scientist describes their prediction task and raw data structure. The model suggests relevant feature categories: aggregations, ratios, time-based patterns, interaction terms.

The scientist implements promising suggestions, tests them, and requests refinements. This accelerates implementation but maintains the same iterative structure. The scientist still drives hypothesis generation, testing, and selection.

Agentic Approach

An agentic feature engineering system brings multiple specialized agents to the task:

A pattern detection agent analyzes historical data for recurring patterns. A feature generation agent creates candidate features based on detected patterns and domain templates. An evaluation agent tests feature utility using multiple criteria: predictive power, correlation with existing features, computational cost. A synthesis agent combines promising features into higher-order interactions.

The coordinator manages this exploration, balancing breadth and depth. Rather than testing every possible combination, it uses reinforcement learning to focus on promising directions.

The data scientist sets constraints: prediction targets, computational budgets, interpretability requirements. The system explores the feature space and reports promising candidates.

Practical Results

Enterprise teams reported mixed outcomes. Agentic systems handled standard feature engineering patterns well: time-based aggregations, ratio calculations, categorical encodings. They generated features faster than manual approaches.

However, they struggled with domain-specific innovation. An agentic system might create hundreds of mathematically valid features without understanding which ones capture business-relevant patterns. In specialized domains like healthcare or manufacturing, the systems provided less value than in general business contexts.

The most successful implementations combined agentic exploration with human curation, reducing feature engineering time while maintaining quality.

Workflow 3: Experiment Analysis

Running experiments, particularly A/B tests, involves structured analysis workflows. Multiple companies in 2025 began automating these workflows with agentic systems, achieving clearer results than in less structured domains.

Manual Experiment Analysis

A product analyst runs an A/B test comparing two checkout flows. They pull experiment data, join it with user metadata and conversion events, calculate sample sizes, check for sample ratio mismatches, perform statistical tests, evaluate multiple metrics including guardrails, investigate interaction effects, generate visualizations, and document findings.

This workflow requires statistical expertise and takes several hours per experiment. Large organizations run hundreds of experiments simultaneously, creating analysis backlogs.

Single-Prompt Assistance

The analyst describes their experiment structure to a language model: sample sizes, metrics, and test duration. The model suggests appropriate statistical tests, provides code for calculating p-values and confidence intervals, and helps interpret results.

This accelerates specific analysis steps but leaves orchestration to the analyst.

Agentic Approach

An agentic experiment analysis system coordinates multiple specialized agents:

A data ingestion agent retrieves raw data and validates data quality. An analysis agent runs appropriate statistical tests, checking assumptions and handling edge cases. A quality assurance agent looks for common pitfalls: sample ratio mismatches, novelty effects, interaction effects. A reporting agent synthesizes findings and generates visualizations.

The coordinator ensures completeness. If the analysis agent detects potential issues, it triggers deeper investigation. If results seem ambiguous, it runs additional statistical approaches to triangulate conclusions.

One data scientist at a major tech company built exactly this system using Cursor and the Model Context Protocol. After initial setup requiring example analyses and careful prompting, the system handled end-to-end experiment analysis. The implementation reduced per-experiment analysis time though setup and validation still required significant upfront investment.

Measured Impact

This workflow showed some of the clearest value. Structured experiments follow predictable analysis patterns, making them well-suited for automation. Companies using agentic analysis completed experiments faster while maintaining statistical rigor.

The systems excelled at standardization, catching common mistakes like insufficient power calculations or multiple comparison problems. They ensured every experiment received consistent analysis, reducing analyst-to-analyst variability.

Limitations emerged with complex experiments: interactions between multiple treatments, heterogeneous treatment effects, or unusual metric distributions. These cases still required expert review.

Comparing the Three Workflows

Here’s how agentic systems performed across these workflows:

Workflow	Best Use Cases	Key Limitations	Adoption Readiness
Data Cleaning	Structured data with known patterns, standardized checks, audit trail requirements	Novel data quality issues, domain-specific rules, edge cases requiring judgment	Medium (needs validation guardrails)
Feature Engineering	Standard transformations, exploring feature combinations at scale, testing variations	Domain-specific innovation, interpreting business relevance, specialized industries	Medium (requires expert curation)
Experiment Analysis	A/B tests, standard statistical procedures, repetitive analysis tasks	Complex interactions, heterogeneous effects, unusual distributions	High (clear structure and validation criteria)

Common Patterns Across Workflows

Examining these three workflows reveals consistent themes about where agentic systems delivered value and where they fell short in 2025.

Where Agentic Systems Helped

Agentic approaches proved most valuable for structured, multi-step workflows with clear success criteria. Data cleaning followed predictable patterns. Experiment analysis had defined statistical procedures. In these contexts, agents could orchestrate sub-tasks without constant human direction.

The systems excelled at consistency. They applied the same thorough approach to every task, catching issues that humans might miss through fatigue or oversight.

Documentation improved substantially. Agentic systems naturally logged every step, decision, and transformation. This created audit trails that manual processes often lacked.

Where They Struggled

Novel situations exposed limitations. When agentic systems encountered data patterns outside their training distribution, they either failed obviously or, worse, confidently produced incorrect results. Human oversight remained essential.

Domain expertise proved difficult to encode. While systems could apply general analytical techniques, they lacked the nuanced understanding that experienced analysts brought to specific industries or business contexts.

Interpretation and judgment remained human tasks. Agentic systems could identify that sales dropped in the Northeast region but couldn’t assess whether this represented a serious strategic concern or a temporary fluctuation requiring no action.

The Governance Challenge

Organizations implementing agentic systems discovered that technology deployment was the easy part. The harder challenges involved governance, monitoring, and accountability.

When an agentic data cleaning system removed 10,000 records as duplicates, who verified that decision was correct? When experiment analysis suggested launching a feature, how did teams trace the reasoning?

Successful deployments invested heavily in monitoring, validation, and rollback capabilities. They treated agentic systems more like production software requiring testing and safeguards than like analytical assistants.

These challenges showed up in industry forecasts. Gartner predicted that over 40% of agentic AI projects would be canceled by the end of 2027, citing governance complexity and unclear ROI as primary factors. Organizations that succeeded treated implementation as an organizational challenge, not just a technical one.

Looking Forward

The practical experience of 2025 revealed that agentic analytics represents evolution rather than revolution. The technology genuinely helps with specific, well-defined workflows but doesn’t eliminate the need for skilled analysts.

Strategic frameworks from consulting firms emphasized this measured approach. McKinsey’s research found that organizations achieving value from agentic systems shared three characteristics: they started with tightly scoped workflows, maintained strong validation processes, and invested in upskilling analysts to work alongside agents rather than being replaced by them.

The most effective applications combined agentic automation with human expertise. Systems handled orchestration, routine checking, and documentation. Humans provided domain knowledge, strategic judgment, and validation.

As organizations move into 2026, the clearest opportunity lies in identifying workflows that match agentic system strengths: structured processes, clear success criteria, repetitive orchestration. Teams that carefully select these use cases while maintaining appropriate human oversight will realize sustained value.

The hype suggested agentic systems would transform data work immediately and completely. The reality proved more measured. They became useful tools for specific tasks rather than autonomous replacements for analysts. That more modest outcome still represents meaningful progress.

For data professionals, the shift means adapting rather than being replaced. Understanding how to direct and review agentic systems becomes as important as traditional analytical skills. The role evolves from executing every analytical step to orchestrating human-agent collaboration, focusing human effort where judgment and creativity matter most.

The agents are here. They’re helping. But they’re not doing it alone.