How to Build an AI Analyst Agent for Business Teams

Answering "what was revenue last month" is a question; answering "why did revenue fall in the Northeast and what should we look at" is an analysis. The second requires a plan: compare periods, decompose by segment, check for data anomalies, chart the movers, and write it up with every figure sourced. An AI analyst agent does that on governed data, with the review discipline a human analyst's work gets. This playbook builds it on top of how to build a text-to-SQL agent and the governance in the data agents and text-to-SQL whitepaper; the reporting context is in AI analytics dashboards.

Step 1: What is the agent's role?

Element	Definition
Purpose	Perform routine analyses on governed data so analysts spend time on judgment and executives get sourced narratives faster
Scope	Standard analysis plans over the certified semantic layer: period comparisons, variance decomposition, cohort and segment breakdowns, anomaly checks, chart and narrative generation
Non-scope	Causal claims; uncertified data; forecasts presented as facts; distribution without review
Owner	Head of analytics

Step 2: What are the standard analysis plans?

Plan	Steps	Output
Period comparison	Metric for two periods; delta; drivers by dimension	Table, chart, narrative
Variance decomposition	Change attributed across a chosen dimension hierarchy	Waterfall chart, narrative
Cohort breakdown	Metric by cohort over time	Cohort table, chart
Anomaly check	Metric against expected range from history	Flags with context
Segment health	Several metrics by segment against targets	Scorecard

Each plan is specified with its steps, the semantic objects it uses, its checks, and its narrative structure, and each has its own golden cases.

Step 3: How does the planner work?

The agent interprets the business question, selects a plan or composes a small sequence of plans, lists the queries it will run and the definitions it will use, and, for non-routine questions, presents the plan for the analyst's confirmation before running. Every query goes through the text-to-SQL agent's validation and executes under the user's permissions.

Step 4: How are data quality checks built in?

Before writing a narrative, the agent checks for late-arriving data, unusual nulls, duplicate loads, and definition changes in the semantic layer, and states any caveat in the narrative. An analysis that ignores a partial load is worse than no analysis.

Step 5: How is the narrative generated?

In a fixed structure: what was asked, what the data shows (every figure linked to its query), what moved (decomposition), caveats (freshness, data issues, coverage), hypotheses (clearly labeled as proposals for investigation), and suggested next analyses. Description and explanation are separated by structure, not by tone. Charts carry the same source links.

Step 6: What review gates apply?

Routine plans on stable metrics can be reviewed by sampling; non-routine analyses and anything distributed to executives require an analyst's approval before release. The review interface shows the narrative beside the queries and figures. Corrections feed the evaluation set. Oversight design follows human-in-the-loop AI explained.

Step 7: What is the evaluation set?

Real analyses with analyst-verified plans, figures, charts, and narratives, plus questions the data cannot support. Score plan selection, query correctness, figure accuracy, chart correctness, narrative faithfulness, separation of description from hypothesis, and refusal correctness. Gate changes on the set; sample distributed narratives weekly.

Step 8: Rollout

1. Period comparison for the analytics team only.
2. Variance decomposition and anomaly checks.
3. Narratives with analyst review for one executive audience.
4. Cohort and segment plans.
5. Self-service for business users on routine plans, still sampled.

What does the agent look like in daily operation?

A regional director asks why Northeast revenue fell last month. The agent selects period comparison plus variance decomposition, lists the queries and the certified definitions, runs them under the director's role, and finds the decline concentrated in two product lines and one large account, with a note that the last two days' data loaded late. The narrative shows the figures with links, charts the decomposition, flags the partial load, and proposes two hypotheses, a price change and a lost renewal, as things to check. An analyst reviews, confirms the figures, adds a sentence about a known renewal, and releases it. The director has a sourced brief in an hour instead of a week, and the analyst spent ten minutes on judgment rather than a day on queries.

How does the agent handle forecasts and targets?

Business questions often drift from what happened to what will happen. The analyst agent may present projections only when a certified forecasting method exists in the semantic layer, and it labels every projected figure as a projection with its method and confidence interval, never as a fact. Comparisons against targets use the certified target table with its version date. Where no certified method exists, the agent states that a forecast was requested and none is available, and offers the descriptive trend instead. This boundary keeps the agent inside what the data can support and keeps forecasting a governed capability rather than a model's improvisation; the forecasting build is covered separately in how to build a demand forecasting system.

How does this fit the wider platform?

The analyst agent is a consumer of the semantic layer, the text-to-SQL capability, the gateway, and the evaluation platform, and it adds only the planner, the narrative generator, and the review gate. Building those on the shared foundation is what makes the second audience's analyses cheap; the foundation is described in the data agents and text-to-SQL whitepaper.

What are the common mistakes?

1. Causal claims from observational data.
2. Unsourced figures in narratives.
3. No data-quality checks before writing.
4. Distribution without review.
5. Building the analyst before the text-to-SQL foundation.

How does FISTA Solutions help?

FISTA Solutions builds analyst agents on governed semantic layers as part of its AI agents practice, with forward deployed engineers working inside analytics teams to specify plans and build evaluation sets, on the platform its AI enablement practice establishes. FISTA has delivered 150+ projects for 50+ companies across 12+ countries.

To build an analyst agent for one audience, message FISTA on WhatsApp, or read why AI agents need a semantic layer for the foundation it depends on.