The best division of labor is not handing an entire project to AI. Let an Agent accelerate schema exploration, initial SQL, repeated queries, and Report drafts. Analysts retain problem definition, business metrics, validation design, anomaly interpretation, and decision communication. The first group saves operating time; the second determines whether results deserve trust.
Divide work by verifiability, not job title
Agents fit tasks with explicit inputs, observable outputs, and rerunnable checks. Humans should lead work requiring context, accountability, or cross-functional negotiation.
| Work | Agent can start | Analyst must own |
|---|---|---|
| Data understanding | List schemas, types, NULLs, candidate keys | Determine business meaning and gaps |
| SQL | Draft JOINs, aggregations, repeated steps | Review grain, denominator, windows, fan-out |
| Validation | Run row, distinct, and outlier checks | Decide acceptable differences and added tests |
| Reporting | Draft findings, tables, and prose | Explain causes, separate correlation from causation, recommend |
| Communication | Rewrite for audience templates | Answer challenges and own commitments and tradeoffs |
"AI wrote it faster" is therefore incomplete. Saving 20 minutes on drafting but spending two hours investigating opaque numbers makes the workflow worse.
Quick path: use the Agent as an exploration accelerator
Quick tasks include listing columns, finding status values, producing initial distributions, validating SQL, and sizing whether a question deserves deeper work. Give a short acceptance condition:
List distinct order_status values and counts first.
Flag NULLs, unknown values, and the latest data date; do not draw business conclusions yet.
Quick results can stay in chat and SQL tabs. Upgrade when denominators, cross-table JOINs, formal comparisons, or external numbers appear.
Project path: let the Agent execute inside a contract
Formal analysis has five handoffs:
- Analyst defines the question: decision, population, grain, success criteria.
- Agent explores and drafts a Plan: sources, query steps, checkpoints.
- Analyst reviews the Plan: definitions, exclusions, risks.
- Agent executes: approved SQL, evidence, initial Report.
- Analyst validates and interprets: samples, alternative explanations, limitations, stakeholder communication.
OpenAI's 2026 Codex research indicates Agent use is moving toward longer tasks across tools. Its labor-duration thresholds are model estimates, and its sample and OpenAI setting have explicit limits, so treat it as a signal about work patterns rather than a universal productivity guarantee. See How agents are transforming work.
Long tasks also challenge context management, not just model intelligence. Anthropic treats context as finite and recommends just-in-time retrieval, structured notes, and compaction. See Effective context engineering for AI agents. For analysts, Plans, SQL, checkpoints, and Reports are more stable external state than an ever-growing chat.
Lantide Data's collaboration model
Lantide Data offers two paths: Quick Analysis for exploration and one-off validation, and Project Analysis for Plan → comments → user Approve & Execute → SQL evidence → Report. See the analyst Plan → Execute → Report guide.
The point is not to remove analysts but to leave reviewable artifacts from Agent work: SQL is retrieval evidence, Plan is the pre-execution contract, and Report explains conclusions and limitations. Highly programmatic models, non-tabular data, or notebook-heavy exploration may still suit Jupyter and specialized tools better. Lantide's focus is SQL-first tabular analysis requiring governance and delivery.
Evaluate collaboration with four metrics, not time alone
- definition corrections before Execute;
- issues found by SQL/data-quality checkpoints;
- time to rerun or hand off the analysis;
- Report content accepted directly versus returned for overclaiming.
These show whether the Agent reduces rework or merely packages unfinished work as complete.
Conclusion
Choose one low-risk but real retrospective. Let the Agent explore, query, and draft; have the analyst validate at Plan, checkpoint, and Report. The ideal division is not less human work, but concentrating human time where context, responsibility, and professional judgment are indispensable.