Pattern: Agent as Analyst
🔵 Optional
Use this when: You have data (research transcripts, backlog items, metrics, support tickets, survey responses) and need to find patterns, surface insights, or identify risks.
How it works
You provide data. The agent processes it and surfaces patterns. You validate with your domain knowledge.
You → upload or paste raw data
Agent → identifies patterns, themes, outliers, risks
You → validate: Does this match what I know? Is this real?
You → ask follow-up questions to go deeper
Agent → refines analysis
You → decide what to act on
This pattern is fundamentally different from "Agent as Drafter." Here, the agent isn't creating something new — it's finding structure in something that already exists. Your job is validation, not revision.
When to use it
- Synthesizing user interview transcripts
- Analyzing support ticket trends
- Reviewing backlog health (stale stories, missing criteria, duplicates)
- Velocity and sprint trend analysis
- Competitive landscape mapping
- Analyzing survey responses (qualitative and quantitative)
- Identifying risks in the current iteration plan
When NOT to use it
- Judgment calls — The agent can tell you what the data shows. It can't tell you what to prioritize, which stakeholder to listen to, or when to pivot.
- Small data sets — If you have 2 interviews, you don't need an analyst. Read them yourself.
- Data you haven't verified — Garbage in, garbage out. Don't ask the agent to analyze data you haven't sanity-checked.
How to get good analysis
1. State your questions explicitly
Bad: "Analyze these interview transcripts."
Good: "Analyze these 6 interview transcripts. I'm trying to understand: (1) What are the top pain points in the onboarding flow? (2) Where do users get stuck? (3) What do users wish they could do that they currently can't?"
2. Ask for evidence
Always require the agent to cite specific examples from the data.
For each theme you identify, include:
- At least 2 specific quotes from different participants
- Which participants expressed this theme
- How confident you are in this pattern (strong, moderate, emerging)
3. Challenge the findings
After the first analysis, push back:
Now play devil's advocate on your own analysis:
- Which of these themes might be artifacts of my interview questions rather than genuine patterns?
- What alternative explanations exist for the patterns you identified?
- What does this data NOT tell us?
4. Connect to action
Analysis without action is wasted effort. Always end with:
Based on this analysis, what specific changes would you recommend to our backlog?
Which existing stories are validated, challenged, or missing?
Validation framework
The agent's analysis needs to pass your judgment filter:
| Question | If "no" → |
|---|---|
| Do the patterns match my observations? | Investigate the discrepancy — either you missed something or the agent is wrong |
| Are the quotes representative? | The agent may have cherry-picked. Check the source. |
| Is the sample size sufficient for this claim? | Downgrade "strong pattern" to "emerging signal" for small samples |
| Does this change what we build? | If not, the analysis is interesting but not actionable |
| Am I being told what I want to hear? | Ask: "What in this data contradicts our current plan?" |
The confirmation bias trap
This is the biggest risk with Agent as Analyst. If you already believe something, the agent will find evidence for it in almost any dataset. Protect against this by:
- Asking for contradictory evidence first — "What in this data suggests our current approach is wrong?"
- Asking for alternative interpretations — "What else could explain these patterns?"
- Having someone else review — A second pair of eyes catches bias the agent reinforces.
RAG as a reliability amplifier
When using agents with large context (uploaded documents, project files), remember that Retrieval-Augmented Generation (RAG) amplifies reliability but doesn't guarantee truth. The agent retrieves and reasons over your documents, but it can still:
- Misinterpret ambiguous text
- Overweight recent or prominent data
- Miss connections across disparate documents
- Hallucinate connections that don't exist
Always treat agent analysis as a hypothesis to validate, not a conclusion to accept.
AI in practice: analyst patterns across industries
The Agent as Analyst pattern powers a wide range of client use cases:
- Revenue forecasting with confidence intervals — ARIMA models and time-series analysis with automated uncertainty quantification help finance teams make better predictions. See Finance & Analysis use cases.
- Customer churn prediction — Models identify at-risk accounts from usage patterns, support trends, and engagement signals. One team retained $5M in annual revenue. See Customer Success use cases.
- Cohort analysis and segmentation — Customer profile clustering reveals natural customer groups and what drives them, informing product and marketing strategy. See Data Analysis use cases.
Evaluating analyst agents in production
When the Agent as Analyst pattern runs at scale — processing production data, generating reports, or feeding dashboards — you need systematic evaluation beyond manual validation. The Continuous Alignment Techniques (CAT) framework covers instrumentation, golden datasets, and offline/online evaluation strategies for exactly this scenario.
Connecting to workflows
This pattern is used in: