Business & Research Use Cases Best Practices
Ten practices for keeping research, analysis, CI, meeting, and internal KB agents accurate and properly sourced.
Use this list when demos look fluent but you are not yet sure the claims would survive a skeptical reviewer.
How to Use This Checklist
- Work top to bottom; earlier items prevent expensive rework.
- Apply the list per use case (research ≠ meeting notes ≠ RAG) with the same evidence bar.
- Tick only what is true in the running system, not the slide deck.
- Revisit after the first real user-visible error - that is when teams usually skip citations "just this once."
- Keep a filled copy next to the design doc and eval plan.
A - Evidence First
- 1. Define done as verifiable claims, not vibes. "Three cited pricing changes in 90 days" beats "good market overview."
- 2. Force claim → source links in the output schema. Freeform "trust me" paragraphs are a defect even when they sound right.
- 3. Rank sources explicitly (primary / secondary / tertiary). Official pages and systems of record outrank undated SEO posts and hallway Slack lore.
B - Runtime Discipline
- 4. Bound every run: max turns, max sources, max cost, max time. Curiosity without a budget becomes a browsing bill.
- 5. Prefer read-only tools until review quality is proven. Draft emails, draft CRM updates, and draft wiki pages before any auto-send or overwrite.
- 6. Log the evidence trail, not only the final memo. Queries, hits, chosen chunks, code/SQL, and stop reasons are the unit of debug and audit.
C - Task-Specific Quality
- 7. Match verification to task shape. Research needs citations; data analysis needs executed code or governed metrics; meetings need quote/timestamp grounding; CI needs diffs against baseline.
- 8. Refuse or partial-answer when evidence is thin. "Open questions" and "insufficient sources" are successful outcomes, not product failures.
- 9. Separate synthesis from collection. Do not let a single unreviewed pass both scrape and declare strategy; structure stages or agents so conflicts stay visible.
D - People and Process
- 10. Name a human owner for publish standards and evals. Someone must score faithfulness samples, maintain golden questions, and can lower autonomy after incidents.
Applying the Habits by Use Case
| Use case | Habits to stress | Extra gate |
|---|---|---|
| Multi-source research | 1-4, 6, 8-9 | Min independent sources |
| Data analysis | 1, 4, 6-8 | Code/SQL + row-count checks |
| Market / CI monitors | 3-6, 8, 10 | Diff severity thresholds |
| Meeting / comms | 2, 5-8, 10 | No external auto-send |
| Internal KB RAG | 2-3, 6, 8, 10 | ACL-aware retrieval + refuse |
Quick "Do Not Ship" Signals
- Citations point to URLs never fetched in the trace.
- The agent averages conflicting numbers into one false figure.
- Warehouse or sandbox credentials are broader than the question requires.
- Meeting actions lack owners but still create tickets.
- RAG answers cite docs the user is not allowed to open.
- No golden set exists, only executive enthusiasm.
FAQs
How many checklist items are mandatory?
Treat 1, 2, 4, 5, and 10 as launch gates. The rest scale quality; skipping the gates scales incidents.
Can we ship without structured citations if users hate footnotes?
Keep citations in the trace and offer a "show sources" toggle. Hiding UI is fine; deleting the evidence model is not.
What is a good minimum eval set?
Twenty graded tasks spanning easy, multi-hop, conflict, and refuse cases - plus ten permission tests for internal KB agents.
How do these practices differ from coding-agent best practices?
Coding agents lean on tests and CI. Knowledge-work agents lean on sources, faithfulness, and human publish gates. Both need budgets and traces.
Should every knowledge agent be multi-agent?
No. Split roles when collection, analysis, and writing need different tools or models. Complexity must pay rent.
How often should we sample production answers?
Weekly for new systems; monthly once stable - and on every incident. Sample both thumbs-down and random successes.
What is the most common accuracy failure?
Fluent synthesis over thin or secondary evidence, delivered without open questions.
Can cheap models follow these habits?
They can if schemas, retrieval, and runtime checks carry the load. Do not rely on a small model to "remember" to cite.
How do we handle regulated industries?
Same ten practices plus mandatory human approval, retention rules, and industry-specific deny topics. See industry use-case pages for risk context.
When should a successful agent path become a non-agent pipeline?
When the question and sources stabilize (weekly metric, fixed FAQ). Promote to scheduled jobs; keep agents for ad-hoc work.
Who should fill this checklist?
Pair product, engineering, and a domain reviewer (analyst, SE, HR ops, etc.). Accuracy is a shared property.
How does this connect to the rest of the use-cases group?
Overview pages decide if an agent fits; this section's practices decide if a knowledge-work agent is trustworthy enough to keep.
Related
- How Knowledge-Work Agents Differ from Coding Agents - why these habits differ from green-test culture
- Business & Research Use Cases Basics - first cited-brief walkthrough
- Research Agents: Multi-Source Investigation and Synthesis - research quality gate detail
- Internal Knowledge-Base Agents with RAG Grounding - closed-corpus grounding
- Use Cases for AI Agents Best Practices - broader scoping habits
Stack versions: Pins from the category manifest (verify at build): OpenRouter (~315+ models, July 2026 pricing/fees); LangGraph 1.0+; CrewAI 1.14+; Microsoft Agent Framework 1.0; Vercel AI SDK 6; Pydantic AI (latest); LlamaIndex (latest); OpenAI Agents SDK (latest + MCP); MCP (Linux Foundation governance); A2A (HTTP+SSE+JSON-RPC 2.0); Solana
@solana/web3.js+@solana/spl-token.