Use-case library
What agents get asked, function by function
For each function: questions people would put to an analytics agent, the data and the written definitions needed to answer them correctly, what can go wrong, and the controls that prevent it. Use it to choose a first domain and to see how much definition work sits behind a question that sounds simple.
Vendor-neutral. The numbers inside the example questions (14 days, 10%, 90 days) are placeholders for the thresholds your business already uses.
The same pattern in every function
The questions differ, but the hard part does not. Each one hides three or four definition choices (which date, which currency, which version of the hierarchy, which attribution model) that an analyst resolves from memory and an agent resolves by guessing unless they are written into the semantic layer. And each function holds data that some of its own people should not see. That is why the readiness assessment asks about definitions and access before it asks about models.
Finance
Finance questions have exact answers and an audit trail behind them, which makes finance a demanding first domain and a useful test of whether definitions are really written down.
Questions a user would ask
- Why did gross margin in EMEA fall compared with plan last quarter, and which product lines drove it?
- Which cost centers are more than 10% over budget year to date, and on which expense lines?
- What is days sales outstanding by region this month, and which customers moved it most?
- Where does recognised revenue in the ledger differ from bookings in the CRM for September? List the largest differences.
- If the mid tier had been priced 5% higher last quarter at the same volume, what would revenue have been?
Data needed
- General ledger and sub-ledgers
- Budget and plan versions
- Accounts receivable and billing
- CRM bookings
- Exchange rates
- Cost center and legal entity hierarchies, with effective dates
Definitions to write down first
- Revenue, kept separate as booked, billed, and recognised
- Gross margin: which cost lines are included
- Fiscal calendar and period status (open or closed)
- Currency conversion: which rate type and which date
- Days sales outstanding: formula and averaging window
- Cost center roll-ups as of a given date, since hierarchies change
Governance risks
- Material non-public figures reached before an earnings release
- Booked and recognised revenue mixed in one answer
- Open-period numbers quoted as final
- Payroll and compensation detail exposed through expense questions
- Scenario answers (the pricing question) read as actuals
Controls
- Row-level rules by legal entity and cost center, applied to the person asking
- Compensation columns masked for everyone outside payroll
- Every answer states the definition used, the period status, and the as-of time
- Scenario answers labelled as estimates, with their assumptions listed
- Read-only access; nothing the agent produces feeds external reporting without the normal review
Where to start: Budget-versus-actual questions for closed periods. The definitions usually exist already, the answers are checkable, and the audience is small.
Sales and revenue operations
Revenue operations teams field a steady stream of pipeline and performance questions that each take an analyst a few hours. The catch is that CRM data changes daily and definitions of stages and wins vary by team.
Questions a user would ask
- Which deals in the commit category have had no logged activity in the last 14 days?
- How has win rate by segment changed this quarter compared with last, and where did it change most?
- Why is pipeline coverage for next quarter lower in the West region than in the others?
- Which accounts grew product usage by more than 20% but have no open expansion opportunity?
- What is the median sales cycle for partner-sourced deals compared with direct deals?
Data needed
- CRM opportunities, with stage history
- Activities (calls, meetings, emails logged)
- Accounts and contacts
- Product usage
- Quotas and territory assignments
- Marketing source and attribution fields
Definitions to write down first
- Pipeline stages and the criteria for entering each one
- Win rate: by count or by value, and which outcomes are excluded
- Pipeline coverage: which quarter, which stages, against which target
- ARR, bookings, and contract value, kept distinct
- Segment and territory as of a given date, since accounts move
- Deal source: partner, direct, or marketing, and who decides
Governance risks
- Reps seeing peers' quota attainment or compensation
- Current CRM values used for trend questions that need stage history
- Stale or incomplete CRM entries treated as fact
- Forecast figures shared outside the forecast audience
- An agent with write access editing opportunity records
Controls
- Row-level rules following the sales hierarchy and territory
- Quota and compensation fields masked outside sales leadership and finance
- Trend questions answered from snapshotted stage history, not current values
- Answers show record counts and the time of the last CRM sync
- Read-only access; any suggested update goes to a person to make
Where to start: Pipeline hygiene questions (stale deals, missing fields) for one region. Wrong answers are cheap, and the value is visible to sales managers within weeks.
Supply chain
Supply chain questions cut across purchasing, inventory, logistics, and demand systems that rarely share keys or units. That makes the definitions work harder, and it is also why an agent that gets it right saves real analyst time.
Questions a user would ask
- Which items will run out within 21 days at the current rate of sale, and at which locations?
- Why did on-time delivery from our largest supplier fall last month?
- What is the landed cost difference between our two sourcing options for this part?
- Which distribution centers hold more than 90 days of supply for slow-moving items?
- How did the port delay affect order fill rate by region?
Data needed
- ERP inventory positions and item master
- Purchase orders and supplier master
- Shipments, advance ship notices, and logistics events
- Sales orders
- Demand forecast versions
- Location hierarchy
Definitions to write down first
- On-time: against the promised or the requested date, and the tolerance window
- Fill rate: by line, by unit, or by order
- Days of supply and the demand figure it divides by
- Landed cost: which freight, duty, and handling components are included
- Units of measure and the conversions between them
- Forecast version: which one counts as the plan
Governance risks
- Supplier prices and contract terms seen by people who should not see them
- Forecast figures presented as actuals
- Unit-of-measure mistakes producing answers that are off by a case pack
- Inventory snapshots hours old treated as live
- An agent placing or changing purchase orders
Controls
- Contract terms and supplier pricing restricted by column
- Unit conversions defined in the semantic layer, never left to the model
- Answers state the snapshot time of the inventory data
- Forecast and actual figures carry distinct names in every answer
- Reorder suggestions go to a planner for approval; the agent does not transact
Where to start: Days-of-supply and stock-out questions for one region or product family, checked against what planners already track.
Customer support
Support leaders want trends: why contact volume moved, which customers are struggling, whether a process change helped. The data holds a lot of personal information in free text, so the governance design matters as much as the answers.
Questions a user would ask
- What were the top five contact reasons this week, and which grew fastest?
- Which enterprise customers have had more than three escalations in the last 30 days?
- Did first response time change after the queue restructure at the start of the month?
- What share of billing tickets were reopened after being marked resolved?
- Which product release lines up with the rise in login-related tickets?
Data needed
- Ticketing system: tickets, status history, assignments
- Customer satisfaction surveys
- Customer and account master, with tier
- Product release history
- Contact center call records
Definitions to write down first
- First response time: business hours or calendar hours, and which clock starts it
- Resolved, closed, and reopened, and the time window for a reopen
- Contact reason categories, and how they changed over time
- Escalation: which transitions count
- Customer tier as of the ticket date
Governance risks
- Ticket text containing personal, payment, or health details
- An agent quoting one customer's messages in an answer about another
- Timing presented as cause (the release question)
- Agent performance data used for individual evaluation without the right access
Controls
- Aggregate answers by default; ticket bodies behind separate, narrower access
- Free text redacted or masked before it reaches the agent
- Answers about releases and incidents labelled as association, with the dates shown
- Row-level rules on per-agent performance data
- Audit trail on any query that reads ticket text
Where to start: Contact-reason trends and response-time questions using structured fields only. Leave ticket text out of the first phase.
Marketing
Marketing questions often have two defensible answers because the attribution model is a choice. An agent is useful here when it names the model it used, and harmful when it quietly picks one.
Questions a user would ask
- Which campaigns sourced pipeline that became closed-won this half, by channel?
- What is cost per qualified lead by channel over the last 90 days?
- How often did webinar attendance lead to an opportunity within 60 days?
- Which segments had unusually low email engagement this month?
- How does trial-to-paid conversion compare for users from paid search and from organic search?
Data needed
- Ad platform spend
- Marketing automation: campaigns, emails, events
- Web analytics
- CRM leads, contacts, and opportunities
- Product sign-ups and trials
- Consent and preference records
Definitions to write down first
- Attribution model (first touch, last touch, or multi-touch), named in every answer
- Qualified lead: the criteria and who sets them
- Conversion windows for each question type
- Channel categories, and how sources map into them
- Spend: currency and the date it counts against
Governance risks
- Individual-level data used in ways the person did not consent to
- Two attribution models producing two answers, with no sign of which was used
- Ad-platform reported conversions mixed with internal CRM conversions
- Spend data arriving days late and understating recent cost
Controls
- Consent flags enforced as row filters where the data is read
- Aggregate answers by default for anything touching individuals
- Attribution model and conversion window stated in every answer
- Platform-reported and internal figures kept under different names
- Answers show the latest spend date loaded
Where to start: Spend and cost-per-lead questions by channel, where the definitions are simplest and the data is already aggregated.
How engineers build the controls
The controls above are the same few mechanisms applied to different data. The builder side is on Agentic Lakehouse:row-level security,column-level security,data masking,AI agent authorization, andthe AI semantic layer.