Web/app events
Pseudonymous user, session, timestamp, event, item array, campaign, device, consent
Data Analysis Agent for E-commerce
An agent that finds where profit disappears.
Connect spend, sessions, orders, product margin, returns, fulfillment, and repeat purchases so the agent can look past attributed revenue and investigate the economics underneath it.
The recurring problem
Revenue is up. The economics are less clear.
Channel, behavioral, order, margin, fulfillment, return, and support systems use different identities, clocks, and attribution rules.
The representative scenario follows a channel-profitability review: reconcile spend, sessions, orders, margin, returns, and repeat purchase behavior, then examine why attributed revenue and contribution rank disagree. It is a synthetic composite, not a claimed e-commerce deployment.
Event
Session
Attribution window
CAC
Contribution margin
Cohort
Repeat rate
LTV horizon
On-time delivery
What the agent connects
The agent has to follow the customer across systems.
Illustrative dataset
Northstar Home
A fictional direct-to-consumer home-goods retailer testing whether apparent channel growth remains attractive after margin, returns, and repeat behavior.
- Date range
- 1 November 2025–31 July 2026
- Disclosure
- Synthetic event, ad, order, cost, shipment, return, support, and pseudonymous identity records shaped around documented e-commerce schemas.
Advertising
Spend, impressions, clicks, reported conversions, currency
Orders & product cost
Revenue, discount, quantity, COGS, pick/pack, warehouse
Shipment events
Carrier, service, scan, location, promise, actual time
Returns & refunds
Reason, condition, requested/received/refunded dates, amount
Support & identity
Order, reason, resolution, consent, cohort, approved link
Item and transaction IDs connect events to catalog and orders; approved pseudonymous identity links sessions and customers; order, parcel, return, and case IDs connect the post-purchase journey.
Session, channel, attribution model/window, CAC, contribution margin, cohort, repeat purchase, LTV horizon, on-time delivery, return rate, and refund cycle time.
A campaign naming change breaks spend-to-session mapping for ten days and some returns arrive after the reporting cutoff. Both effects must remain visible.
Watch the agent investigate
Which channels still pay after returns?
Representative Interface · Illustrative Analysis. Every value uses synthetic records; this shows representative agent behavior, not a client result.
Growth and finance lead
Which acquisition channels remain profitable after product margin, discounts, returns, and repeat purchases are included?
The agent’s interpretation
Use paid-and-organic last-touch for the primary view, a 90-day repeat window, contribution after COGS, discount, payment, pick/pack, and observed returns, with sensitivity for returns not yet received.
web_events
ad_spend_daily
orders_lines
product_costs
returns_refunds
customer_identity
- Reconcile campaign and transaction identifiers
- Apply the stated attribution window
- Calculate order and return-adjusted contribution
- Build acquisition cohorts and compare sensitivity cases
Example Output
Revenue rank and contribution rank do not match.
Illustrative analysis shows Paid Social first in attributed revenue but third in 90-day contribution after discounts and returns. Organic Search has lower acquisition volume but stronger repeat contribution.
Channel
CAC
Return rate
90-day contribution
Organic search
$18
8.4%
$82k
Paid search
$41
10.1%
$63k
Paid social
$56
17.8%
$38k
Evidence & definitions
6 sources · last-touch / 90 days · observed returns + sensitivity · synthetic records
Follow-up question
Compare 30- and 90-day windows and exclude platform-modeled conversions.
The agent narrows the answer
Paid Social falls below the illustrative review threshold at 30 days and remains most sensitive to unreceived returns; the result is flagged for measurement and creative review.
Channel table · cohort view · definition sensitivity reportReport / table output available in an implementation
Human decisionGrowth and finance teams review measurement and spend assumptions before changing budgets.
Where else the agent works
Three jobs beyond the revenue dashboard.
01 Agent job Acquisition, funnel, and customer economics Channel dashboards show revenue or conversions without product margin, discounts, returns, repeat purchases, or the assumptions behind attribution.
People asking
Heads of growth, CFOs, performance marketers, e-commerce directors, and customer analysts.
Real questions
- Which channels appear profitable after margin, discounts, returns, and repeat purchases?
- How does the answer change by attribution window?
- Which cohorts recover acquisition cost under the approved definition?
Sources the agent uses
Ad spend and clicks · Web/app events and sessions · Orders, customers, and discounts · Product cost, payments, and returns
Work the agent performs
Reconcile spend and campaign IDs, sessionize events, apply an explicit attribution model/window, connect consented identities to orders, calculate agreed contribution and cohort repeat behavior, and compare definitions rather than presenting one as truth.
Output it prepares
A channel contribution table, funnel, cohort curve, and assumption panel.
Human decision or workflow
Inform budget and measurement review, creative or landing-page work—not autonomous media buying.
What must be validated
Identity loss, modeled conversions, attribution scope, late returns, cost allocation, and LTV horizon materially change the answer.
02 Agent job Product and merchandising performance Products receive strong traffic but lose buyers between view, cart, and checkout while price, variant availability, and delivery-promise effects are unclear.
People asking
Merchandising leaders, digital product managers, e-commerce directors, category managers, and inventory planners.
Real questions
- Which products have strong views but weak progression?
- What patterns distinguish the lost buyers?
- Which bundles increase contribution rather than only order value?
Sources the agent uses
Item-level events · Catalog and content versions · Price, discount, and inventory · Orders, returns, search, and delivery promise
Work the agent performs
Build item funnels, deduplicate events, segment by device/source/cohort, join availability and price at event time, connect completed orders and returns, and compare products with sufficient exposure.
Output it prepares
A product funnel, abandonment comparison, availability exceptions, and merchandising worklist.
Human decision or workflow
Prioritize content, price tests, assortment, inventory, checkout, or instrumentation work.
What must be validated
Instrumentation, bot filtering, cross-device identity, exposure, experiment design, and sample size require validation.
03 Agent job Fulfillment, returns, and customer experience Late deliveries, returns, refunds, contacts, and poor reviews rise, but each system describes the order differently.
People asking
Operations directors, fulfillment and CX leaders, carrier managers, and finance partners.
Real questions
- Which routes, products, warehouses, or cohorts drive late delivery and contacts?
- Are returns tied to product fit, damage, delay, or expectation?
- Where does refund cycle time exceed policy?
Sources the agent uses
Orders and lines · Warehouse and carrier events · Returns, refunds, and reviews · Support contacts and customer cohorts
Work the agent performs
Construct order and parcel timelines, compare promise with actual delivery, apply reviewable reason mappings, calculate rates by product/warehouse/carrier/lane/cohort, and trace exceptions to source events.
Output it prepares
A fulfillment exception table, delivery-time comparison, return-reason matrix, and case queue.
Human decision or workflow
Review carrier/service, warehouse process, product information, packaging, or customer recovery.
What must be validated
Scan completeness, split shipments, reason-code quality, review bias, and promised-date logic must be visible.
From scenario to implementation
Begin with an agent that can investigate beyond attributed revenue.
Working Data Analysis Agent
The reusable foundation already handles connected-source questions, follow-up analysis, evidence, visual and downloadable outputs, permissions, and guidance. E-commerce configuration makes identity, time, margin, attribution, and post-purchase assumptions explicit.
- Conversational investigation across acquisition, behavior, order, product, return, and fulfillment sources
- Cross-source identity and transaction joins with visible assumptions
- Funnel, cohort, contribution, exception, report, and download outputs
- Permissions and guidance for customer, support, commercial, and finance users
Configure the economics—not just the event feed.
Attribution window, session logic, identity scope, cost boundary, return treatment, cohort horizon, and delivery promise determine what the analysis means. They are agreed, configured, and exposed with the answer.
- Analytics, advertising, commerce, catalog, cost, fulfillment, returns, support, and approved identity connections
- Session, channel, attribution, CAC, contribution, cohort, LTV-horizon, return, and delivery definitions
- Campaign, item, order, parcel, customer, warehouse, carrier, and time relationships
- Consent boundaries, team permissions, late-data handling, sensitivity views, and decision outputs
Keep the agent reviewable
Keep attribution assumptions inside the answer.
Display attribution model, lookback window, identity scope, cohort horizon, and cost allocation with the result.
Separate observed, modeled, unattributed, and late-arriving events rather than blending them invisibly.
Use consented and minimized identity links; scope access to customer and support information.
Treat attribution, LTV, incrementality, and causal explanations as assumptions to inspect—not unquestionable truth.
A bounded client implementation
Start with one commercial decision.
Choose the decision and intended result
Select channel economics, product funnel, or fulfillment experience; identify the users; and record the current analysis effort, cycle time, or commercial/operating measure.
Agree questions and horizons
Define attribution model/window, cohort horizon, cost boundary, return treatment, and expected outputs.
Connect approved source samples
Use representative or client-approved events, spend, orders, product cost, shipments, returns, and support data.
Define identity and calculations
Document consented identity links, session logic, campaign mapping, contribution, CAC, and delivery/return definitions.
Set permissions and quality rules
Scope customer/support access and handling of modeled events, missing IDs, late returns, and incomplete scans.
Configure analysis and output
Shape the agent, channel/funnel/cohort views, sensitivity panel, and review report.
Validate against known periods
Reconcile orders and spend, replay agreed questions, and test attribution and late-data sensitivity.
Evaluate and choose the next step
Review the capability against the agreed questions and measures, then choose the next instrumentation, workflow, report, or commerce-system integration where justified.
Define the result before deployment
Measure whether better answers improve commercial work.
Record the effort and delay behind recurring growth, merchandising, and fulfillment questions, then evaluate whether the agent shortens analysis cycles and helps teams make better-supported decisions about conversion, contribution, returns, and service.
The baseline, review period, and acceptable evidence are agreed before deployment; no improvement is assumed.
- Time to reconcile recurring channel, product, or cohort questions
- Reporting and manual data-preparation effort
- Conversion, contribution margin, return, or resolution measures selected for the use case
- Adoption, repeat investigation, and reviewed-output completion
Extend where the business requires
Connect the analysis to the next commerce decision.
Company A can add proprietary platform connections, experimentation or merchandising workspaces, recurring reports, return and support workflows, operational alerts, data pipelines, and controlled actions in commerce or fulfillment systems.
Possible next workflows
- Add experimentation and merchandising workspaces
- Connect fulfillment exception workflows
- Build customer-service or returns applications
- Schedule margin and cohort reporting
- Operate instrumentation and definition quality checks
What custom development can add
- Analytics, advertising, commerce, and fulfillment integrations
- E-commerce operations applications
- Returns and support workflows
- Identity and permission-aware interfaces
- Managed data quality and reporting
Identify a bounded first implementation