Ecommerce Performance Analysis — AI and Data Workflow
Applied AI · Data Analytics · Prompt Engineering · Business Intelligence
Project Overview
Ecommerce Performance Analysis is an interactive proof of concept exploring how AI-assisted data analysis can transform a standard ecommerce sales export into clear, decision-ready business intelligence.
The project demonstrates a structured workflow for validating data, calculating supported commercial metrics, comparing product performance, identifying significant patterns and translating findings into prioritised recommendations.
Rather than presenting AI as an autonomous decision-maker, the system is designed around human review. Calculations remain connected to their source data, analytical limitations are made visible and every recommendation requires commercial judgement before action is taken.
All figures and commercial records used in the prototype are simulated.
The Challenge
Ecommerce platforms can generate large quantities of transactional data, but exported reports do not automatically provide meaningful commercial insight.
A typical sales export may contain order dates, product names, categories, quantities, prices, discounts, refunds and customer identifiers. Before this information can support a decision, several questions must be resolved:
Is the dataset complete and internally consistent?
Are duplicate or invalid records affecting the totals?
Which calculations are genuinely supported by the available fields?
Which products produce revenue without excessive returns?
Which categories combine sales volume with healthy margins?
Are observed patterns commercially meaningful or simply coincidental?
What additional evidence is required before action is taken?
Manually answering these questions can involve multiple spreadsheets, repeated calculations and significant interpretation.
The challenge was therefore to design an AI-assisted workflow that could accelerate the analytical process without disguising uncertainty or presenting recommendations as established facts.
Project Objective
The objective was to create a reusable ecommerce analysis system that could:
Inspect the structure of a Shopify-style CSV export.
Identify missing values, duplicates and inconsistent records.
Calculate only the metrics supported by the dataset.
Compare revenue, orders, margins and return signals.
Highlight commercially important patterns.
Separate verified findings from assumptions.
Translate approved insights into a concise executive report.
Preserve human responsibility throughout the process.
The project also demonstrates how prompt engineering can be used to create a repeatable analytical workflow rather than a collection of isolated AI queries.
The Approach
The project was organised into five connected stages:
Validate the dataset.
Explore commercial performance.
Compare products and categories.
Interpret significant patterns.
Report approved findings and next steps.
Each stage adds a new layer of understanding while retaining the evidence, limitations and decisions established during the previous stage.
Stage One — Data Validation
The first stage examines whether the uploaded data is suitable for analysis.
The validation process checks:
Available columns and data types.
Required order and product fields.
Missing values.
Duplicate order identifiers.
Negative or unexpected transaction values.
Inconsistent product categories.
Currency inconsistencies.
Records requiring manual review.
The interface distinguishes valid records from unresolved issues and explains how each problem affects the analysis.
For example, duplicate order identifiers are excluded from headline calculations until they have been reviewed. Missing product categories can be mapped against a catalogue, while negative transaction values may require confirmation as refunds.
The prototype uses a simulated dataset containing 4,235 rows. Of these, 4,218 are treated as analytically usable, while 17 are routed to a review queue.
These figures demonstrate the workflow and should not be interpreted as results from a real business.
Stage Two — Exploratory Performance Analysis
Once the usable dataset has been established, the workflow calculates a supported set of commercial measures.
These include:
Net revenue.
Total valid orders.
Average order value.
Revenue by category.
Revenue by product.
Units sold.
Illustrative gross margin.
Return signals.
Weekly performance trends.
The workspace provides an immediate overview of performance while keeping the data-quality status visible.
The aim is not simply to display KPIs. It is to help the reviewer understand which measures are reliable, which are illustrative and which cannot be calculated from the available data.
Customer lifetime value, marketing attribution and causal campaign performance, for example, cannot be verified from a short transactional export alone.
Stage Three — Product and Category Comparison
The product-performance interface compares several commercial signals together.
Revenue is considered alongside:
Units sold.
Gross-margin contribution.
Return rates.
Category performance.
Potential merchandising roles.
This prevents a high-revenue product from automatically being classified as a strong performer.
A product may generate substantial sales while also producing high returns, weak margins or operational friction. Conversely, a lower-revenue accessory may provide healthy margins, low returns and an opportunity to increase basket value.
The simulated analysis identifies illustrative patterns such as:
Outerwear leading category revenue.
Knitwear showing a strong margin signal.
Footwear generating significant revenue but elevated returns.
Accessories presenting potential bundle opportunities.
Certain products requiring further investigation before increased promotion.
Stage Four — Evidence-Based Insight Review
The insight-review stage translates analytical patterns into commercially meaningful findings.
Every insight contains four elements:
Evidence: What the dataset shows.
Interpretation: Why the pattern may matter.
Limitation: What the current data cannot confirm.
Next step: What the business should investigate or test.
This structure is designed to prevent the system from turning correlation into certainty.
For example, the prototype may identify a higher average order value among returning customers. However, the short reporting period and anonymised customer identifiers limit the strength of that finding. The appropriate recommendation is therefore to validate the pattern using a longer customer-history dataset.
The resulting recommendations are organised by potential impact, strength of evidence and implementation effort.
(Placeholder – Evidence-Based Insight Review and Recommendation Matrix)
Stage Five — Executive Reporting
The final stage turns approved findings into a concise executive summary.
The report communicates:
Overall commercial performance.
Leading categories and products.
Material risks.
Potential growth opportunities.
Outstanding evidence gaps.
Recommended next actions.
A proposed 30-day testing and review plan.
The report maintains a clear distinction between verified calculations, analytical interpretation and proposed business action.
The system does not issue automated commercial decisions. It prepares evidence so that a human reviewer can decide what should happen next.
The AI and Data Workflow
The analytical process is supported by a reusable prompt pattern based on task decomposition and structured outputs.
The prompt instructs the AI to:
Inspect the uploaded dataset before calculating anything.
Identify columns, types, missing values and duplicates.
Explain limitations affecting the analysis.
Calculate only metrics supported by the available data.
Produce appropriate visualisations.
Identify commercially significant findings.
Explain why each finding matters.
Request additional information where necessary.
Separate facts, assumptions and recommendations.
communicate results in plain British English.
This approach applies principles from the Cognitive Verifier, Template and Semantic Filter prompt patterns.
It treats prompt engineering as part of the analytical system rather than as a one-off instruction.
Interface Design
The website was designed as a connected analytical experience rather than a static collection of charts.
Its principal views are:
Overview
Introduces the commercial problem, project scope, analytical workflow and responsible-use model.
Workspace
Presents supported KPIs, revenue trends, category distribution and important product signals.
Validate
Documents data-quality checks, unresolved records, analytical readiness and limitations.
Products
Compares products and categories using revenue, units, margin and return signals.
Insights
Connects significant findings to their evidence, limitations and recommended next questions.
Report
Presents the approved findings as a business-focused executive summary and action plan.
The responsive design allows the workflow to remain usable across desktop and mobile layouts.
Human Oversight and Responsible Use
Human review is treated as a functional part of the workflow.
The system is designed around several constraints:
AI-generated findings are proposals rather than confirmed decisions.
Unresolved data-quality issues remain visible.
Unsupported calculations are not presented as reliable metrics.
Recommendations include evidence and limitations.
Commercial actions require human approval.
Simulated figures are clearly identified.
The original dataset remains the primary evidence source.
The central principle is:
AI accelerates analysis. People own the decision.
Outcome
The completed project demonstrates how Advanced Data Analysis can support an end-to-end ecommerce intelligence workflow.
It shows how raw transactional information can be transformed into:
A dataset inventory.
A data-quality report.
An exploratory performance dashboard.
Product and category comparisons.
Evidence-based findings.
A recommendation matrix.
An executive summary.
A reusable analytical prompt pattern.
Because the project uses simulated data, its outcome is the workflow and interface itself—not a claim of measurable business improvement.
Skills Demonstrated
AI-assisted data analysis.
Exploratory data analysis.
CSV and spreadsheet interpretation.
Data validation and quality assessment.
Ecommerce KPI design.
Product and category analysis.
Data visualisation.
Business intelligence.
Prompt engineering.
Structured output design.
Task decomposition.
Human-in-the-loop workflow design.
Commercial interpretation.
Executive reporting.
Responsive UX and interface design.
Static website development and deployment.
Tools and Technologies
ChatGPT Advanced Data Analysis.
CSV and spreadsheet-style datasets.
Structured prompt patterns.
HTML.
CSS.
JavaScript.
Responsive interface design.
Hostinger static website hosting.
Reflection
The project reinforced that effective AI-assisted analysis depends on more than producing charts or summary statistics.
The most important design decisions involved establishing:
Which questions the dataset could reasonably answer.
How uncertainty should be communicated.
Where human judgement remained necessary.
How findings could be traced back to evidence.
How analytical outputs could become practical business decisions.
The result is a portfolio project that demonstrates AI-assisted analysis as a governed business workflow rather than an automated answer generator.
Potential Future Development
The concept could be extended by adding:
Direct Shopify data imports.
Longer-term customer cohort analysis.
Stock and inventory coverage.
Marketing-channel performance.
Product-page conversion data.
Return-reason analysis.
Forecasting and scenario modelling.
Downloadable Excel and PDF reports.
Saved decision logs.
Role-based review and approval.
Scheduled analysis of new sales exports.
