Data Science
Same data, two tools, two different answers
A dashboard tells the seller what happened. A decision engine tells the seller what to do next, ranks the options by projected profit lift, and shows the math behind every recommendation. The clearest way to feel the difference is to put ten common operating decisions through both tools and read the columns side by side.
| Operating decision | Dashboard view (descriptive) | Decision engine view (prescriptive) |
|---|---|---|
| Should I scale Campaign A? | Campaign A produced THB 80,000 in attributed sales last week. | Campaign A's true ROAS is 1.2 vs SKU break-even 2.4 → REDUCE budget by ~40%; reallocate to Campaign D (true ROAS 3.8 vs break-even 2.1). |
| Should I reorder SKU B? | SKU B sold 500 units last month; current stock 240 units. | Days-of-supply 14d; lead time 18d; 11.11 demand-spike window in 21d → REORDER NOW (450 units, ranked #2 on profit-loss-on-stockout list). |
| Should I discount SKU C for Pay Day? | SKU C average ROAS during last Pay Day: 4.6×. | Contribution margin 18%; required volume lift to break even on 12% discount: +200%; historical Pay Day lift on this SKU: ~80% → DECLINE Pay Day participation; protect margin. |
| Should I match competitor price drop? | Competitor SKU now THB 290 (vs my THB 320). | Competitor margin estimate ~12% (vs my 28%); likely inventory clearance; price-elasticity history suggests <5% volume loss at THB 320 → HOLD price; competitor reverts in 7–10 days. |
| Is TikTok Shop earning more budget? | TikTok Shop ROAS 3.4× this month, +18% MoM. | True ROAS net of affiliate commission + elevated return rate is 2.1× vs break-even 2.6× → HOLD TikTok budget; reallocate to Shopee where marginal-ROAS curve sits higher at current spend. |
| Why is my hero SKU losing impressions? | Sponsored Search position dropped 3.2 → 5.8 over 14 days. | Keyword-cluster CPC +42% during pre-Pay-Day window; competitor raised bid 1 day after inventory clearance → INCREASE bid 18% for 11 days OR shift to Sponsored Discovery on broader query (same conversion at 31% lower CPC). |
| Should I stay in Free Shipping Programme? | Free Shipping Programme orders +24% YoY. | FSP costs THB 8.40/order seller-funded; FSP-margin 2.1pp lower than non-FSP, but conversion +14% → KEEP for SKUs A/B (margin absorbs); EXIT for SKU C (margin already 9%; FSP turns it loss-making). |
| Should I shift SKU D from Lazada to Shopee? | SKU D Lazada GMV THB 89K; Shopee GMV THB 152K. | Lazada per-unit contribution THB 18 (after Voucher Wallet seller portion + LazMall fee); Shopee per-unit THB 27. Lazada needs 21% LazMall conversion-lift to break even, currently 11% → REDIRECT 60% of remaining Lazada inventory to Shopee over 30d. |
| Should I bundle SKU E with SKU F? | SKU E and SKU F bought together by 18% of SKU E buyers. | SKU E margin 12% (margin-trap); SKU F margin 31%. Bundling lifts blended margin to 22% if F attaches 1:1; historical attach 18% with 1.4× lift at bundle-discount pricing → LAUNCH bundle at THB 480; model predicts 6.4% margin recovery over 90d. |
| Should I exit Category Z? | Category Z produced THB 2.1M GMV last quarter. | Aggregate contribution margin −3.4% after ads, fees, returns, 27-day payout lag. Three of fourteen SKUs profitable; carry 84% of category gross margin on <12% of volume → EXIT 11 loss-making SKUs over 60d; retain 3-SKU positive cohort with focused ad budget. |
The dashboard column answers "what happened" for each entity. The decision-engine column answers "what to do, ranked by projected profit lift, with the math attached." The seller approves the decision; the engine eliminates the analysis tax of getting to it.
The pattern repeats across every cell of the table. The dashboard surfaces an observation; the decision engine resolves it into an instruction the seller can approve, override with reason, or escalate. The instruction comes with three things the dashboard cannot supply: a comparison to the right break-even, a ranking against the other live decisions, and the math behind the recommendation.
What is behind that prescriptive column
Every recommendation in the right column is the output of a five-layer stack, each layer producing the inputs the next layer needs to be honest. Most failed seller-software deployments fail at Layer 2 — canonical SKU mapping across shops and platforms — because the data fragmentation is worse than vendors anticipate. Quality at that layer dominates everything above it.
Layer 1 — Ingestion
Pulls data from marketplaces, ad platforms, inventory systems, cost files, pricing history, and settlement records. Without it, every layer above is missing inputs.
Layer 2 — Normalisation
Maps SKUs, variations, campaigns, costs, and orders into one consistent structure across shops and platforms. The unglamorous step that makes everything else possible — and the step most spreadsheet-based setups never quite finish.
Layer 3 — Metric calculation
Computes contribution margin, true ROAS, break-even ROAS, ad waste, stockout risk, days-of-supply, promotion impact, price-elasticity signals, and channel margin. These are the numbers the recommendation column actually compares against.
Layer 4 — Pattern detection
Watches for the recognised failure modes — high revenue with low profit, ads on stockout-risk SKUs, discount-driven revenue, deteriorating margin, campaigns underperforming break-even, channel mismatches where the same SKU prints money on one marketplace and bleeds on another.
Layer 5 — Recommendation
Outputs ranked actions attached to specific entities — the right column of the table. The seller still decides; the engine eliminates the analysis tax of getting to a defensible decision in the first place.
Decision-making as a cost problem
Every seller decision has an opportunity cost. If budget goes to SKU A, it cannot go to SKU B. If cash is used to reorder slow inventory, it cannot be used for high-margin winners. If a seller discounts unnecessarily, margin is permanently gone. If a seller ignores stockout risk, future campaign performance pays the bill. The total cost of suboptimal decision-making is the sum of four kinds of leakage:
wasted spend + lost margin + missed upside + operational dragA decision engine reduces that cost by surfacing the conflict between two decisions early — before either one has fully fired.
Why decision engines matter more now
Commerce is becoming more algorithmic. Marketplaces optimize search, ads, recommendations, logistics, and pricing signals at increasing resolution. Video commerce and retail media make discovery faster and more competitive. Major platforms are investing heavily in AI-driven shopping and advertising. For sellers, the practical effect is that the operating environment is moving faster than monthly P&L review cycles can keep up with.
Manual reporting cannot keep up with real-time decisions.
Human judgment still matters
A decision engine should not replace the seller. It should improve the seller's judgment. The seller still understands brand, supplier relationships, product quality, customer expectations, and the longer-term strategy that does not show up in any dataset. The system contributes the things software is better at — connecting data, calculating metrics, detecting patterns, monitoring change, and surfacing anomalies the human eye would miss in a spreadsheet at 2 a.m. The best operating model combines human context with machine consistency.
Limitations and where this argument breaks
- Account-size lower bound. Decision engines compound on data quality. Below ~THB 200K monthly revenue, the operational overhead of running the architecture exceeds recovered margin; simpler heuristics (a per-category margin lookup, manual weekly reconciliation) outperform.
- Layer 2 is the silent failure point. The architecture sketched here looks clean in a diagram and is brutal to implement in production. Most failed seller-software deployments fail at Layer 2 (canonical SKU mapping across shops and platforms) because the data fragmentation is worse than vendors anticipate. Quality at this layer dominates everything above it.
- Recommendation framing risk. A recommendation layer that overclaims confidence will burn trust faster than no recommendations at all. The output should be 'here is what we think you should do, here is the math, here is the confidence' — not 'do this' as a directive. Auditable math is the trust layer.
- Pattern-detection lag. Layer 4 detects recognised failure modes but cannot anticipate genuinely novel patterns (a new campaign mechanic, a sudden category disruption, a one-off platform policy change). Human judgment still owns the long tail.
- Cross-platform coordination. The architecture works cleanly within one platform's data. Cross-platform coordination (Shopee × Lazada × TikTok Shop) requires the canonical product catalog at Layer 2 to bind the same physical SKU across platforms — an additional step that scales the operational complexity meaningfully.
- Internal-data scope. The 6–10 percentage-point margin-recovery figure is aggregated across the SEA-6 Thai marketplace seller accounts we model directly. Not a population claim about all decision-engine deployments; explicitly excludes accounts below the size bound and the negotiated-rate enterprise tier.
Methodology
Public-data citations are taken from the 2025 ScienceDirect systematic review of ML approaches in inventory control (which formalises the multi-layer ingest → forecast → optimise → policy stack), Reuters reporting on Sea Limited's AI investment cadence (3 March 2026) and the Google–Sea agentic shopping partnership (19 February 2026), Sea Limited's 4Q25 / 1Q26 investor disclosures, and the Bain e-Conomy SEA 2025 commentary on regional commerce velocity.
Internal-data claims — the 6–10 percentage-point margin-recovery figure, the layer-by-layer value-compounding pattern in the chart, the typical failure-mode distribution — are aggregated across approximately 400 active marketplace seller accounts across the DataGlass research methodology sample frame (Jan 2024 – Apr 2026, 28-month observation window). Margin-gap measurements are computed on rolling 90-day windows.
FAQ
Four short, query-shaped questions and answers that compress the report for skimmers and AI search. The "vs dashboard" and "when is it not worth using?" questions live next to the relevant evidence in the body above.
Not another dashboard. A profit-aware decision engine — five layers, ranked recommendations, auditable math, human judgment on top.