“Pure-AI inventory forecasting overcommits Q4 every year. The model has not lived through enough Q4s to know what ‘pull-forward demand’ actually looks like in October.”
Inventory forecasting on Amazon is the function where the cost of getting it wrong shows up most cleanly on the brand’s P&L. Underforecast and the brand stocks out. They lose Buy Box for a four-to-six week recovery cycle. They forfeit the organic-rank position the listing had built. Overforecast and the brand pays FBA storage fees, aged-inventory surcharges, and the cost of pulling unsold inventory back out of the network when Q1 comes.
The inventory forecast workflow we run is built around one observation. AI is genuinely good at the math underneath inventory forecasting. Exponential smoothing. Seasonality decomposition. Multi-variate sell-through projection. And genuinely bad at the judgment that wraps the math.
Most replenishment errors are not math errors. They are context errors. The model did not know about the upcoming promotion. The model did not know the brand was holding back launch volume for a competitor’s stockout. The model did not know that the prior Q4 was an anomaly. The inventory forecast workflow handles these failure modes by keeping humans in the context loop.
The inventory forecast workflow has eight steps. AI does the math in two of them. Humans do the context in five. Rule engine catches drift in the eighth.
Step 1 (parsers), sell-through ingestion from H10, Intentwise, and Windsor
Sell-through data comes from three sources. Ingested daily by the parser layer in the ClearSight portal:
- Helium 10, daily and weekly unit-sales pulls per ASIN, alongside the Inventory Protector readouts
- Intentwise, the daily account snapshot, which includes sell-through plus advertising-driven unit lift and returns rate
- Windsor, the secondary Amazon seller/ads feed used as a cross-reference. Also the primary source for accounts not yet fully on Intentwise.
The parsers normalize each source’s unit and date conventions into a single schema. Daily, weekly, and trailing-90-day rolling windows are computed at ingest time. The strategist sees one merged dataset. The source-of-truth reconciliation is handled in the parser layer. This is the input the inventory forecast workflow runs on.
Step 2 (rule engine), weeks-of-cover thresholds and IPI scoring
The inventory_rules module runs daily against the merged sell-through dataset. The rules surface:
- Weeks-of-cover below 4 on any top-5 revenue SKU, flagged as a critical replenishment risk
- Weeks-of-cover below 6 on any top-15 SKU, flagged as a high-priority replenishment risk
- Weeks-of-cover above 16 on any SKU, flagged as an overstock risk, particularly relevant when FBA aged-inventory surcharges apply
- IPI score trending below 500, flagged as a capacity-cap risk
- Sell-through velocity deviation from the trailing-90-day baseline by more than 30% in either direction, flagged for context investigation
These flags route to Slack within an hour of detection. The inventory forecast workflow does not allow a top-5 SKU to drop below four weeks of cover without the strategist knowing about it within the same business day.
The reason is straightforward. By the time the brand’s accounting team notices a stockout in the books, the Buy Box recovery cycle is already running. The cost is locked in.
Step 3 (human), seasonality and brand-context calibration
Before any forecast runs, the distribution strategist annotates the trailing-12-month sell-through data with the context that affects future periods. The annotations cluster in five categories:
- Promotional history, every Lightning Deal, Best Deal, Prime Day, Black Friday Cyber Monday week, and brand-internal promotion that ran in the trailing year
- Out-of-stock periods, windows where sell-through was artificially low because the SKU was not available
- Competitor events, stockouts at major competitors that pulled demand toward the brand, or competitor launches that pulled demand away
- Listing changes, major A+ rebuilds, price changes, image refreshes, anything that may have shifted the conversion baseline
- Inventory capacity decisions, periods where the brand deliberately throttled sell-through because of upstream supply constraints
This annotation step is human work and the most underrated piece of the inventory forecast workflow. AI cannot do it from the sell-through data alone. The events are not in the dataset. The brand’s marketing calendar, the competitor’s stockout history, and the brand’s supply-chain constraints are external to Amazon’s data feeds.
The strategist sources this context from the brand directly during the monthly forecast review. This is the step the ai-only Amazon agency model entirely lacks. Same shape problem we wrote about in the offshore VA + ChatGPT post.
Step 4 (AI compile), forecast modeling
With the annotated history loaded, the forecast modeling step runs. The model is Claude with explicit instructions. Use exponential smoothing with multiplicative seasonality. Use Holt-Winters where the SKU has enough history. Produce a 90-day, 180-day, and 360-day forecast per SKU. With confidence intervals.
The instructions are specific. The model is told to ignore promotional periods unless the strategist has flagged a similar promotion for the forecast window. The model is told to weight the trailing 90 days more heavily than the prior periods, except where seasonality demands otherwise. The model is told to flag any forecast where the implied unit demand is more than 2x the trailing-90-day baseline. These get human review before the forecast is accepted.
The forecast output is a structured table per SKU. Expected units. Lower-bound units (P10). Upper-bound units (P90). The assumptions the model used to produce each estimate. The strategist reviews the assumptions, not just the numbers. This is the math half of the inventory forecast workflow. The judgment half lives in Steps 3 and 5.
Step 5 (human), replenishment PO sizing
The strategist takes the forecast. Applies the brand’s inventory policy. Safety stock target. Lead-time variance. Capacity cap headroom. And sizes the replenishment PO.
This is where the inventory forecast workflow keeps human judgment as the binding constraint. PO sizing is the decision the brand pays for if it is wrong.
The strategist is weighing the forecast (math) against the brand’s cash position, the supplier’s lead-time variability, the FBA capacity cap on the SKU, the seasonality of the brand’s category, and any upcoming promotional commitments. The model produces a number. The strategist produces a PO. The two are not the same.
Roughly 65% of the time, the strategist sizes the PO within 10% of the model’s central estimate. Roughly 25% of the time, the strategist sizes more conservatively. Usually because of cash-position or capacity-cap constraints. Roughly 10% of the time, the strategist sizes more aggressively. Usually because the brand has a known event in the forecast window that the model could not see.
Step 6 (rule engine), lead-time and capacity-cap validation
Before the strategist ships the PO, the inventory_rules module validates the proposed quantity. Against the brand’s current FBA capacity cap and the supplier’s documented lead-time variance. Any PO that would exceed the cap, or that would land too late given the supplier’s lead time, is flagged for revision.
This catch is what the ai-only Amazon agency model lacks. The model produces a forecast. The agency sends the brand a recommended PO size. The agency does not check whether the PO can physically fit in FBA when it lands. Or whether the supplier’s known lead-time variance gives it a meaningful chance of arriving on time. The brand discovers the gap when their inventory hits the cap and the shipment is rejected. The inventory forecast workflow catches this before the PO is sent.
Step 7 (human), PO ship and FBA shipment plan
The strategist ships the PO to the brand’s supply chain team. With the replenishment plan, the FBA shipment-plan timeline, and the Buy Box defense plan for the period leading up to the new inventory landing in the network. The PO is logged in ClickUp with the forecast assumptions and the success criteria for the post-action review.
Step 8 (rule engine), forecast-vs-actual MAPE tracking
Every 30 days, the inventory_rules module computes the Mean Absolute Percentage Error (MAPE) of the prior month’s forecast against actual sell-through. The trailing-12-month MAPE on the active client book runs about 11-15% on top-5 revenue SKUs. And 18-24% on the long-tail.
The 11-15% figure is competitive with what the best stand-alone forecasting tools claim. With the difference that the ClearSight numbers are net of human PO-sizing adjustments. That is the version that actually matters. The human-adjusted PO is what gets shipped. Pure-AI tools cannot match this because they skip the Step 5 human in the inventory forecast workflow.
Forecasts that miss by more than 25% on a top-5 SKU trigger a strategist review. The most common root cause: the strategist missed a competitor event in the seasonality annotation step. The forecast inherited the gap. The next most common: an LLM hallucination on a less-common seasonal pattern that the strategist did not catch on review. Both are caught by Step 8 before the next forecast cycle runs.
Why pure-AI forecasting overcommits Q4 every year
The structural reason a pure-AI forecast overcommits Q4 is that the model fits to the trailing-12-month average. It assumes Q4 will be a repeat of the prior Q4.
The strategist knows that Q4 demand is not stable year-over-year. Pull-forward effects from earlier Black Friday timing. Prime Day cannibalization. Category-level shifts in consumer behavior. The prior Q4 is often the worst possible single-year base for a forecast.
The inventory forecast workflow handles this by having the strategist explicitly flag the prior Q4 as a “context-anomaly” period during the annotation step. Which down-weights it in the model’s projection.
Pure-AI forecasting tools do not have this annotation step. They treat the prior Q4 as canonical history. The result is the predictable Q4 overcommit pattern we see across brands running on those tools. The inventory forecast workflow’s annotation discipline is the single biggest reason our forecasts beat theirs in MAPE.
The stack-reference post earlier in this series describes the broader architecture. The ads workflow post shows the same hybrid shape applied to PPC. The catalog workflow post applies it to listings. This post is the inventory forecast workflow specific instance.
The next and final post in this series, the results recap, walks through 12 months of what this stack has produced across the book. The specific numbers on TACoS, ROAS, ranking, and the hours we did not spend.
Reviewed by the Amazon Growth Team.
The Hybrid Stack, 10-post series. You are reading post 9 of 10.
Black-hat track, three archetypes of the ai-only Amazon agency:
- The three black-hat shapes of the 2026 Amazon agency
- Why offshore VA + ChatGPT shops are the most expensive cheap option
- Why AI listing-builder SaaS can’t get a 7-figure brand to actually index
- Why AI-first consultants who never log into Seller Central miss the work that moves money
White-hat track, the ClearSight hybrid stack:
- The ClearSight intelligence layer, stack reference
- Catalog AI/human workflow
- Ads AI/human workflow
- Creative AI/human workflow
- Inventory AI/human workflow (you are here)
- 12 months of the hybrid stack, results recap
← Previous: Creative AI/human workflow | Next: 12 months of the hybrid stack, results recap →