Safety Stock for Lumpy Demand: S&S, Promos, Agentic Commerce
Variances add, standard deviations do not. How to size safety stock when demand arrives in lumps from Subscribe and Save, promos, and agentic commerce.
Andrew Erickson is the founder of Inventory Hero. He has spent years working with Amazon FBA sellers on demand forecasting, restock planning, and the cash flow side of running a private-label brand. Inventory Hero exists because every spreadsheet-based inventory system he tried eventually broke — usually right before Q4.
How do you calculate safety stock when demand arrives in lumps?
Split demand into components, buffer only the ones whose timing you do not know, and combine them as the square root of the sum of their variances. Standard deviations do not add. A lump whose date you DO know, like a weekly Subscribe and Save shipment or a scheduled wholesale PO, belongs in cycle demand as a planned draw, not in the buffer, because a naive daily STDEV reads a predictable lump as volatility and buys you insurance you do not need.
Why does a small demand component change my safety stock so much?
Safety stock scales with the standard deviation of demand over lead time, not with average demand, so a component's danger has almost nothing to do with its share of your sales velocity. A slice averaging half a unit a day that arrives 10 units at a time carries far more variance per unit of mean than a steady 30 units a day, and variance is what the buffer is priced against.
What is agentic commerce for Amazon sellers?
Agentic commerce is when a shopper's AI assistant places the order instead of the shopper clicking buy. On Amazon that means Alexa for Shopping features like Auto Buy, which purchases at a target price, and Scheduled Actions, which handle recurring tasks. It is a different topic from using AI to forecast your own demand: one creates demand, the other predicts it.
Can I see which of my Amazon orders were placed by an AI agent?
No. Amazon publishes no Seller Central report that flags an order as assistant-placed or agent-placed, and Business Reports has no traffic source that isolates it. The one adjacent thing you can see is Subscribe and Save: SnS-tagged orders in Manage Orders and active subscription counts on the program page. Treat agent volume as unmeasurable and size your buffer from the demand shape you can observe.
Should I stop using a repricer because of Auto Buy?
No. Set a floor you would be happy to sell your entire remaining runway at, which is worth doing even if agent volume on your ASINs is zero. Auto Buy is reported to check price on a roughly 30 minute cycle, though that figure comes from trade press rather than Amazon, so a short price test could in principle fire queued orders. The general rule stands regardless: avoid base-price cuts on any SKU with less than lead time plus safety days of cover.
If any part of your demand arrives in lumps, the standard deviation you pull off a spreadsheet is buying you the wrong amount of insurance. The fix is two rules. Variances add, standard deviations do not, so demand components combine as sigma_total = square root of (sigma_a squared + sigma_b squared), and a small, spiky slice can dominate the result. And a lump whose date you know is cycle stock, not safety stock, so blending it into one average sigma produces a bigger number that is still wrong.
This matters for Subscribe and Save shipments, wholesale POs, promo clusters, viral spikes, and, at the newest and smallest end, orders placed by a shopper's AI agent.
Safety stock = Z x sigma x square root of lead time
Z is the service-level factor (1.65 for 95 percent), sigma is the standard deviation of daily demand, and lead time is in days. If your lead time swings too, use the combined form in lead time variability.
Everything hangs on sigma, and sigma is where blending goes wrong. Take a SKU with a 45 day lead time doing about 40 units a day, split three ways:
Component
Mean units/day
Daily sigma
Where the number comes from
Measured or assumed?
Browse-driven baseline
30
6.0
STDEV of daily units ordered, Business Reports > Detail Page Sales and Traffic by Child Item
Measured
Recurring, known date
10
24.5
Population sigma of a perfectly regular 70 / 0 / 0 / 0 / 0 / 0 / 0 week
Assumed pattern
Clustered, unknown date
0.5
2.2
Scenario: a 5 percent daily chance of 10 units landing at once
Assumed scenario
Only the first row is data. The other two are constructed, and the last column is there so you do not mistake them for a report you can pull. The recurring row is the headline anyway: it is a quarter of the average and carries a sigma four times the baseline, purely because it is all-or-nothing by day, and it is an idealization (a real Subscribe and Save week is not exactly 70 / 0 / 0 / 0 / 0 / 0 / 0). The clustered row is modeled as a Bernoulli spike, so sigma = Q x square root of (p x (1 - p)) = 10 x square root of 0.0475 = 2.2 units/day on a mean of half a unit.
Now size the buffer three ways, with square root of 45 = 6.71:
Approach
Daily sigma
Safety stock
Everything blended together, no cluster in the sample
sqrt(6.0² + 24.5²) = 25.2
1.65 x 25.2 x 6.71 = 279 units
Browse only, ignoring cluster exposure entirely
6.0
1.65 x 6.0 x 6.71 = 66 units
Two components: browse + cluster, recurring planned rather than buffered
sqrt(6.0² + 2.2²) = 6.4
1.65 x 6.4 x 6.71 = 71 units
That first row is a recombination of the two component sigmas, not a number you would get by running STDEV() on real combined daily orders. It stands in for the blending mistake; a real spreadsheet pull on messy data would land somewhere near it, not on it.
The blended 279 is four times the honest answer and still wrong. It is insurance against a cadence you could have put on a calendar, while carrying nothing sized for the cluster your sample never contained. Over a 45 day lead time the recurring slice delivers six or seven lumps and you know which days, so it belongs in cycle demand (45 x 10 = 450 units of planned draw), not in the buffer. That is the same instinct as pulling seasonality out before you smooth, covered in is my product seasonal and the inventory forecasting methods comparison.
Be clear about what that third row does: it sets the recurring component's buffer contribution to exactly zero, not merely "less." That is only defensible when the dates are genuinely known, which is the case for a Subscribe and Save shipment schedule or a confirmed wholesale PO. If the dates are approximate, the honest move is a small residual sigma for timing jitter rather than the full 24.5. A lump that can land a day early is a one-day exposure, not a week of volatility. Say a residual of 5 for a lump that wanders by a day: sqrt(6.0² + 5.0² + 2.2²) = 8.1, which is 90 units. That is the shape of the correction, and it is nowhere near 279.
The cluster slice does belong in the buffer, and it moves 66 to 71 on a component worth half a unit a day out of 40. That is a real effect and a modest one. Push the scenario both ways and the honest range is narrow: a 2 percent chance of 5 units gives 67, a 10 percent chance of 30 units gives 120. Run your own numbers in the safety stock calculator, or see the safety stock definition for the short version.
In rough order of how much volume it moves for a typical private-label seller today:
Subscribe and Save. The largest recurring channel and the one you can actually see. Known cadence, known dates: cycle stock.
Wholesale or B2B POs. Large, infrequent, and usually known days in advance once you have the relationship. Cycle stock when you know, buffer when you do not.
Promotions and deal events. You set the date, so plan the lump. What you cannot plan is the size, which is genuine buffer territory.
Viral or off-platform spikes. Pure variance with no known date. This is what safety stock exists for.
Agent-placed orders. Newest and smallest. Worth understanding mechanically, not worth re-engineering planning around.
Note that four of those five have nothing to do with AI. The method is the point.
Two things, roughly nine months apart, and the timeline gets misreported constantly.
Auto Buy launched in November 2025 under the Rufus brand. A US Prime member tells the assistant to buy an item when it hits a target price or discount level, the assistant monitors price on a roughly 30 minute cycle, and when the condition is met it places the order using the shopper's default payment and shipping details, with a 24 hour cancellation window before shipment.34 Trade coverage consistently reports it as limited to items Fulfilled by Amazon, one unit per request and one active request per item, with coupons and promotional discounts excluded from the trigger.56
That per-request limit is the number that keeps this in proportion. A ten unit Auto Buy cluster on one ASIN requires ten separate Prime shoppers holding queued requests that all trip on the same price move. On a 40 unit a day SKU that is about six hours of demand, which is why the scenario above uses 10 and not 60.
Scheduled Actions arrived on May 13, 2026, when Amazon merged Rufus into Alexa for Shopping and retired the Rufus name.[^dc360]1 A shopper sets a recurring task like restocking pet food or detergent each month. Amazon's own examples are mostly cart adds and alerts rather than completed purchases.2 A cart add is intent, not an order, which is why Scheduled Actions is a weaker demand signal than a Subscribe and Save subscription today.
So one feature places orders and the other mostly stages them. Sellers who already track Rufus and how it surfaces products are looking at the discovery side of the same assistant; this is the transaction side. None of it is AI demand forecasting, which lives on your side of the transaction and predicts the signal rather than creating it.
Agentic commerce
Purchases initiated and completed by a shopper's AI assistant acting on a standing instruction, rather than by the shopper actively browsing and checking out at that moment.
Less than the headlines suggest, and it is standard hygiene rather than a new discipline. A price cut on a SKU with queued price-triggered requests pulls demand forward at the price you least wanted to sell at. That is a one-shot event, not a feedback loop: the orders fire, and they do not move your price again unless your repricer reacts to velocity or inventory, which most do not, because they track competitor Buy Box price. Mechanically it is identical to any promo that clears your cover faster than expected.
Two guardrails that cost nothing and are worth setting even at zero agent volume:
Set a repricer floor you would happily sell your entire remaining runway at. Not break-even. The price at which clearing every unit you have left is a good outcome.
Do not price-test downward on a SKU with less than lead time plus safety days of cover. This is true of any discount mechanism and does not depend on how Auto Buy works.
There is a third, weaker guardrail worth knowing but not worth restructuring around: Auto Buy is reported to trigger on base price with coupons and promotional discounts excluded, which would make a coupon a safer way to move effective price than a base-price cut when cover is thin.56 Two independent trade sources say so and Amazon's own help page reportedly does too, but we could not retrieve that page directly. Verify it before you build a discounting policy on it.
No, and saying so plainly is more useful than pretending otherwise. Amazon publishes no Seller Central report that labels an order as assistant-placed, and Business Reports has no traffic source that isolates it.
The one adjacent thing you can see is Subscribe and Save. Be careful which surface you point at: Amazon deprecated the Subscribe and Save Forecast and Performance reports (empty responses from July 25, 2025, removed from SP-API on December 11, 2025), so older advice to pull them is stale.7 What remains is active subscription counts on the Subscribe and Save program page and SnS-tagged orders in Manage Orders. Those are your real recurring-demand numbers. Scheduled Actions is not measurable at all.
Split before you smooth. Pull anything with a known date out of the sample before you take a STDEV. Plan it as cycle demand. This is where most of the money is.
Combine the rest with root-sum-square, not by averaging. One blended sigma over mixed demand shapes is the error that produced 279 above.
Set the repricer floor. Ten minutes, and it protects you from ordinary repricing races regardless of agents.
Log your own price changes against daily units. You know when you moved price, so this is the one cluster effect you can attribute. Pull daily units ordered by child ASIN from Business Reports > Detail Page Sales and Traffic by Child Item and put your repricer's change log beside it. Orders with no matching jump in sessions is a hint, not a signature: Subscribe and Save shipments, Buy Again taps, sessions from a prior day, multi-unit orders, and Business Reports' own session-attribution gaps all produce the same pattern. Treat it as a reason to look, never as proof.
Check the boring explanation first. Same-weekday clustering is almost always Subscribe and Save or a weekly promo rhythm, not an agent.
Do not rebuild your forecasting for agentic commerce. A model tuned for an unmeasurable channel is worse than the simple one you have.
Agentic commerce is early, no seller can measure the volume through it, and most of the circulating numbers about its scale do not survive a second source. But the statistics it made people ask about are old and durable: variances add, a low-mean high-sigma component can dominate a buffer, and a lump with a known date belongs on the calendar rather than in the safety stock. Apply that to the Subscribe and Save and promo clusters already in your history and you will have fixed something real, whether or not a single agent ever buys from you.
Amazon, "Meet Alexa for Shopping, your personalized, agentic AI assistant on Amazon," aboutamazon.com/news/retail/alexa-for-shopping-ai-assistant (accessed August 2026). Announces the merger of Rufus and Alexa+ into Alexa for Shopping, names Scheduled Actions, Auto Buying, Price History, Shop Direct, and Buy for Me, states the rollout to all US customers "over the coming week" with no Echo device or Prime membership required for the assistant itself, and states that "Rufus helped over 300 million customers in 2025." ↩↩2↩3
Amazon, "How to use Alexa for Shopping to compare products, check price history, auto-buy items at target prices, and more," aboutamazon.com/news/retail/how-to-use-amazon-shopping-ai-assistant (accessed August 2026). Describes Scheduled Actions as recurring shopping tasks that research products and either notify the shopper or add items to the cart, and describes price-triggered actions that either alert or complete the purchase with the default payment method. ↩↩2
eMarketer, "Amazon edges deeper into agentic commerce with Rufus 'Auto Buy' function," November 14, 2025, emarketer.com/content/amazon-edges-deeper-agentic-commerce-rufus-auto-buy. Dates the Auto Buy launch to early November 2025 for US Prime members in the Amazon mobile app, notes the 24 hour cancellation window, and reports 250 million Rufus users for the year. Note the conflict: Amazon's own May 2026 announcement puts the 2025 figure at over 300 million. Both figures measure assistant usage, not agent-placed orders, and neither is a measure of agentic order volume. ↩↩2
Modern Retail, "Amazon's shopping bot Rufus can now automatically buy products for you when prices drop," November 2025, modernretail.co/technology/marketplace-briefing-amazons-shopping-bot-rufus-can-now-automatically-buy-products-for-you-when-prices-drop/. Reports that Rufus monitors prices every 30 minutes and executes the purchase automatically when the condition is met, with 24 hours to cancel before shipment, available to all US Prime members through the Amazon app. The 30 minute figure appears only in trade press; we have not seen it stated by Amazon.
Zon Wizard, "Amazon Auto Buy: the new button that is changing the rules of pricing for sellers," June 16, 2026, zonwizard.com/blog/amazon-auto-buy-the-new-button-that-is-changing-the-rules-of-pricing-for-sellers. Reports Auto Buy as US Prime members only, limited to items Fulfilled by Amazon, one unit per order and one active request per product, a 24 hour cancellation window, and no promotional discounts or coupons applied (base price only triggers the purchase). ↩↩2
My Amazon Guy, "Amazon Add to Auto Buy Button Impacts Seller Pricing," myamazonguy.com/news/amazon-add-to-auto-buy-button. Independently reports the same mechanics: "one active request per item and one unit per request," and that "promotional discounts and coupons are not applied to Auto Buy orders." Amazon's own "About Auto Buy" customer help page (amazon.com/gp/help/customer/display.html?nodeId=TsaUdPSIWqy1tZhF09) is reported to describe the same limits, but returned a 503 on every attempt to retrieve it for this article. Treat the per-request and coupon mechanics as consistently reported by trade press rather than confirmed first-hand, and check the current help page before building a pricing rule on them. ↩↩2
Be Bold Digital, "Amazon Subscribe & Save Report Deprecation," bebolddigital.com/blog/amazon-subscribe-and-save-reporting-deprecation (accessed August 2026). Reports that the Subscribe & Save Forecast Report (8 week projections) and Performance Report (4 week trailing) began returning empty API responses on July 25, 2025 and were removed from SP-API on December 11, 2025, with sellers directed to the Replenishment API v2022-11-07 or the Subscribe & Save program page, where active subscription counts remain visible; SnS-tagged orders remain visible in Manage Orders. ↩