FOXNUT AI

Definition

Can AI do inventory management and replenishment?

A documented rule, with or without AI, can calculate when to reorder and how much. Keep a person on the step that turns it into a real purchase order.

By the Foxnut team · Updated

Can AI do inventory management?

A documented rule, with or without AI, can do the calculating part of inventory management, which is working out when to reorder and how much. The step that should stay reviewed is the one that turns a calculated reorder into a real purchase order.

Foxnut AI is a senior-only practice in Bengaluru and Paris. We built a supply-chain system for a D2C cosmetics brand in India that generates purchase orders from stock thresholds. There’s a link to talk to us at the end of this article. Let’s first work out what your reorder rule is actually reading.

Inventory replenishment is the decision a business makes many times a day without usually calling it a decision. It looks at what is on the shelf right now, what is already on order, how long a vendor takes to deliver and how much of an item moves in a normal week. Then it decides whether a purchase order goes out today, and for how much.

Replenishment is not the forecast that estimates how much of an item will sell. That is a separate problem, solved or unsolved before replenishment ever runs. It is also not the check that an arriving invoice matches what was ordered and what arrived, which happens downstream of a purchase order that already exists. Replenishment sits between the two.

The mechanics are largely the same whether a business calls the method min-max, reorder point or demand-driven material requirements planning. A documented buffer methodology sets a minimum quantity, below which stock is not meant to fall, and a maximum quantity, up to which a replenishment order restocks it. A reorder point between the two triggers the order.

What separates a working rule from a wrong one is rarely the arithmetic, which one vendor’s documentation lays out down to the equation. It is whether the two numbers the arithmetic depends on, the stock on hand and the lead time required, can be trusted the moment the rule reads them. Getting that trust right, and knowing when it has quietly broken, is most of the transfer this studio structures a replenishment build around.

What does a reorder rule actually read?

A reorder rule reads three inputs, and each fails in a different way.

InputWhat it isWhere it usually goes wrong
Stock on handA recorded quantity available or already committed, not a physical countDiverges from the physical shelf whenever a receipt, a pick, or an adjustment is not posted the moment it happens
Lead timeThe time between placing an order and having usable stock, not a vendor’s quoted delivery timeThe quoted lead time and the time actually required to reach a stocked point in the chain are rarely the same number
Vendor termsThe minimum order quantity and the interval a vendor allows between ordersSet once at onboarding and rarely revisited, so a rule can be arithmetically correct and still order in a size or on a schedule the vendor no longer honors

The first two rows get their own sections below, because each goes wrong in a way no generic accuracy claim covers.

What is stock reconciliation, and why does a reorder rule depend on it?

A reorder rule never looks at a shelf. It looks at a number a system already holds, which is why stock reconciliation, comparing the recorded quantity with the counted quantity, decides whether the rule can be trusted.

Every documented method for producing that number is explicit about which transactions count toward it. One demand-driven buffer methodology calculates average daily usage, the figure both its yellow zone and its red zone are built from. It only counts transactions with a status of on order, reserved ordered, reserved physical, picked, deducted or sold, dated within a chosen backward-looking period, and it excludes warehouse work, quarantine, sales quotations and statements.

That list is long and specific because the calculation is not reading a shelf. It is reading what a series of prior transactions, correctly posted and correctly statused, says should be there. A receipt posted a day late, a pick recorded against the wrong location or a quarantine hold that never clears is invisible to the arithmetic. From the rule’s point of view it looks the same as stock that genuinely is not there.

The same documentation flags a narrower version of this problem. When a business reserves batch-tracked inventory below the location level and does not allow batch numbers to mix at a picking location, the replenishment process can generate work that conflicts with itself. Only items reserved above the location level carry the batch information the put step needs. This caution sits inside the feature’s own setup instructions, so the vendor already treats the problem as real.

Classifying, matching or reconciling ledgers the way a finance-side process would does not solve this. The fix is the discipline a warehouse already owes itself: post the receipt when the pallet arrives, resolve the quarantine hold before it ages into a phantom balance, and treat any gap between the recorded quantity and the counted quantity as a defect in the input a reorder rule is about to act on. It is not a rounding error to write off at the next physical count.

How does lead time tracking change the reorder point?

Lead time tracking changes the reorder point because a reorder point is demand during lead time, plus a margin for how much that demand and that lead time vary. One documented buffer methodology separates the quoted, cumulative lead time a vendor states from what it calls the decoupled lead time: the time actually required to reach a point in the chain that is itself kept in stock. Its worked example cuts a 21-day cumulative lead time down to five decoupled days, once a decoupling point already holds inventory partway through the chain.

Using the vendor’s quoted number where the decoupled number belongs does not fail safe. It oversizes every buffer downstream of the substitution, and ties up cash in stock the real, shorter lead time never needed.

Lead time is also not applied evenly across items. The same methodology gives each item a lead time factor between 0 and 1, lower for a longer lead time and higher for a shorter one, and a separate variability factor between 0 and 1 for how much demand swings. The vendor documents these as recommended ranges from an outside body, the Demand Driven Institute, and not as a fixed formula: roughly 0.20 to 0.40 for a long lead time and 0.61 to 1.00 for a short one, with a mirrored pair of bands for low versus high variability.

Two businesses computing what looks like the same reorder point can land on materially different buffer sizes, because the factor each chose inside that range was a judgment call and not a measurement. Automating the arithmetic does not remove that judgment. It moves it to whoever sets the factor, and the factor needs revisiting on the same schedule as the lead time itself, not left at whatever value onboarding assigned.

What does the decision to automate replenishment turn on?

The decision turns on six structural dimensions, and inventory replenishment reads differently from its moving-goods relatives on nearly all of them. The thing being produced is a live threshold, computed from data the business already holds and acted on the moment it is crossed. It is not a forecast of the future or a comparison of documents already written.

Replenishment: process profile
DimensionWhat it reads on inventory replenishmentSource
Exception varianceNot a stockout versus a full shelf, but which of three zones the recorded stock sits in. One documented buffer methodology defines a red zone below a computed minimum, a yellow zone between the minimum and the reorder point, and a green zone between the reorder point and a computed maximum, and a reorder fires the moment the reorder point is crossed, not at the point stock reaches zero. The height of each zone is itself computed, not fixed: the yellow zone is average daily usage times the decoupled lead time, and the red zone pads that base by a separately chosen lead time factor and variability factor. An item's exception rate is a property of how those factors were set for it, not a fixed characteristic of the item, which is a different claim from a document either matching or not matching.Dynamics 365, buffer profile and levels documentation
VolumeNot SKU count and not units shipped, but the number of SKU-location-vendor combinations a business is willing to let a rule decide on without a person looking first. One vendor's own account of why demand-driven planning exists states plainly that a conventional planning run often gives planners thousands of actions with no way to know what to focus on, and a transaction-set standard built for exactly this exchange is written to run across many distribution centers, warehouses or retail outlets at once, not one location at a time. The payback condition follows as a condition rather than a count: automation pays where the number of buffered items, multiplied by the locations each is held at, already exceeds what a person can review on the schedule the business needs, and it is a loss where a handful of SKUs at one location would let a person just look.Dynamics 365 DDMRP overview; ASC X12 852 transaction set reference
Cost of an errorAsymmetric, and only one direction is captured anywhere a business is required to look. An accounting standard governing inventory measured other than by LIFO or the retail method requires it to be carried at the lower of cost and net realizable value, and states that when net realizable value falls below cost, the difference is recognized as a loss in the period it occurs, whether the cause is damage, obsolescence, or simply a change in price levels - exactly the fate of stock a replenishment rule ordered too much of. A stockout has no equivalent entry anywhere in the same framework: a sale that did not happen because the shelf was empty is not a figure any accounting standard asks a business to recognize, it is revenue that never appears, absorbed silently by whichever function loses the customer. The business that overstocks has a number forced onto its books; the business that stocks out has nothing forced onto its books at all, which is not the same as the stockout being cheaper.FASB ASU 2015-11, ASC 330-10-35-1B
ReversibilityReversible up to a named step, and the vendor's own warning marks where that step is. Before a replenishment rule's output is committed, it exists only as a planned order, and a documented calculation screen offers a button that discards some or all of the calculated buffer values outright, at no cost. Converting that planned order into an actual purchase, transfer or production order is called firming, and a firmed order is also described as a released or open order, the language of something that has left the plan. The same documentation carries an explicit warning against automating this step carelessly: uncritical firming of planned orders can create massive numbers of unwanted purchase, transfer and production orders, and its own recommended practice is to always preview the result before running an automated firming job rather than trust the filter criteria blind. The window for a free correction closes at firming, not at the moment the buffer was calculated or the reorder point was crossed.Dynamics 365 planned-order-firming documentation; buffer profile and levels documentation
Regulatory exposureIndirect, and only on the overstock half of the decision. No rule specifies how a business should compute a reorder point, choose a lead time factor, or decide between min-max and demand-driven buffering; that choice is left entirely to the business. What eventually tests the decision is downstream and financial rather than operational: the same accounting standard that forces a write-down on inventory whose net realizable value has fallen below cost is, by its own text, a measurement rule for a balance-sheet figure, applied after the fact to whatever the replenishment rule already ordered. It says nothing about whether the order should have been placed, only what the business must now say the resulting stock is worth. The stockout side of the same decision has no equivalent examiner at all: nothing tests, audits or requires disclosure of an order that should have been placed and was not.FASB ASU 2015-11, ASC 330-10-35-1B
Vendor market maturitySettled at the data layer and comparatively unsettled at the decision layer directly above it. The transaction standard that carries inventory and sales activity between trading partners so that one side can propose a replenishment quantity for the other has been a published, named ANSI standard for decades, built to run across distribution centers, warehouses and retail outlets as a matter of course. The reorder logic that consumes that data is not standardized the same way: the buffer factors a documented planning methodology asks a business to set are published as recommended ranges from an outside institute, not as fixed values the software supplies, which is the vendor's own admission that the judgment call inside the arithmetic has not been reduced to a solved default. A market can standardize how the data moves long before it standardizes what a business should do with it.ASC X12 852 transaction set reference; buffer profile and levels documentation

Two of those rows carry the argument. The reversibility row says the free correction window is real but narrow: everything before firming can be discarded at no cost, and the vendor warns about the step that follows because uncritical automation of it is a documented failure.

The cost-of-error row says the two directions of a bad reorder are not treated the same way anywhere outside the business. Overstock eventually forces a number onto the books, and a stockout forces nothing onto them at all. That explains why so many replenishment projects quietly optimise against overstock and drift toward stockouts without anyone deciding to.

The vendor-maturity row should temper expectations before anything else does. A standard that has moved replenishment data between trading partners for decades is not a settled answer for how large a buffer should be. The data plumbing is mature. The judgment on top of it, expressed as a range and not a number, is not.

What is our position on inventory replenishment?

Our position is in the two records below, and both need more careful reading than the rest of this page. Both trace to one source only, our founder’s own account of a client engagement. That is an operating record, not a cited market statistic.

Named directly

We built a supply-chain management system for a D2C cosmetics brand in India. It generates purchase orders from stock thresholds across vendor terms and lead times. The engagement's own account of scope confirms inventory replenishment and purchase orders as work it actually did

Method The founder's observed pattern across the systems the studio has built and handed over; stated as operating experience, not a cited market statisticSample One client engagement, confirmed against the process list in this pillarPeriod As of August 2026Source Foxnut AI engagement records (not externally retrievable)

Escalates, does not guess

Our founder's stated design principle for this system: it takes reorder decisions on rules it can learn and append from past and ongoing ordering behaviour, and it escalates to a person when it cannot make the call, instead of guessing

Method The founder's account of the system's own design, given as the studio's position rather than a benchmarked findingSample One client engagementPeriod As of August 2026Source Foxnut AI engagement records (not externally retrievable)

What the studio's engagement records support, and what they do not
NumberWhat it measuresPeriodSource
Named directlyWe built a supply-chain management system for a D2C cosmetics brand in India. It generates purchase orders from stock thresholds across vendor terms and lead times. The engagement's own account of scope confirms inventory replenishment and purchase orders as work it actually didAs of August 2026Foxnut AI engagement records (not externally retrievable)
Escalates, does not guessOur founder's stated design principle for this system: it takes reorder decisions on rules it can learn and append from past and ongoing ordering behaviour, and it escalates to a person when it cannot make the call, instead of guessingAs of August 2026Foxnut AI engagement records (not externally retrievable)

The “Named directly” record states which processes the engagement reached and nothing more: not exception counts, build time, cost or which client. The “Escalates, does not guess” record describes the design principle our founder attributes to the system, not a measured accuracy rate. Everything else about how the system performed stays unstated, because our founder did not supply it and this page does not infer it.

What do people usually get wrong about inventory management automation?

The first error is buying a reorder engine before checking whether the stock number it will read can be trusted. A rule built on demand-driven or min-max logic is only as good as the transaction data behind it. A business with unposted receipts, uncleared quarantine holds or picks recorded against the wrong location will get a confidently wrong reorder from perfectly correct software.

The second error is treating a vendor’s quoted lead time as the number the buffer needs. The two numbers are related but not identical. Substituting one for the other either strands a business without stock during the gap, or ties up cash in a buffer sized for a longer wait than the real one.

The third error is setting the lead time and variability factors once, at onboarding, and never touching them again. Both are published as ranges because they are meant to move as a vendor’s actual performance and an item’s actual demand pattern become known. A factor frozen at its first guess drifts further from reality every month it goes unreviewed.

The fourth error is automating the firming step with the same confidence as the calculation step. The arithmetic that proposes a planned order and the decision to convert it into a real, released purchase order carry different consequences. The vendor’s own warning against uncritical, unreviewed firming deserves to be taken at face value.

How should you automate inventory replenishment?

Automate the calculation and keep a reviewed step at firming, with the review weighted toward whichever of the three inputs you trust least: the recorded stock figure, the lead time behind it, or the vendor terms wrapped around both. None of the six dimensions argue against a documented buffer methodology. Several argue for one, since a rule that recomputes a reorder point from current data every cycle is a real improvement over a static number nobody revisits.

What they argue against is skipping the two checks that decide whether the automation is reading anything real. The first is whether the stock figure has been reconciled recently enough to trust. The second is whether the lead time feeding the buffer is the actual time to a usable point in the chain, not a vendor’s quoted number from onboarding.

The reason is structural and not a comment on any particular product. The exception row says a reorder is triggered by a computed zone boundary, not by a shelf going empty, so the boundary is only as honest as the factors that built it. The reversibility row says the free correction window closes at firming, and the vendor warns about that step because uncritical automation of it is a documented failure mode.

The regulatory row says nothing outside the business tests whether a reorder should have happened at all. Only the overstock half of a bad decision ever surfaces on a balance sheet, so the discipline of getting this right has to come from inside the business.

For a team weighing this, the first count is not a demonstration of the arithmetic. Pull the reorder history for one buffered item over the last two review cycles, and check three things: whether the recorded stock figure matched a physical count taken around the same time, whether the lead time used matched how long the last few orders for that item actually took to arrive, and whether the buffer factors have been touched since the item was first set up. The answer says more about whether a replenishment build is ready than any pilot would.

Who can build an inventory replenishment system with you?

We know nobody enjoys finding out the stock number was wrong after the order has gone out. The system we built for a D2C cosmetics brand in India generates purchase orders across vendor terms and lead times, and it is designed to escalate to a person when it cannot make the call. It does not guess.

Whoever signs your proposal at Foxnut AI is the person doing this work, because we have one tier, and we designed, made and launched our own products before advising anyone. The first step is a free expert audit of your setup by a senior practitioner, without an RFP or a pitch deck. You can read the consulting behind this work first. Bring that check to a conversation.

Sources

  1. Microsoft, 'Demand Driven Material Requirements Planning (DDMRP) overview', Dynamics 365 Supply Chain Management product documentation - the description of DDMRP as a planning methodology based on decoupling supply and demand, the statement that companies increasingly experience stockouts or overstocks because they do not know how much inventory to stock, and the statement that conventional MRP tools often give planners thousands of actions with no way to know what to focus onRetrieved
  2. Microsoft, 'Buffer profile and levels', Dynamics 365 Supply Chain Management product documentation - the definition of the red, yellow and green buffer zones and the minimum quantity, reorder point and maximum quantity that bound them; the yellow-zone and red-zone equations (average daily usage times decoupled lead time, padded by a lead time factor and a variability factor); the three methods of calculating average daily usage (past, forward, blended) and the specific transaction statuses each one counts; and the Demand Driven Institute's recommended factor ranges by lead time length and demand variabilityRetrieved
  3. Microsoft, 'Set up a min-max replenishment process', Dynamics 365 Supply Chain Management product documentation - the definition of min-max replenishment (when inventory falls below the minimum level, the system creates work to replenish the location) and the documented caution that replenishing batch-tracked items under certain reservation-hierarchy and mixing settings can create a work conflictRetrieved
  4. Microsoft, 'Firm planned orders', Dynamics 365 Supply Chain Management product documentation - the definition of firming as the step that converts a planned order into an actual purchase, transfer or production order (also described as a released or open order); the three firming methods (manual, auto-firming within a time fence, query-based firming); and the product's own warning that uncritical firming of planned orders can create massive numbers of unwanted orders, with its recommended practice of always previewing the result before running an automated firming jobRetrieved
  5. Financial Accounting Standards Board, Accounting Standards Update No. 2015-11, 'Inventory (Topic 330): Simplifying the Measurement of Inventory', July 2015 - paragraph 330-10-35-1B's requirement that inventory measured using a method other than LIFO or the retail inventory method be measured at the lower of cost and net realizable value, with the difference recognized as a loss in earnings in the period it occurs; the Master Glossary definition of net realizable value; and the scope note excluding inventory measured using LIFO or the retail inventory method. Read as text extracted from the FASB's own PDFRetrieved
  6. ASC X12, '852 - Product Activity Data' transaction set reference - the transaction set's own statement of purpose, that it conveys inventory, sales and other product activity information, and that this data enables a trading partner to plan and ship or propose inventory replenishment quantities for distribution centers, warehouses or retail outletsRetrieved

Foxnut AI takes on work like this from Bengaluru and Paris. If you want to see how we do it before you talk to anyone, here is how an AI engagement is scoped and priced.