The agents arrived at the count before the controls did
Worked example from Glock, Rekik and Syntetos for ECR Retail Loss, February 2026: the same twelve records, measured three ways. Only the first number is usually reported.
In the five weeks to 8 October 2026, three enterprise vendors moved agents closer to the execution path of warehouse and inventory work. SAP's supply-chain product organisation set out its Joule assistant portfolio — planning, manufacturing, logistics, asset operations, business network — with general availability planned for the fourth quarter and a careful formulation attached: agents trigger coordinated action “when authorized,” and leaders should be able to set what an agent may complete. Oracle shipped an AI agentic Warehouse Operations Workspace in Fusion Cloud SCM 26B, with four advisors across stock, inbound, outbound and task execution — and exactly one action that executes without a human: dynamic task assignment. Everything else is a recommendation subject to privilege and approval. On 7 September 2026, Blue Yonder added agentic returns disposition, simulating outcomes at unit level and recommending routing, receipt and resale, with returned items back in sellable stock around 25% faster.
Read those three releases together and the pattern is not capability. It is restraint: the agent may observe, reason and recommend, and a human still holds the write. Gartner has named the direction of travel — surveying 509 supply chain leaders in October 2025, it predicts that by 2031, 60% of supply chain disruptions will be resolved without human intervention, advising that autonomy expand in a controlled manner starting from low-risk decisions.
The inventory adjustment will look like exactly that: small, high-volume, routine, invisible outside the warehouse. It is also the mechanism by which an operation's only independent measurement of its own accuracy is destroyed. A count produces evidence; the posting consumes it. Automate the posting before the attribution and you have not reduced risk — you have removed the instrument that detects it.
The measurement problem is neither speculative nor small. The ECR Retail Loss programme published 9 February 2026 by Christoph Glock of TU Darmstadt, Aris Syntetos of Cardiff Business School and Yacine Rekik of emlyon — 25 interviews, verified in a workshop with 46 experts — reports that over 60% of inventory records contain inaccuracies, and that in grocery only about 35% are exactly right. Correcting records after a stock audit produced sales uplifts of 4% to 11% in test stores.
Their companion paper on metrics is the one an operator should read first. In its worked example of twelve records the signed error is +10 units, or 0.2% — a number any executive would sign off. The absolute error across those same twelve records is 300 units, roughly 18% of physical stock, and the binary error rate, the share of records where system and physical simply disagree, is 92%. One shelf, three numbers, two orders of magnitude apart. The authors are explicit: signed error can look acceptable while individual records are badly wrong, and error should be expressed as a percentage of physical stock, not system stock.
A balance that reads level tells you nothing about the weights. Net accuracy is a balance reading, and it is the figure almost every inventory scorecard carries.
The variance loop: detect, attribute, absorb
An inventory discrepancy is not one decision. It is three, taken in sequence, by different people, in different systems, and recorded with wildly different care. Socradata separates them deliberately, because the automation risk sits in one of the three and the investment sits in another. The Brazilian work on distribution-centre performance by Iuri Destro, Francielly Staudt, Karine Somensi and Carlos Taboada, published in Production in 2023, supplies the decomposition underneath: record inaccuracy splits into ghost inventory the system believes exists and hidden inventory it does not know it has, and the two have different causes, costs and cures.
The count that produces the variance. This layer is funded: cycle-count programmes, scanning discipline, RFID, vision, predictive count scheduling that targets error-prone records. It is also the layer vendors are automating first, and sensibly so. But detection only generates a signal. An operation that improves detection while leaving the next two layers alone simply finds more discrepancies it has no process to explain.
The reason code that explains the variance: miscount, mis-scan, mis-pick, receiving error, damage, theft, unit-of-measure conflict, a put-away to the wrong bin. The ECR maturity work found one retailer that had identified roughly 20 distinct root causes. This is the layer nobody funds, because attribution costs operator time and moves no stock — hence the dashed box below. It is also the only layer that turns a discrepancy into management information.
The posting that closes the variance and moves the record — and, through the goods-movement document, the ledger. This is the layer with write authority, financial consequence and audit exposure, and the one an agent reaches last and most quietly. Absorption without attribution is the failure mode: the record becomes correct, the cause stays unknown, the discrepancy recurs next quarter, and the signed error on the scorecard stays flat throughout.
Detection is instrumented, absorption has the write authority, and attribution — the only layer that feeds anything back — is the one left to an operator with a drop-down and a shift to finish.
So what: a system allowed to close its own variance will always report an accuracy number that nets to zero. The number that matters is the one that does not net.
Two metrics span the framework, and neither requires a new platform. The first is the absolute-to-signed error ratio — gross divided by net, computed on the same records over the same period. Most operations have never calculated it; where they do, it is a multiple, not a margin, and it tells an executive in one figure how much cancellation is hiding inside a reassuring number. The second is reason-code coverage: the share of posted adjustments carrying a specific, actionable cause rather than a default. KPIs before APIs is the whole argument of this edition — an agent granted write authority against an unmeasured gross error will optimise the net, because the net is what it can see.
Three loops where the posting is already close to automatic
The three patterns below are composite, anonymised operating patterns drawn from how these systems are configured in the Southern Cone; they are illustrative, not client cases, and the improvement ranges are labelled as such. Each names the systems of record, the decision loop, the human override path and the measure that would reveal the loop failing.
Multi-client 3PL, Greater Buenos Aires — the correction that is also a liability. Systems: a multi-tenant WMS holding count tasks and a per-client inventory ledger, the client's own ERP on the other side of an interface, and the billing engine that prices storage and handling. Decision loop: an operator reports a short at the pick face; the system proposes either a bin-to-bin correction, which is free, or a variance posting against that client's stock, which is a claim. In a multi-tenant warehouse the adjustment is not a record change — it is a commercial event with a counterparty. The structural risk is offsetting: a shortage in one client's stock and an overage in another's net to zero at warehouse level while two clients are each wrong. Human override: a client-facing inventory controller approves any variance above a per-client value band, and cross-client offsets are impossible by construction rather than discouraged by policy. Measure: absolute-to-signed ratio reported per client, never aggregated. Illustrative target: claim-to-close cycle time down 20–30% once reason codes are mandatory, because disputes are then argued on cause rather than quantity.
Supermarket chain, CABA — the phantom stockout the system fixes by agreeing with itself. Systems: ERP inventory management, the store replenishment engine that reorders against on-hand, and the point-of-sale stream. Decision loop: a record shows stock the shelf does not have, so replenishment does not trigger, so the item does not sell, so consumption data confirms low demand — and the forecast learns the error. Grocery is where this bites hardest: on the ECR figures only about 35% of records are exactly right, and the 4–11% uplift measured after correction is the prize sitting in that loop. Automatic negative-on-hand correction is the tempting fix and the dangerous one, because it silences the symptom without touching the cause. Human override: below a small unit band a store manager may accept an auto-correction; above it the system forces a count task before any posting, and the forecast is flagged rather than fed. Measure: binary accuracy at item-location and phantom-stockout rate, with absolute error in value reported alongside shrink.
Industrial spare-parts distributor, Southern Cone — where the materiality band has to be in value, not units. Systems: WMS bin and serial records, ERP inventory management, and the financial posting the goods movement generates in the ledger. Decision loop: a single-unit variance on a low-volume, high-value part is trivial in quantity and material in money, yet lands in the same queue as a carton of consumables. Where inventory is revalued often, the peso value of an unexplained adjustment can differ between the count and the posting — so a band defined in units is not a control at all. Human override: finance co-signature above a value materiality band, with the reason code mandatory and the serial reconciled before the posting clears; the agent that proposes the adjustment may not approve it. Measure: absolute adjustment value per period against inventory carrying value, and the share of adjustment value closed without human review — which should be zero above the band, by construction rather than by review.
None of these needs a new vendor, and none is solved by a better model. Reviewing why agent pilots underperformed, Arturo Torres Arpi Acero put the diagnosis plainly in Supply Chain Management Review in February 2026: “the failure was rarely model accuracy. It was decision design.” Teams could not say which decisions the agent could execute automatically, what threshold triggered escalation, or who owned the outcome. An adjustment queue with no value bands and no reason-code discipline is POC theater waiting to happen — and it fails quietly, in the one place where quiet failure compounds.
Measure gross before granting write
Three actions, in order, and the order is the point. First, publish both numbers. Compute signed and absolute error on the same records over the same period, express both as a percentage of physical stock, and put the ratio on the operations review beside the accuracy figure already reported. That costs a query, not a budget. Second, write the adjustment authority matrix — value bands, permitted reason codes, a named approver per band, an expiry date on every band — and enforce it at the write credential rather than in the user interface, so it binds an agent and a person identically. Third, fix attribution before autonomy. Drive reason-code coverage above 90% and map root causes, then grant an agent write authority inside the lowest band only, and only where divergence between proposed and confirmed cause has been measured for a quarter.
State the boundary as plainly as the finding. The ECR figures come from retail, mostly European; a grocery shelf is not a spare-parts bin, and the uplift measured in test stores will not transfer unchanged to a Southern Cone distribution centre. What does transfer is the arithmetic: signed error cancels and absolute error does not, in any inventory, in any currency. The falsifier is cheap — compute both numbers for one quarter, and if the ratio comes back near one, this operation does not have the problem and the control can be sized accordingly. Run that test before the authority is granted, not after the first audit finding.
So what: do not ask whether the agent can post the adjustment. Ask what it may not post, who signed that limit, and when the limit expires.
Governance
One control: an adjustment authority matrix enforced at the agent's write credential, owned by inventory control in the COO line with finance as co-signer above the materiality band. Each row names the value band, the permitted reason codes, the approver, and an expiry date. Three properties make it a control rather than a policy: the proposer may not approve its own variance; the band is defined in value, not units; and an unreviewed band lapses, so neglect fails closed. It is the operational form of what SAP calls setting what an agent may complete, and of Gartner's advice to expand autonomy from low-risk decisions — with “low-risk” defined in money and signed by a named person.
KPIs
Absolute-to-signed error ratio: gross over net on identical records; typically never computed; baseline within one count cycle. Binary accuracy: share of records where system equals physical; ECR reports over 60% of records inaccurate and roughly 35% correct in grocery. Reason-code coverage: share of postings with a specific actionable cause; baseline usually low; gate agent write authority at 90%. Auto-closed share by value band: volume and value posted without human review; target zero above the band. Absolute adjustment value per period: gross, not net, against inventory carrying value — the figure an auditor will eventually ask for.
Four-quarter roadmap
Q4 2026 — publish both numbers. Signed and absolute error on the same records, as a percentage of physical stock, with the ratio on the operations review. Q1 2027 — write and wire the matrix. Value bands, reason codes, approvers and expiry dates, enforced at the credential so it binds agents and people alike. Q2 2027 — earn attribution. Reason-code coverage above 90%, root causes mapped, and divergence measured between the cause an agent proposes and the cause a count confirms. Q3 2027 — grant narrowly and review. Agent write authority inside the lowest value band only, with the first band review raising a limit solely where divergence justifies it and letting unreviewed bands expire.
The adjustment is the control, not the clean-up
Operational decision intelligence tends to be sold at the top of the stack — the forecast, the plan, the allocation — because that is where the money appears to be. The inventory adjustment sits at the bottom, in a queue no executive reads, and is the better place to start. It is the highest-frequency decision in the warehouse, it writes to the financial record, and it is the only routine point at which an operation measures itself against physical reality rather than its own expectation. Every downstream model, every agent included, inherits its quality.
That inheritance is why the measurement choice is a governance choice. Gartner's survey of 140 senior supply chain leaders found only 17% pursuing transformational redesign of processes and workflows, against 83% applying AI incrementally to specific use cases — while a separate Gartner survey reports 55% of chief supply chain officers unclear on the return from AI investment, with 67% of supply-chain digital spend now going to AI. Those are the same finding. Incremental deployment into decisions measured by a netting metric cannot produce a legible return, because the metric was never capable of showing one. A WMS count, an ERP posting and a ledger entry that cannot be reconciled on a common reason code will never produce a number anyone can act on.
Socradata transforms ERP, WMS and supply-chain data into predictive intelligence and governed operational decision systems — and a governed decision system is one in which every recurring decision has a record, an owner, a gross number and an expiry. The inventory adjustment is the smallest decision that qualifies, and the one most likely to be automated before it is governed.
Find out what your accuracy number is cancelling
Every Wednesday, The Operational AI Dispatch translates one consequential AI signal into an operating model, a KPI set and an action plan for leaders running enterprise operations, ERP, WMS, supply chains and public systems. Published weekly by Socradata.