The decision your forecast does not make
Three developments this year point at the same neglected object. On 6 May 2026, Gartner published a survey of 140 senior supply-chain leaders showing that only 17% are pursuing transformational redesign of processes and workflows; the remaining 83% are applying AI to isolated use cases or scaling it gradually into processes designed for something else. At the same symposium, Gartner named low data quality the leading barrier to scaling AI and told chief supply-chain officers that orchestration depends on foundational master-data alignment that technology alone cannot fix.
On 19 May 2026 at ICON, Blue Yonder shipped new skills for its Inventory Ops Agent: an agentic ordering workflow that accelerates supplier order approvals while giving planners explainable insight into how each decision was made. Reporting on the release noted that the agent's skills now go beyond setting inventory targets from a demand plan to testing whether that demand can be fulfilled under production and supply constraints. Setting an inventory target is a parameter operation, not a forecasting one. Then on 30 June 2026, Gartner's top supply-chain technology trends put agentic AI and physical AI at the head of a list organised around three themes: autonomy and agency, specialisation and intelligence, and trust and governance. The trust theme is not decoration. It is there because of what the first theme now does.
Now put the regional evidence beside it. A January 2026 study of 155 supply-chain leaders across Latin America found that 58.1% suffer stockouts that lose direct sales while 54.2% carry oversupply that immobilises working capital. Those are not two populations. They are largely the same companies, holding roughly the right amount of total inventory in demonstrably the wrong places. The same study found 47.1% operating with forecast accuracy below 70%, 63.5% reporting critical system-integration problems, and — the number that should worry an operating executive most — 45.8% who do not periodically review their digital strategy against the capabilities actually installed in their ERP or WMS.
The distribution of maturity explains the mechanism. In that sample, 31% still plan in spreadsheets, 28% are standardising, 31% have partial integration, 8% have reached automation and 2% report autonomous decisioning. Most of the region, in other words, runs replenishment through an ERP or WMS whose planning fields were populated during implementation, tuned once during stabilisation and left alone ever since. The demand signal has been re-forecast a thousand times since then. The parameters have not moved.
A company that is out of stock and overstocked in the same week does not have a forecasting problem. It has a placement problem, and placement is decided by parameters.
The Parameter Layer
Socradata uses a three-layer model to locate where a replenishment decision is actually made. It is deliberately narrow. It describes the path a single order takes from signal to commitment, and asks one question at each step: is this layer owned, measured and reconstructable?
The forecast and the demand sensing behind it. This layer is mature and heavily instrumented: accuracy, bias and forecast value added are standard vocabulary. The peer-reviewed evidence is sobering. Fildes, Goodwin and De Baets, analysing roughly 147,000 forecasts in the International Journal of Forecasting, found that judgmental adjustments improved accuracy and bias for only just over half of stock-keeping units, that upward adjustments were more likely to make things worse, and that the evidence forecasters were exploiting information unavailable to the algorithm was weak. Layer 1 is not where the remaining value sits.
The configuration that converts a signal into a quantity: safety stock, planned delivery time, reorder point, lot-size procedure, minimum order quantity, target service level, and at the pick face the WMS minimums and slotting rules. In SAP these are material-master fields; in a planning engine they are policy records; in a WMS they are replenishment thresholds. They are typically undated, unattributed and unreviewed. Celonis built a master-data improvement capability precisely because supplier lead times, safety-stock levels and reorder points drift out of correspondence with reality, and SAPinsider's guidance on static safety stock is blunt: manually entered values should be reviewed, and any preset service level should be justified. Almost no enterprise can name the owner of this layer.
Who may write to Layer 2, under what evidence, and whether the write can be reconstructed afterwards. This layer was dormant while three named planners changed parameters once a year. It is contested now that agents write into it. A Cloud Security Alliance study commissioned by Aembit, surveying 228 IT and security professionals in January 2026, found that 68% of organisations cannot clearly distinguish AI agent activity from human activity. Kiteworks' 2026 annual survey adds the operational consequence: half of organisations cannot produce a complete AI data access audit record within one business day.
The spanning metric is Parameter Value Added. Forecast value added asks whether a human improved on the algorithm's forecast. Parameter value added asks a harder and more useful question: did this change to a planning parameter improve the decision, measured as the joint movement in service level and working capital over one full replenishment cycle, against the counterfactual of leaving the prior value alone? It applies identically to a planner and to an agent, which is why it is the right instrument for 2026. KPIs before APIs: if you cannot compute parameter value added today, you are not ready to grant write access to anything, human or otherwise.
So what: an organisation that cannot say who last changed a safety-stock value, when, and on what evidence does not have a forecasting problem. It has an unowned control that quietly overrides every model it buys.
Three operating patterns
The patterns below are anonymised composites from operating work in the Southern Cone and Brazil. Ranges are illustrative rather than audited client results, and are stated as targets an operator can set. Each names the system, the decision loop, the human override path and the measure that decides whether the change worked.
Consumer-goods distributor, CABA — the parameter register. System: SAP S/4HANA material master and MRP. Decision loop: nightly MRP run, planned order, purchase requisition. The intervention is administrative before it is analytical. Every active SKU-location combination in the A and B classes gets a register row carrying the current value of safety stock, planned delivery time and lot-size procedure, the date it was set, the evidence behind it and a named owner. Safety stock and planned delivery time are then recomputed quarterly from actual goods-receipt history rather than from the contractual lead time in the vendor record. Override path: the demand planner may adjust within a ±15% band; anything outside the band routes to the S&OP governance lead with a written rationale and a mandatory expiry date. Target for the first two quarters: parameter freshness above 90% on A and B class, and every stockout event attributed to either forecast error or parameter setting. Illustrative range for working-capital release on this pattern is 8% to 15% at constant service level; treat it as an industry estimate until your own baseline says otherwise.
Regional grocery distribution centre, Southern Cone — the pick-face minimum. System: WMS replenishment thresholds and slotting rules, integrated to the ERP planning calendar. Decision loop: replenishment task generation at the pick face, which is where labour cost and short-pick risk are actually set. The failure mode here is subtler than in the ERP, because pick-face minimums are usually tuned by a supervisor responding to yesterday's short picks and never reverted. The intervention is to make every threshold effective-dated and every manual suspension time-boxed: a supervisor may suspend a slotting move or lift a minimum for one shift, the suspension carries a reason code, and it expires automatically rather than persisting as silent policy. Target: suspension rate below 10% of generated tasks within two quarters, with 100% of suspensions carrying a reason code. Illustrative range for replenishment travel reduction is 10% to 20%, and it is entirely conditional on order-profile stability.
Industrial manufacturer, São Paulo — scoped write access for the agent. System: planning engine with an agentic ordering workflow, writing back to the ERP material master. Decision loop: supplier order proposal and approval. The governing principle is that an agent gets no standing write access to the parameter layer. It receives a scoped, revocable token covering one parameter class — safety stock for C-class consumables — and runs in shadow mode for a full quarter, proposing writes that are recorded and scored but not applied. Every proposed write carries the agent's registered identity, the prior value, the new value, the evidence and an expiry. Promotion to production requires positive parameter value added on at least 70% of shadow proposals and a demonstrated ability to reconstruct any proposal within one business day. Override path: a named planner approves any write outside tolerance, and any human override of an agent proposal is itself logged as a parameter event with a reason code. This is the pattern that separates productionisation from POC theater, and the shadow quarter is the non-negotiable part.
From configuration to control
The sequence is administrative before it is technical, and that is not a concession — it is the whole point. Begin by building a parameter register: an inventory of every planning parameter currently in force by SKU-location, with its value, its effective date, its evidence and a named owner. In most enterprises this is a matter of weeks rather than quarters, and it produces two uncomfortable findings. The first is that a material share of active parameters have no discoverable rationale. The second is that several people have write access who did not know they had it.
Only then instrument the layer. Compute parameter value added retrospectively for the last four quarters using data you already hold, and attribute each stockout and excess-inventory event to either forecast error or parameter setting. That attribution is the single most valuable artefact this exercise produces, because it converts an argument about model quality into a measurable claim about configuration. It also gives the CFO something that a forecast-accuracy chart never has: a line item. From pilot to policy.
So what: agents are about to get write access to the numbers that place your orders. Grant that access to a register you can audit, or you will be reconstructing six months of replenishment decisions from memory.
Governance
One control, owned by the business. Every write to a planning parameter passes through effective-dated, dual-approved change control carrying prior value, new value, effective date, evidence, actor identity and expiry. The owner is the S&OP or planning governance lead, not IT, because the decision rights are commercial. Agents receive scoped, revocable write tokens by parameter class, never standing access, and their identity appears in the log as a registered agent rather than as the service account of whoever installed them. Overrides are time-boxed and expire by default. Interoperability or it doesn't scale: the register must reconcile across ERP, planning engine and WMS or it will be three registers disagreeing.
KPIs
Parameter freshness: share of active A and B class parameters reviewed or recomputed within 90 days. Baseline unmeasured; the Datup regional proxy is the 45.8% that never review strategy against installed ERP and WMS capability. Target above 90% in two quarters. Parameter-attributable variance: share of stockout and excess events traced to parameter setting rather than forecast error. Baseline is the Datup 58.1% and 54.2% coexistence; target is attribution of at least 80% of events within one quarter. Decision reconstructibility: share of orders whose full parameter state and acting identity can be rebuilt within one business day. Baseline unmeasured; nearest proxy is the half of organisations that, per Kiteworks, cannot produce a complete AI data access audit record in a business day. Target 100% on A class. Parameter value added: positive on at least 70% of approved changes. Override rate and half-life: below 10%, with every override expiring inside one planning cycle.
12-month roadmap
0–90: build the parameter register for A and B class, freeze and re-grant write access, baseline parameter freshness and decision reconstructibility, and compute parameter value added retrospectively on four quarters of existing data. 90–180: place A and B class under effective-dated change control with a named owner, publish parameter-attributable variance beside forecast accuracy on the supply-chain scorecard, and run the first agent in shadow mode against one parameter class. 180–360: promote agent writes to production inside tolerance bands where shadow-mode parameter value added justifies it, extend the register to WMS pick-face minimums and slotting rules, and move C class onto recomputation by exception rather than by review.
The last unmeasured control.
The industry spent a decade instrumenting the forecast and an afternoon thinking about what happens to it afterwards. That asymmetry was survivable while parameters changed rarely, and only by people whose names appeared on a change ticket. It stops being survivable the moment a planning agent can propose, and then apply, a change to a safety-stock value at machine cadence across ten thousand SKU-location combinations. The capability arrived first, as it usually does. The register did not.
There is a useful precedent here, and it is not from artificial intelligence. Financial controllers solved this problem thirty years ago by refusing to let anyone change a general-ledger account without an effective date, an approver and a reason. Nobody called that innovation. It was the price of being able to say, later, what happened and why. The planning parameter layer now carries decision authority comparable to a ledger posting — it commits cash, occupies space and determines whether a customer is served — and it is governed with roughly the rigour of a shared spreadsheet.
This is the layer Socradata works in, and the entry point is deliberately unglamorous: find the parameters that decide your orders, give each one an owner and a date, and measure whether changing it helped before you let anything else change it. Socradata transforms ERP, WMS and supply-chain data into predictive intelligence and governed operational decision systems — and governance, in this domain, begins with a register rather than a model.
Find out who owns your safety stock
Every Wednesday, The Operational AI Dispatch takes one consequential AI signal and translates it into an operating model, a KPI set and a practical action plan for leaders running enterprise operations, ERP, WMS, supply chains and public systems. Published weekly by Socradata from Buenos Aires and New York.