The category arrived before the arithmetic
The sequence matters more than any single announcement. At Oracle AI World on 26 October 2026, Oracle introduced Fusion Claw — in its own words “a governed agentic execution runtime” — with three agentic applications in Fusion Cloud SCM: Supply Planning Allocation, Shipment Load Planning and Procurement Spend. Two sit on opposite sides of the same trade-off.
Two days later Gartner published its Emerging Market Quadrant for Integrated Supply Chain AI Agents — Established Vendors, by analysts Tess Frenzel and Carlton Sapp. On 3 November, SAP announced its placement in the leading tier, Market Shapers, which Gartner defines by “advanced multiagent solutions available across supply chain functions” whose agents “reason across functions and systems, sense and respond to shifts in data, and execute autonomously”.
Read that definition as a procurement specification, because that is what it now is. A solution confined to one function is placed in a lower tier: the highest rating in the category is earned by crossing functional boundaries.
That is a change in the unit of purchase, not only in capability. Until this quarter an operator bought a forecasting module, a wave-release optimiser, a load builder — each scoped to one function, measured by that function’s KPI, paid from that function’s budget. SAP’s architecture makes the shift explicit: a Supply Chain Planning Assistant within Joule coordinating demand, inventory and capacity agents, under the sequence “People direct. Assistants coordinate. Agents execute.”
The adoption evidence does not point at capability. Gartner has predicted that more than 40% of agentic AI projects will be cancelled by the end of 2027, attributing the failures to escalating costs, unclear business value and inadequate risk controls. Model quality is not on the list. In Latin America and the Caribbean the asymmetry is starker: the Inter-American Development Bank, drawing on its fAIr Tech Radar work with NTT DATA, reports around 80% of regional firms already using AI tools, 23% reporting economic benefits and just 6% seeing significant impact. Four months earlier, Gartner’s own Top Supply Chain Technology Trends for 2026, published 30 June, had named Collaborative Multiagent Systems and Decision Governance in the same list of eight: the capability and its missing control, in one document.
The capability was procured across functions. The accountability was not.
The Scope Congruence Test
Socradata separates every automated operational decision into three scopes, set by three groups of people who rarely sit in the same room and moving at three different speeds. The failure mode of an integrated supply chain agent is not that one of them is wrong; it is that they are not congruent.
The systems and transaction types the agent may write to: goods movements and requisitions in the ERP, wave release and task allocation in the WMS, load tenders in the TMS, planned orders in the planning engine. Set jointly by IT and the vendor’s integration surface, it expands fastest, because it is what release notes describe and what a proof of concept demonstrates. Oracle’s Claw runtime and SAP’s AI Agent Hub are both instruments for widening and policing this layer — identity enforcement, audit trails, observability. Necessary, and not the binding constraint.
The variables the agent optimises and, critically, the relative prices among them: service level against freight, overtime against on-time dispatch, unit price against holding cost. In practice it is set by whoever accepted the vendor’s default weighting at configuration — frequently, therefore, by nobody. It is also the only one of the three an operator can change within a quarter, and the one almost none has written down in money.
Whose cost centre, profit centre and KPI actually moves when the agent acts. Set by the chart of accounts and the annual budget, it changes once a year at best. In third-party logistics it is contractual rather than managerial, each client a separate profit centre with its own service credits; in public administration it moves only by a formal act of reallocation. Nothing in the vendor’s roadmap touches this layer.
The test is simple to state and uncomfortable to apply. A cross-functional agent is production-ready when objective scope and ledger scope are congruent: when every trade-off it may make has an approved monetary price and a cost centre that carries it. Where the two diverge nothing breaks visibly; the agent keeps working, producing decisions whose cost lands on a manager who never agreed to it and whose benefit lands on one who will never be asked to fund it. Operations has a name for that arrangement: a transfer price. Finance has a name for what happens when nobody sets one: an unexplained variance.
The public-sector mirror of this failure was published in the same month as the quadrant. ILIA 2026, the fourth edition of the Latin American AI index produced by CEPAL with Chile’s CENIA, covering 20 countries across 126 sub-indicators, finds that 85% of countries in the region have a national AI strategy in force or in preparation — and that only Brazil and Colombia have budgets assigned to theirs. Objective scope declared at national level; ledger scope empty. The same index puts average firm-level AI use at 9.1%, rising to 27.6% among large firms and falling to 6.8% among small ones — the constraint binds hardest where the chart of accounts has least room to absorb a cross-functional trade-off.
So what: an agent that reasons across four functions is making a transfer-pricing decision every time it acts, and transfer pricing is not a model problem — it is a signature problem.
Three loops where the cost and the benefit land in different books
Each pattern names the systems, the decision loop, the override path and the outcome. Outcomes are labelled illustrative where no public figure exists; the structural claim — that cost and benefit post to different ledgers — is an accounting fact checkable against any chart of accounts.
Southern Cone grocery retail — the split load that was never priced. Systems: inventory management in SAP S/4HANA and the distribution-requirements run in a planning engine, feeding load building in a TMS. Loop: overnight allocation from a Buenos Aires province distribution centre to some two hundred stores. The agent chooses between splitting a load to protect availability on a short-shelf-life line — incremental freight, posting to the logistics cost centre — and consolidating while accepting a partial stock-out, which lands as lost margin in the commercial profit centre. Oracle’s two agents released on 26 October automate the halves of this trade independently, and nothing in either defines the exchange rate between them. Override: the logistics planner may reject a split up to the dock cut-off, and any split breaching the approved freight-per-incremental-case ceiling escalates to the commercial director before release, not at month-end. Outcome, illustrative: with the rate priced and the ceiling enforced, freight per case and on-shelf availability move together rather than in opposition, recovering 2–4 points of availability on short-shelf-life lines.
Multi-client third-party logistics in AMBA — the shift that borrowed from another client’s contract. Systems: multi-tenant WMS wave release with engineered labour standards, a workforce-management module, and the provider’s per-client billing engine. Loop: wave release and labour allocation across client accounts within one shift. An agent minimising network-level labour cost will under-staff one client’s picking window to protect another’s dispatch commitment — rational against a single objective function, and a contract breach in a ledger where every client is a distinct profit centre with its own service credits. This is the sharpest form of the problem: ledger scope here is contractual, so no internal reallocation can cure it. Override: the shift supervisor holds release authority per account, and any reallocation moving hours across accounts needs prior approval from the account manager losing them. Outcome, illustrative: service-credit exposure concentrates in a few accounts, and naming the authority removes most cross-account reallocations without reducing throughput.
Provincial public health procurement, Argentina — the consolidation that moved a cost nobody could post. Systems: a public financial-administration ERP handling requisition, framework agreement and budget commitment; hospital-level stock records; a central procurement catalogue. Loop: a procurement agent consolidates requisitions across facilities to reach a better unit price. The price gain accrues centrally; the holding cost and stock-out risk transfer to individual facilities whose budget lines are legally distinct. Here the problem is explicit rather than implicit: in a public ledger a cross-functional trade-off cannot be waved through at an operating review. Override: the facility’s named budget authority may refuse a consolidation raising its holding requirement beyond an agreed threshold, escalating to the provincial finance secretariat — the only body able to post the offset. Outcome, illustrative: consolidation gains are real but overstated when read from unit price alone, and a priced transfer retains most of the gain while removing the facility-level stock-outs unpriced consolidation produces. From pilot to policy here means writing the transfer into the procurement instrument, not into a slide.
What to settle before the agent is switched on
The artefact to build is not a governance policy or another monitoring layer. It is a priced trade-off register with a named arbitrator, and it takes three moves. First, map execution scope against the cost-centre map. For every action class in the agent’s write scope, record which cost centre bears the cost and which records the benefit; the classes where those differ are the entire problem, and the list is usually short. Second, price those trade-offs in money and have each owning function countersign. Pesos per incremental case of availability; pesos per overtime hour against the service credit it avoids; pesos of holding cost accepted per point of unit-price gain. Anything unpriced becomes a mandatory escalation, not a vendor default. Third, name one arbitrator with in-quarter reallocation authority and put arbitration latency on the operating review. An escalation path ending in next year’s budget cycle is not a control; it is a queue.
So what: autonomy should expand only as fast as the ledger that absorbs its trade-offs. Price the trade-off, name the arbitrator, then raise the setting — KPIs before APIs.
Governance: the priced trade-off register
One entry per action class, enforced at the agent’s write credential rather than in a policy document: a class with no approved price blocks and escalates instead of falling back to a vendor default. Each entry names the action class, the two ledgers affected, the approved monetary weight, the countersigning function heads, the escalation owner and the review date. Re-priced quarterly, because a rate set in one freight market is wrong in the next. Owner: the COO, or in a public body the finance secretary — never IT and never the vendor, because the owner must be able to move budget, not merely approve a request. Where a vendor holds the weighting internally and will not expose it, that becomes a procurement clause: an objective function only one party can read cannot govern a decision two parties pay for.
KPIs: five numbers, each with a baseline
Cross-ledger decision share — share of agent-executed decisions whose cost and benefit post to different cost centres. Baseline: unmeasured in most operations, hence measured first. Trade-off price coverage — share of that population carrying an approved, countersigned monetary weight; baseline 0%, target 100% of the top decision-value decile before autonomy is raised. Arbitration latency — median hours from an escalated unpriced trade-off to a named human decision; target under 4 hours for wave release, 24 for allocation. Unattributed variance — month-end variance in affected cost centres untraceable to a named decision owner; target under 2% of controllable spend in scope. Out-of-scope override rate — share of overrides whose stated reason names a cost centre absent from the agent’s objective; the direct instrument for scope mismatch. A rising number means the objective is narrower than the business.
Roadmap: twelve months from capability to congruence
Q1 — Map and count. Execution scope against the cost-centre map; publish the cross-ledger action-class list; baseline the five KPIs before any autonomy setting is touched. Q2 — Price and sign. Build the register for the top decision-value decile, countersigned by each affected function head; replace vendor defaults with escalation defaults. Q3 — Widen only where the register holds. Raise autonomy class by class, with arbitration latency reported at the operating review, opening a shared cost centre where an offset has nowhere to post. Q4 — Move it into procurement. The register becomes a stated requirement in the next agent tender, and quarterly re-pricing a standing agenda item. Interoperability or it doesn’t scale applies to the ledger as much as to the API.
The integration moved. The accountability did not.
What Gartner named in October is a real architectural shift. For two decades the hard part of operational AI was getting the systems to speak to one another: planning engine to ERP, ERP to warehouse, warehouse to carrier. Vendors have now absorbed much of that work into their own agents, and Claw and the AI Agent Hub are serious engineering aimed at it. But in solving integration inside the product, they moved the binding constraint one layer up — out of the integration layer, where IT could act on it, into the chart of accounts, where IT cannot. Supply chain integration is no longer something the enterprise assembles. It arrives pre-assembled, optimising a joint objective the enterprise never specified and cannot see.
This is the most plausible reading of the regional numbers. Eighty per cent of firms in Latin America and the Caribbean using AI tools while six per cent see significant impact is not a story about model quality or data readiness, both of which have improved markedly. It is the signature of capability installed on an unchanged accountability structure — the same shape ILIA 2026 shows one level up, where eighty-five per cent of countries hold a strategy and two hold a budget line for it. Objective declared, ledger untouched. POC theater is usually called a demonstration problem; it is also what an organisation produces when a pilot’s benefit was never assigned to anyone who had to defend it in a budget review.
Operational decision intelligence sits at this seam — between the model that proposes an action, the system of record that executes it, and the ledger that absorbs the consequence. The vendors build the first two competently. The third is the operator’s own work, cannot be bought in a quadrant, and decides whether an agent produces impact or variance. Socradata transforms ERP, WMS and supply-chain data into predictive intelligence and governed operational decision systems.
Find out which of your agent’s decisions cross a ledger
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.