The regulatory deadline moved. The deadline at the dock did not.
Three dated facts describe a gap most operations have not noticed they are standing in.
On 13 October 2026, reporting on September shipping performance put the share of vessel arrivals inside their scheduled window at just 27%, two points below August, average container-ship delay at six days, and reliability deteriorating for a fourth month. Inside that picture sits a precise Southern Cone event: a three-day interruption of pilotage services in Argentine ports added 1.6 days to average berth access in Buenos Aires and pushed cargo to Uruguay and Brazil. For three days, every downstream system consuming a vessel ETA — yard, gate, appointment grid, carrier dispatch, labour plan — was fed an input that had quietly stopped meaning anything.
On 13 July 2026, Oracle published its Fusion Cloud SCM 26C summary, naming agentic applications for inventory planning, procurement, supplier qualification and Kanban, plus self-service dock appointment scheduling in transportation. The note says what the agents do. It does not state an approval threshold, a confidence score, a timeout, or what happens to a dock appointment when the agent cannot decide. That is the industry convention, not an Oracle failing. It holds at FourKites too, whose YardWorks AI announcement of 6 May 2025 describes a scheduling digital worker running around the clock and staff redeployed to “exception handling” — without saying where the exception boundary sits.
And on 6 May 2026, confirmed in Council on 13 May, the EU’s digital omnibus agreement moved the Annex III high-risk obligations from 2 August 2026 to 2 December 2027, Annex I products to 2 August 2028. The clock that would have obliged many deployers to document robustness and fail-safe measures slipped sixteen months. Nothing about the operation slipped with it.
Gartner predicted on 25 June 2025 that over 40% of agentic AI projects would be cancelled by the end of 2027, naming inadequate risk controls alongside escalating cost and unclear value, and expects 15% of day-to-day work decisions to be taken autonomously by 2028, from zero in 2024. Read the two together: a material share of routine decisions moves to machines within twenty-four months, inside programmes analysts expect to fail on controls rather than on model quality.
The missing control is not the one most governance programmes are building. Approval thresholds, confidence bands and autonomy levels describe a system that is functioning. Almost nobody writes what happens when the system is late, starved of inputs or unavailable — yet that behaviour exists, in every integration carrying a default timeout, a retry policy and an error branch written for a connectivity problem rather than an operational one.
Every automated decision has a second path. In most operations it was configured, not decided.
Deadline, budget, default
An automated operational decision is usually discussed as one thing: can the agent decide this, and how confident must it be? Socradata separates it into three, set by different parties at different times, only one of which is normally written down. We call the model the decision deadline; risk concentrates in the layer with the least documentation.
Regulation gestures at this without settling it. Article 15(4) of the EU AI Act requires high-risk systems to be “as resilient as possible regarding errors, faults or inconsistencies”, and notes that robustness “may be achieved through technical redundancy solutions, which may include backup or fail-safe plans.” The verb is permissive and the standard is relative rather than numeric, and the deadline is now December 2027. None of that helps a yard supervisor at 03:00. What follows converts the idea into three artefacts an operator can sign.
The moment after which the decision no longer has value, set by physics or by contract: the truck is at the gate, the wave must release for the store cut-off, the berth window opens, the declaration must be lodged. It is not an IT parameter, cannot be negotiated by the people building the automation, and is almost never recorded in the same document as the model meant to meet it. The first diagnostic question in any agent deployment is therefore simple, and almost never answered on the spot: by when must this decision exist, and who set that time?
The share of the deadline the automated path may consume before the system stops waiting. It is an architecture decision with an operational consequence, and where the two disciplines fail to meet. A retrieval step, a planning solve, a model call and three integration hops each carry a timeout inherited from a library default, and nobody has summed them against the deadline above. A budget exceeding the deadline is not a performance problem; it is a loop that will fail on its worst day, which is the day the decision mattered most. State it as a number, measured at the ninety-fifth percentile rather than the mean, because the mean describes the quiet Tuesday.
What executes when the budget expires. There are two families, and the choice between them is a business decision in technical clothing. Fail-open proceeds on a lower-fidelity rule — last known good, yesterday’s plan, a static heuristic — protecting throughput and risking a wrong action at scale. Fail-closed stops the flow and routes the decision to a named human, protecting correctness and risking a queue nobody sized. Most deployments have neither, which in practice means an accidental fail-open: the cached value is used, the previous plan persists, the grid holds. That is the second decision system, and it has no owner, no limit and no log.
So what: your agent’s autonomy setting describes what it does when it works. Your timeout default describes what your operation does when it does not — and only one of the two is written down.
Two measures span the framework, and neither needs a new platform. The first is decision-deadline headroom: the gap, at the ninety-fifth percentile, between the automated path’s elapsed time and the deadline that governs the decision. Headroom that collapses under volume means the fallback is the real operating mode at peak, which is when peak decisions are taken. The second is the fallback invocation rate: the share of decisions executed by the degraded path. Most operations cannot produce it, because the degraded path is logged as an error rather than as a decision. KPIs before APIs is not a slogan here but a sequence: an operation that cannot count how often its agent failed to decide has no basis on which to widen what the agent may decide.
Three Southern Cone loops where the second path is already running
The three patterns below are composite, anonymised operating patterns from the Southern Cone — illustrative, not client cases, with ranges labelled as such. Each names the systems of record, the decision loop, its deadline, the human override path and the measure that reveals the degraded path in use.
Port-adjacent 3PL, Buenos Aires–Campana corridor — the appointment grid with a stale input. Systems: the WMS dock-appointment and yard module, a TMS feed carrying vessel and truck ETAs, and the port community system upstream. Decision loop: re-sequencing inbound slots when berth access slips. The gate sets the deadline, not the planner — a truck arriving against a slot that no longer exists becomes detention within the hour. During the pilotage interruption the ETA feed did not fail; it kept returning values that had stopped being true. An accidental fail-open re-sequences on last-known ETAs and sends trucks to a gate that cannot receive them; a blunt fail-closed freezes the grid and fills the yard. The specified degraded state is neither: hold the published grid, suspend automated re-sequencing, release only pre-cleared loads, widen windows rather than reorder them. Human override: the yard supervisor may re-sequence manually inside a documented window, and the agent may not act on ETA data older than a stated freshness bound. Measure: detention exposure per disrupted day, split by path. Illustrative target: detention hours down 10–18% on disrupted days, achieved by refusing to act rather than acting better.
Grocery distribution centre, CABA — the wave that released on yesterday’s plan. Systems: WMS wave release, the labour-planning service sizing the wave against available headcount, and ERP replenishment downstream. Decision loop: how large a wave to release at shift start. The deadline is the store delivery cut-off, fixed by route departure times and indifferent to model availability. When the labour-planning service does not answer in time, the usual configured behaviour is to release on the previous day’s plan. That is a fail-open, invisible on every dashboard, and on a day with absenteeism it under-staffs the pick and misses the cut-off while reporting a normal release. The specified degraded state releases a reduced wave sized to confirmed headcount and flags the shortfall, so the replenishment plan records a constrained day instead of letting the forecast absorb it as demand. Human override: the shift manager holds fifteen minutes to accept, amend or escalate before the conservative plan auto-applies. Measure: on-time wave completion reported separately for model-path and fallback-path releases.
Agro-export trucking, Rosario–Santa Fe corridor — the fallback that leaves the system of record. Systems: TMS trip assignment, the carrier portal, and ERP contract and settlement. Decision loop: which carrier wins a trip, at what rate. Here the degraded path is not hypothetical but the market norm: the fourth Avancargo transport digitalisation index, on 378 responses in June 2026, found that 86.5% of companies still coordinate trips by telephone, spreadsheet or WhatsApp, only 13.5% use a specialised TMS, and 56.6% use no artificial intelligence at all. When an assignment agent cannot decide, the decision does not degrade inside the system — it leaves it, for a channel with no rate control, no audit trail and no link to settlement. The specified degraded state is a published standby rate card, a named dispatcher holding authority bands, and a write-back rule: a trip not recorded in the TMS within a fixed window is not settled. Human override: the logistics manager approves above the rate band. Measure: the shadow-assignment rate, the share of trips awarded outside the system of record, counted in most of this market for the first time.
Specify the second path before widening the first
Three actions, in order. First, time the loop against the physical deadline. For each automated decision running or about to run, record four things: the deadline, who sets it, the p95 elapsed time of the automated path today, and the resulting headroom. That is instrumentation, not investment, and in most estates it can be assembled from integration logs within a fortnight. Second, choose the default explicitly and in writing. Fail-open requires a stated value limit and an input freshness bound; fail-closed requires a sized queue with a named approver and a response-time commitment. A loop satisfying neither is not ready for autonomy, whatever its testing accuracy. Third, rehearse the degraded state. Once a quarter, run the loop with the model deliberately withheld and observe what the operation does, as distinct from what the runbook says. A fallback never exercised is POC theater with worse consequences, because it is believed in.
State the boundaries as plainly as the finding. The Avancargo index describes Argentine freight transport on 378 responses: a market, not any single operation, and the ratios differ in Brazil or Chile. The FourKites figures are vendor-reported and facility-specific. The three patterns are composite. What generalises is the structure, not the numbers: every automated decision bounded by a physical deadline has a second path, specified or inherited. The falsifier is cheap — instrument one loop for a quarter, and if the fallback never fires and headroom never approaches zero, the control can be sized down. That test costs a query. Discovering the answer during a three-day pilotage interruption costs considerably more.
So what: a fallback nobody specified is still a policy. It was written by whoever set the default in the integration layer, and it is executing today.
Governance
One control: a degradation specification, one page per automated decision loop, owned by the operations director in the COO line with the enterprise architect as co-signer, because the timeout lives in the integration, not the user interface. Each names the deadline and its source, the latency budget, the default on expiry stated as fail-open or fail-closed, the input freshness bound, the named approver and that approver’s capacity, and an expiry date. Three properties make it a control rather than an intention: the default is chosen by the business owner, not inherited from a client library; fail-open requires a written value limit, fail-closed a sized queue; and an unreviewed specification lapses, reverting the loop to human decision, so neglect fails closed.
KPIs
Four measures, each with a baseline obtainable in one cycle. Decision-deadline headroom (p95) — baseline unmeasured; target positive and stable at peak volume, not at mean. Fallback invocation rate — baseline unlogged, because the degraded path is recorded as an error; steady-state target below 2%, every excursion reviewed. Escalation queue load factor — escalations per approver-hour against the capacity the queue was sized for; degradation at design load should not exceed 70% of named-approver capacity. Shadow-decision rate — decisions taken outside the system of record; the Argentine freight baseline of 86.5% shows how large this runs before anyone counts it.
Twelve-month roadmap
Q1 — Inventory. List every automated or soon-to-be-automated decision loop with its deadline, deadline owner and p95 timing; publish headroom. Q2 — Specify. Sign degradation specifications for the five loops with the highest value at risk; instrument the fallback so it logs as a decision with its own identifier, not as an exception. Q3 — Rehearse. Run quarterly degradation drills; size the escalation queue against measured rather than assumed fallback rates; publish the shadow-decision rate. Q4 — Widen. Extend autonomy only on loops whose fallback has been exercised and measured; re-approve or lapse every specification. This is what from pilot to policy looks like when the policy is about failure rather than capability.
The system you did not design is the one that runs on the bad day
Operational decision intelligence is normally sold as an improvement to the decision itself: a better forecast, a tighter allocation, a faster disposition. That is the right place to find value and the wrong place to find risk. Risk concentrates at the boundaries — where a decision crosses from a model to a record, from automation to a person, or from the specified path to whatever was configured beneath it. Each was built by a different team solving a different problem, and none owns the operational consequence.
This is also the reading of why agent programmes stall. Gartner’s forecast names inadequate risk controls alongside unclear value; the two are one finding seen from opposite ends, because the expected value of an automated decision includes the cost of its degraded path. Interoperability or it doesn’t scale applies in a specific sense here: the systems must exchange obligations, not only data. A WMS assuming the planning service will answer, a planning service assuming the ETA feed is current, and an ETA feed with no freshness contract are not a decision system. They are a chain of assumptions that holds until a pilotage service stops for three days.
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 the failure path has an owner, a limit and a rehearsal, exactly like the success path. The degradation specification is the smallest artefact that makes that true, and the one most likely to be skipped while autonomy is being widened.
Find out what your decision system does when it cannot decide
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.