01 · Context

One function finished last. Almost nobody knows what it costs.

At CSCMP EDGE in Nashville in early October, Roberto Canevari, chief value chain officer at The Estée Lauder Companies, described building AI tools to support roughly 1.2 million weekly supply-chain decisions. The figure matters as an arithmetic constraint, not an achievement. No operation automates 1.2 million decisions a week by enthusiasm. It automates a subset, in a sequence, and that sequence is a portfolio decision almost nobody treats as one.

Three studies published in the following fortnight show what the sequence has been. On 6 October 2026, released at that conference, the 2027 Third-Party Logistics Study — founded by Dr John Langley of the University of Tennessee, published with Penske Logistics, NTT DATA and CSCMP — put data quality and availability at the top of the AI barrier list at 61%, ahead of legacy integration at 54% and budget at 46%. NTT DATA’s Jonathan Whitaker summarised it in a line: the technology is ready, and what lags is the foundation underneath it.

On 14 October, ASCM and the Digital Supply Chain Institute released The People Paradox in Supply Chain AI Integration, which scored AI adoption across thirteen core functions on a five-point scale from no use to fully autonomous execution. The average was 1.94: demand planning led at 2.28, supply and inventory planning followed at 2.12, and reverse logistics finished last at 1.59. The same study found the 17% rating their workforce ready reported a 71% project success rate against 28% for everyone else, with 64% of that second group still in pilot or planning.

And on 20 October, CEPAL and Chile’s national AI centre, Cenia, presented the Índice Latinoamericano de Inteligencia Artificial, putting company AI adoption in the region at 9.1% — 27.6% for large firms, 6.8% for small ones, with Argentine large firms at 34.7%. The report is explicit that adoption needs three things together: available data, trained people, and the management capability to put the result inside a production process.

Read as a maturity story, these are three measurements of the same lag, and the familiar response follows: upskill, integrate, wait. Read as an operating story they say something narrower. The 3PL study names data the binding constraint; the ASCM table shows which functions got the agents. Side by side, the ranking explains itself: the functions at the top are the ones whose decisions were already structured, timestamped, numeric fact long before anyone used the word agent. Demand planning has a time series. Reverse logistics has a photograph and a free-text note.

The decision this affects is not a model choice. It is the one an operating committee makes when it signs a technology plan: which decision loops enter the roadmap, in what order, on what evidence. On 29 April 2026, from 140 supply-chain leaders surveyed October to November 2025, Gartner found 56% naming legacy integration a major challenge and 50% citing limited expertise; Snigdha Dewal’s reading was that the friction sits in the legacy environments rather than the technology. A week later Caleb Thomson reported only 17% pursuing transformational redesign against 83% applying AI incrementally. Incremental selection without a ranking rule is not prudence. It is drift, and it has a direction.

Adoption did not rank your decisions. It ranked your data, and you funded the result.

02 · Framework

The Adoption Inversion

Socradata scores a candidate decision loop on three dimensions. Enterprises measure the first, assume the second, and discover the third in production — and that order produces the inversion visible in the ASCM table.

Layer 1 — Instrumentation

Is the decisive input already a field? Not a document, an inspector’s note, a photograph, or a value an experienced person infers from three screens — a field, with a type and a timestamp. This is the only layer the adoption table actually measures. The 61% naming data quality the leading barrier are not complaining about reporting. They are describing the filter that already selected their roadmap for them.

Layer 2 — Consequence

What one wrong decision costs, and who absorbs it. Almost never scored — and where it is, the ranking inverts. Revista Logistec reported in August, citing CEVA Logistics Chile’s operational experience, that handling a return in fifteen days or more rather than two or three can cost up to 26% of recoverable value; in the same piece WeStorage observed that Chilean companies frequently do not measure the full cost of returns at all. The function that finished last on automation holds a decision clock with a quarter-of-value penalty and no instrument pointed at it.

Layer 3 — Exception density

The share of cases that fall outside the rule. High exception density is what makes a function look unautomatable, and it is also where the human hours sit. It decides whether an agent produces leverage or relocates a queue to an approver’s inbox. It is measurable on the manual path today, before anything is procured — and a loop whose exception density is unknown is not a candidate, however good the demo.

The selection rule follows, and is simple enough to run in a spreadsheet: rank candidate loops by consequence multiplied by volume and divided by exception density, then treat instrumentation as a funded gap rather than a disqualifier. In practice the sequence runs backwards. Layer 1 acts as a gate, loops that fail it are quietly dropped, and Layers 2 and 3 are never reached. The result is a roadmap that is internally coherent, defensible in a steering committee, and ordered by the wrong variable.

So what: the adoption curve is a map of your data estate, not a map of your margin. Every year you fund it without a consequence ranking, the two drift further apart.

Two measures span the framework, and neither requires a new platform. Decision-value coverage is the share of annual decision cost running through governed automation, against the share of decision count. Count coverage is what adoption surveys report and what the 1.94 average expresses; value coverage is what the business actually bought. Where the two diverge sharply, the programme is optimising throughput of cheap decisions. The consequence-rank correlation then ranks a function’s automation score against the priced cost of a wrong decision inside it. The ASCM spread from 2.28 to 1.59 is one half of that calculation; most operations have never computed the other half. KPIs before APIs is a sequence here, not a slogan.

03 · Use Cases

Three Southern Cone loops the ranking skipped

The three patterns below are anonymised composites of operating structures common in Argentina and the Southern Cone, not client cases, and their ranges are illustrative.

01

Omnichannel retailer, CABA and AMBA — the disposition decision with no structured input. Systems: ERP sales returns and credit memo, the WMS returns disposition and putaway module, the order-management system, and a TMS reverse-pickup leg. Decision loop: on receipt, choose restock, rework, liquidate or scrap, and authorise or refuse the credit, against a clock. The loop fails the instrumentation gate for a precise reason: condition grading is a human judgement recorded as free text and reason codes are optional, so there is almost no structured input for a model to read. Its 1.59 score is a measurement artefact, not a verdict on value: the Chilean evidence puts up to 26% of recoverable value on the same clock. Make the condition code mandatory and structured, and start the clock, before any agent is procured. Human override: the returns supervisor may set a disposition below a stated value band; above it, finance co-signature is required and the proposer may not approve its own write-off. Illustrative target: recovery rate up 8–15 points once condition codes are structured and days-to-disposition is measured at p95 rather than as a monthly mean.

02

Food and beverage manufacturer, Southern Cone — the inbound shortfall nobody can model. Systems: ERP purchasing and MRP, an MES holding the production schedule and changeover matrix, and a supplier portal. Decision loop: when a confirmed inbound quantity falls short, resequence the line, substitute an approved component, or expedite at premium freight. The loop fails the instrumentation gate because the supplier commitment is not a field. It arrives by email, telephone and WhatsApp and is re-keyed later if at all — the pattern Argentine freight evidence has documented at market scale, where most trip coordination happens outside the system of record. The consequence, by contrast, is among the easiest in the plant to price: a changeover has a known cost and duration. Instrumentation here is cheap — capture the commitment at the portal with quantity, date and a confidence state, and refuse one arriving any other way. Human override: the planning manager decides inside a published substitution whitelist, and any substitution touching a regulated specification requires quality sign-off before the MES accepts it. Illustrative target: changeover hours down 10–18% on disrupted weeks.

03

Public water utility, AMBA — the maintenance release where the consequence leaves the balance sheet. Systems: an EAM or CMMS work-order backlog, SCADA condition telemetry, and ERP finance and procurement for parts and contracted crews. Decision loop: which work orders to release into next week’s crew capacity. Two instrumentation failures compound: asset condition lives in inspector narrative rather than a graded field, and criticality is inherited from a register written for a different network. But the deeper reason this loop never reaches a roadmap is that its consequence — service interruption, pressure loss, regulatory exposure — is absorbed by citizens rather than a cost centre, so it cannot enter a business case in the form the business case expects. A public operational system is another complex operational system, and here it exposes the general flaw: the selection rule is blind to any consequence without an owner in the ledger. Human override: the operations director approves releases above a criticality band, and a release deferring a critical-asset order must record a reason and an expiry after which the order re-enters mandatorily. Illustrative target: the ratio of emergency to planned work down 15–20% over three planning cycles.

04 · Implementation

Rank the portfolio before you buy the next agent

Three actions, in order, none requiring procurement. First, inventory the loops and price the error. Take two functions and list every recurring decision with four attributes: annual volume, the cost of one wrong decision, who absorbs it, and whether the decisive input is a field today. In most estates that is two weeks of work with people who already know the answers, and it produces the ranking the roadmap has been missing. Second, measure exception density on the manual path for the top quartile, which sizes the escalation queue before go-live. Third, fund instrumentation as a line item, not a prerequisite. Dropping a high-consequence loop because the data is not ready is the decision that produced the inversion.

State the boundaries as plainly as the finding. The ASCM study rests on 144 responses of which 107 met its criteria, only four of the thirteen function scores were published, and its job-level results are labelled directional. The 3PL release discloses no respondent count. The CEPAL index measures firms, not decision loops, and the Chilean returns figure is one provider’s experience in one market. What generalises is the structure, not the numbers: where consequence is unmeasured, instrumentation becomes the selection rule by default. The falsifier is cheap — rank one function’s loops by consequence, and if the order matches what you already funded, the inversion is absent from your estate.

So what: “the data is not ready” is not a decision to wait. It is a decision to leave your most expensive loop running on no instrument at all, indefinitely, and to call that caution.

Governance: the decision portfolio register

One row per decision loop, automated or candidate, owned by the COO with the CIO as co-signer — instrumentation debt is funded in the technology budget while consequence is owned in the operating line, and a register signed by only one of them gets ranked by that one’s variable. Fields: decision_id, system_of_record, consequence_value, annual_volume, exception_rate, instrumentation_state, owner, expires_on. Three properties make it a control rather than a slide: the ranking key is consequence, so no loop is funded on feasibility alone; a high-consequence loop with weak inputs is recorded as funded instrumentation debt rather than dropped; and an unreviewed row lapses, so neglect fails closed. The anchor is already written — EU AI Act Article 10 requires data sets for high-risk systems to be relevant, sufficiently representative, to the best extent possible free of errors, complete for the intended purpose, reflective of the setting of actual use, and, in paragraph 2, to have data gaps identified. Inside Annex III scope an instrumentation gap is a compliance finding, not an engineering inconvenience.

Metrics: five numbers, all obtainable this quarter

Decision-value coverage — annual decision cost under governed automation against share of decision count; baseline unmeasured, the 1.94 average standing in as the count-weighted proxy; target value above count within two planning cycles. Consequence-rank correlation — automation score against priced error cost; negative means you are automating the cheap decisions. Instrumentation debt — top-quartile loops whose decisive input is not a field, each with an owner and a dated remediation. Exception density per loop — measured on the manual path first, with a declared band above which the rule is rewritten, not the agent widened. Readiness-weighted success rate — ASCM’s baseline is stark at 71% against 28%.

Roadmap: four quarters from ranking to re-ranking

Q1 — Rank. Inventory the loops in two functions; price the error; publish version one of the register. Approve no new agent procurement in the quarter — the hardest requirement and the whole point. Q2 — Instrument. Fund structured fields for the top consequence quartile: mandatory condition codes, captured supplier commitments, graded criticality. Measure exception density on the manual path. Q3 — Deploy one. Put a single high-consequence, newly instrumented loop into governed production with declared exception bands, a named override and a sized escalation queue; keep the low-consequence pilot as the control, so the comparison is measured rather than argued. Q4 — Re-rank. Lapse unreviewed rows, publish the consequence-rank correlation to the operating committee, and build next year’s plan from measured consequence instead of vendor roadmap. This is from pilot to policy when the policy governs selection rather than capability.

Socradata Perspective

A portfolio nobody ranked is still a portfolio

Operational decision intelligence is usually sold one loop at a time: a better forecast here, a tighter allocation there, a faster disposition elsewhere. That is the right altitude for building and the wrong one for deciding. Where budgets are set, an enterprise is not choosing a model. It is allocating scarce engineering and change capacity across thousands of recurring decisions, and that allocation rule is almost never written down. When no rule is written, one is inherited — and the inherited rule is always the same: start where the data is clean.

This also explains a result that otherwise reads as a contradiction. The 3PL study finds data the leading barrier; ASCM finds readiness; Gartner finds legacy integration and talent. These are not competing diagnoses but three descriptions of one selection filter at different points in the same funnel, all consistent with an organisation executing well and ranking badly. Interoperability or it doesn’t scale has a specific sense here: the systems must exchange priorities, not only records. An ERP that knows an order’s value, a WMS that knows a disposition’s recoverable value and a ledger that knows neither are not a decision portfolio. They are three views with no shared ranking.

Socradata transforms ERP, WMS and supply-chain data into predictive intelligence and governed operational decision systems — and the first thing a governed system requires is an explicit answer to which decisions it governs, and why those and not others. The decision portfolio register is the smallest artefact that forces that answer, and the one most reliably skipped while the next pilot is scoped. For a structured read on where your own ranking has inverted, request an operational diagnostic at info@socradata.com.

Find out whether your roadmap is ranked by consequence or by convenience

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.