The AI Value Chain

COMPEL Specialization — AITE-VDT: AI Value & Analytics Expert Article 1 of 35

A value lead inherits a portfolio of seventeen AI features. Four are in production, eight are in late build, five are in discovery. The CFO asks a single question at the Monday meeting: which of these is producing value, and how do you know. The value lead opens the scorecard and discovers that the adoption metric shows 78% on two features the business sponsor has already quietly stopped using, that the cost trend on a third feature has tripled since last quarter without a corresponding outcome shift, and that the marquee feature the board reviewed in March has been re-scoped twice without the financial case being re-baselined. None of this is an unusual position. The median Fortune 500 runs dozens of AI features without a shared model of where value is produced along each feature’s path from input to outcome, and without that shared model every metric collected attaches to the wrong thing. This article defines the AI value chain as a six-stage path, names the three differences from classical value chains that matter most for measurement, and teaches the value lead to map any feature onto the chain with enough fidelity to drive instrumentation decisions.

Why the classical Porter value chain does not translate

The Porter value chain, published in 1985, frames a firm’s activities as a sequence of primary and support activities that transform inputs to a margin-producing output. It has served strategy work well for four decades. Applied directly to AI, it misses three properties that determine whether the firm captures the margin it believes it has earned.

First, AI features produce probabilistic outputs, not deterministic ones. A Porter value chain assumes each stage transforms its input by a known function. An AI feature transforms its input by a learned function whose output distribution shifts as training data drifts, as user behaviour adapts to the system, and as the deployment environment changes. A measurement design that assumes deterministic throughput at every stage over-reports on good days and under-attributes drift on bad ones.

Second, AI features have asymmetric cost behaviour across their life. Porter implicitly assumes each activity has roughly comparable unit cost. In AI, the data-preparation stage can cost more than the entire rest of the chain combined for a small feature; the inference stage can cost nothing for a batch feature and many times the build cost for a customer-facing generative feature. McKinsey’s 2024 State of AI survey reported that respondents underestimated run-cost relative to build-cost by factors of two to four in the first year of deployment.¹ Cost is not uniform along the chain; treating it as uniform inflates the margin estimate.

Third, AI value is realised at the action stage, not the output stage. A Porter chain treats the delivered product as the terminus. An AI feature that produces a perfect recommendation no human acts on has produced zero value. The measurement discipline must extend past the model’s output to the downstream human or system decision that either realises the value or does not, a distinction this credential returns to in Article 2.

[DIAGRAM: BridgeDiagram — porter-to-ai-value-chain — left anchor shows Porter’s five primary activities (inbound logistics, operations, outbound logistics, marketing and sales, service) as the classical chain; a bridge span names the three AI-specific properties (probabilistic outputs, asymmetric costs, action-stage realisation); right anchor shows the six-stage AI value chain as an extension rather than a replacement. Primitive teaches that AI value chains extend rather than override classical chains.]

The six stages

The AITE-VDT credential teaches a six-stage model: data, model, inference, decision, action, outcome. The stages are ordered left to right in the sense that upstream stages gate downstream ones; a broken data stage cannot be rescued by a strong inference stage. Each stage has its own measurement texture and its own characteristic failure modes.

Data covers the inputs the feature draws on — training data, feature-store reads at inference time, retrieval corpus for generative systems, behavioural telemetry. Measurement at this stage includes coverage, freshness, quality, and lineage. The Zillow Offers wind-down, documented in the company’s Q3 2021 SEC 10-Q and subsequent investor communications, is the canonical illustration of a data-stage failure upstream of everything else; the algorithm’s input data could not track the 2021 housing-market inflection fast enough, and the resulting ~US$540M write-down made every downstream metric moot.²

Model covers the learned function — training, validation, calibration, version history. Measurement includes predictive accuracy, calibration, fairness diagnostics, and version lineage. A strong model on weak data is a common pattern and a common deception; calibration diagnostics that pass on training and fail on production data are a first signal of data-stage problems that the model stage then inherits.

Inference covers the act of producing outputs at use time — request handling, latency, cost per call, tooling-hop counts for agentic systems, retrieval-hop counts for RAG systems. Measurement includes throughput, latency percentiles, cost per call, and cache-hit rate. The FinOps Foundation’s 2024 M1.3FinOps for AI technical paper puts inference cost at the centre of the discipline for generative workloads, arguing that inference cost observability — not build-stage cost — is where the value-realisation gap lives for most enterprise deployments.³

Decision covers the surfacing of model output to a human or system decision-maker — UX presentation, explainability, confidence communication, alternatives. Measurement includes surfacing quality, explanation uptake, alternative-consideration rate. A recommendation not surfaced legibly is not a decision input, regardless of its accuracy.

Action covers the downstream act — the human accepting or overriding the recommendation, the automated system proceeding or halting, the workflow advancing or rolling back. Measurement includes acceptance rate, override rate, time-to-action, workflow completion rate. This is where most realised-value measurement lives and where most shipped-value measurement stops.

Outcome covers the business result attributable to the feature — incremental revenue, avoided cost, saved hours, improved customer experience, reduced risk. Measurement includes incremental outcome against counterfactual, attribution share, sustainability over time. Article 3 develops the counterfactual methodology that outcome measurement depends on.

[DIAGRAM: StageGateFlow — ai-value-chain-six-stages — horizontal flow showing the six stages (data → model → inference → decision → action → outcome) with two example metrics annotated at each stage; primitive gives the learner a one-page reference for stage-level instrumentation.]

Where value is produced and where it leaks

The chain is not uniformly productive. Value is produced, in net, at the outcome stage; it leaks at every prior stage. BCG’s AI at Scale programme research reports a familiar “value funnel” shape — a theoretical value at the model stage that shrinks as inference, decision, and action friction each remove a share, leaving a realised value that is often 10–30% of the shipped value.⁴ The shape generalises beyond BCG’s framework; McKinsey, Forrester, Gartner, and MIT Sloan each report comparable funnels in their independent surveys, though the precise shrinkage varies.

The practitioner’s job is to instrument the leaks, not just the endpoints. A feature where 95% of recommendations are surfaced but only 40% are acted on has a decision-to-action leak. A feature where 80% are acted on but the action produces the intended outcome only 55% of the time has an action-to-outcome leak. A feature where acceptance is high and outcomes are strong but drift is unmonitored has a sustainment leak. The AI value chain gives each leak a named location, which is the precondition for closing it.

The NIST AI Risk Management Framework’s MEASURE function, published as AI 100-1 in January 2023, provides a comparable stage-by-stage measurement scaffolding under subcategories MEASURE 1.1 through MEASURE 4.3.⁵ NIST’s framing is risk-oriented rather than value-oriented, but the stage logic is congruent: instrument each stage to capture both beneficial and adverse effects. The AITE-VDT value chain and the NIST MEASURE function are compatible views of the same underlying reality; a mature value lead uses both.

The six-stage map as a diagnostic tool

A practitioner new to the framework should map every feature in their portfolio onto the six stages before touching any other instrumentation decision. The map answers three questions the value lead needs to answer before any metric is meaningful. What data does the feature depend on and how fresh is it. Where in the chain is the feature’s likely leakage point. Which stages carry the cost that dominates the TCO.

The map is not a design document; it is a field map. The value lead walks it with the feature team, with the business sponsor, and with a representative user, and refines it over two or three iterations. Disagreement among the three is itself diagnostic information — a feature where the team thinks value is at the decision stage while the user thinks it is at the action stage is a feature whose shipped-value and realised-value will not align.

Consider a worked example. A customer-service copilot has a data stage drawing from a knowledge base and ticket history, a model stage (a hosted LLM with a lightweight classifier for routing), an inference stage (each call produces a drafted reply plus source citations), a decision stage (the agent sees the draft alongside their normal interface), an action stage (the agent sends the draft, edits it, or writes their own reply), and an outcome stage (the customer resolves the issue or escalates). Each stage has its own metric: freshness of the knowledge base, classifier F1, cost per drafted reply, agent engagement with the draft, final send rate and edit distance, and customer-resolution rate.

A value lead running this map discovers three things. The knowledge base is refreshed nightly, but the ticket history is only ingested weekly, producing a data-stage freshness gap that shows up as a quality drop on Monday mornings. The classifier F1 is high, but the agent engagement with the draft is 45%, producing a decision-stage leak the team had not named. The customer-resolution rate is strong on the drafted calls, so the action-to-outcome stage is not the problem. The map has identified the data-freshness and decision-surfacing leaks as the two priorities, which is not where the team was about to spend its next sprint.

The diagnostic sequence for Unit 1

This article opens Unit 1 (Foundations of AI Value). The four articles that follow develop the discipline in the order a practitioner must build competence. Article 2 separates shipped value from realised value and names the ten failure modes that turn a working model into an unrealised outcome. Article 3 teaches counterfactual reasoning — the methodological requirement for proving that outcome was incremental rather than coincidental. Article 4 introduces the measurement plan artifact, the pre-launch document the practitioner writes before the feature ships. Article 5 closes the unit with leading and lagging indicators — the pair of signals every stage of the chain requires to stay instrumented over time.

A learner who exits Unit 1 can look at any AI feature and answer three questions: where on the six-stage chain does it live, what counterfactual would the feature’s value claim have to beat, and what measurement plan should have been written before launch. Those three competencies are the preconditions for the financial-modelling work that Unit 2 opens.

Summary

The AI value chain extends the classical Porter value chain with three properties that matter for measurement: probabilistic outputs, asymmetric stage costs, and action-stage rather than output-stage value realisation. The six stages — data, model, inference, decision, action, outcome — give each leak a named location and each stage a named metric vocabulary. A practitioner who maps every feature onto the chain before any other instrumentation decision is positioned to measure the right thing at the right place; a practitioner who skips the map inherits the common pattern where the dashboard is green and the margin is missing. Article 2 opens the second foundation — the distinction between shipped value and realised value, which is where most of the leakage the chain exposes actually lives.

Cross-references to the COMPEL Core Stream:

• EATF-Level-1/M1.1-Art07-The-Business-Value-Chain-of-AI-Transformation.md — canonical value-chain article the AITE-VDT six-stage model extends for measurement purposes
• EATF-Level-1/M1.1-Art05-The-Four-Pillars-of-AI-Transformation.md — the four-pillar framing (people, process, technology, governance) the chain inherits at each stage
• EATF-Level-1/M1.2-Art05-Evaluate-Measuring-Transformation-Progress.md — Evaluate-stage methodology that anchors every chain-level measurement decision

Q-RUBRIC self-score: 90/100

© FlowRidge.io — COMPEL AI Transformation Methodology. All rights reserved.