Applied AI Academy

AI value arrives through complements, not demos

AI as a general-purpose technology: history, hype, and business economics

The question: What makes AI different from ordinary software, and why do firms repeatedly over- and under-estimate it?

You need a mental model for AI as a general-purpose technology: value arrives through complements — data, process redesign, skills, governance — not through the model alone. This single idea explains most AI success and most AI disappointment.

What the lesson covers

AI is best understood in 2026 not as one technology but as a stack of capabilities: classical machine learning for prediction, foundation and reasoning models for cognition, multimodal systems for content and document understanding, retrieval-augmented generation for enterprise knowledge, and agents for tool-using workflows. Separate four managerial problems that get mixed together in hype: automation (doing a defined task), prediction (estimating something unknown), generation (producing drafts and content), and autonomy (deciding and acting). Each creates value differently and fails differently.

A useful economic lens: AI is a prediction technology. When the cost of prediction collapses, the value of the complements — human judgment, data, and the ability to act on predictions — rises. And prediction cost truly collapsed: between late 2022 and late 2024 the inference cost of GPT-3.5-level performance fell roughly 280-fold, from about $20 to about $0.07 per million tokens (Stanford AI Index). Capability got cheap; turning capability into value did not.

AI history alternates between springs and winters: symbolic reasoning, expert systems, statistical learning, deep learning, and now the foundation-model era. Winters happened when expectations exceeded complements — the models were never the whole story. The same pattern explains today's "pilot theatre": Stanford's AI Index reports organisational AI adoption at 88% in 2025, and McKinsey finds most firms still stuck in experimentation, with only ~39% reporting enterprise-level EBIT impact. Broad adoption, shallow value.

The productivity J-curve (Brynjolfsson, Rock & Syverson) explains the lag: measured productivity dips while firms build the complements — data foundations, redesigned workflows, new skills, governance — and only then rises. The firms pulling ahead ("high performers" in the McKinsey survey; BCG's ~5% "future-built") are not simply using AI more; they redesign workflows, assign accountable owners, and measure outcomes.

Practical consequence: diagnose any AI initiative with three questions. Is the capability real for this task? Are the complements in place (data, process, skills)? Is the control adequate (evaluation, governance, ownership)? A demo needs only the first; value needs all three.

Key points

Framework — Capability × Complement × Control

Assess any AI initiative on three axes: model capability for the specific task, organisational complements (data, workflow redesign, skills), and control quality (evaluation, governance, accountable ownership). Value requires all three; theatre only needs the first.

The lab

Learn to tell capability, complement, governance, and hype apart — on real claims.

Deliverable: AI opportunity memo v0.1 (problem · AI type · complement · first metric) — save it in your lesson notes.

Open this lesson, its lab and its quiz

Sources and further reading