INNOVAREModule 2 · Applied

Case 4: The Creative Machine

Every previous AI was a point solution — one model, one task. Foundation models broke that. The same model that writes your email can write code, generate images, summarise legal documents, and tutor a student. That generalisation is new — and it changes everything about how AI gets deployed and what happens when it goes wrong.

July 2026 · Case 4 of 6
As you read — hold this question

When a foundation model can do almost anything, what determines where it shouldn’t be deployed?

2.3M
customer conversations handled by Klarna’s AI in its first month. Then quality dropped. Then they started rehiring. The Klarna arc is the foundation model story in miniature.

Foundation models generate outputs that look correct. At scale, that is commercially powerful — until it meets the edge cases that training data didn’t cover, the regulatory requirements that text generation doesn’t satisfy, and the customer expectations that a confident-but-wrong response fails. The Klarna case is not a story about AI being bad. It is a story about the gap between a model that generalises well and a deployment that succeeds.

The Framework — Storey et al. (2025)
Quiz: Foundation Models

Three properties that define a foundation model

Storey et al. (2025) define foundation models by three properties that distinguish them from all previous AI systems. These are not marketing terms — they are the mechanisms that make foundation models both powerful and risky.

1
Scale — trained on data and compute orders of magnitude larger than any previous model

GPT-4 trained on roughly 1 trillion tokens. This is not incremental improvement — at sufficient scale, qualitatively new capabilities emerge that were not present in smaller models.

2
Emergence — capabilities appear that were not explicitly trained, and were not predicted

GPT-3 could do arithmetic it was never taught. GPT-4 can reason about novel ethical dilemmas. These capabilities were not in the training objectives. They emerged from scale. This is why foundation models surprise even their creators.

3
Homogenisation — a small number of foundation models underpin thousands of downstream applications

Most enterprise AI tools in 2026 are GPT-4, Claude, or Gemini with a wrapper. This concentrates risk: a bias or flaw in the foundation model propagates to every application built on it.

Applied — The Klarna arc

Deployment, 2.3M conversations, quality drop, and what came next

In February 2024, Klarna announced its AI assistant had handled 2.3 million customer service conversations in its first month — the equivalent of 700 full-time agents. The metric was impressive. What happened next was instructive.

Stage
Outcome
Month 1: AI handles 2.3M conversations. Klarna announces 700 agent roles eliminated. Resolves queries in 2 min vs 11 min average.
Commercial win
Months 2–6: Customer satisfaction scores decline. Complex queries — disputes, fraud, nuanced refunds — produce inconsistent outputs. Escalation rate rises.
Quality erosion
2024–25: Klarna quietly begins rehiring customer service staff for complex and regulated queries. CEO acknowledges quality issues in public statements.
Partial reversal
Principle revealed: Foundation model generalises across routine queries. Fails on edge cases where human judgment, regulatory knowledge, and contextual empathy are required.
Key insight
The Principle — Generalisation ≠ Reliability
Foundation models generalise across tasks better than any previous AI. That is the Storey et al. “scale” and “emergence” properties in action. But generalisation in training does not equal reliability in deployment. The gap appears at the edges — regulatory precision, emotional nuance, complex multi-step reasoning. The deployment boundary of a foundation model is not determined by the model’s capability. It is determined by the task’s tolerance for error — and who bears the consequences when the model is wrong.
Take this away

The deployment boundary of a foundation model is not determined by the model’s capability. It is determined by the task’s tolerance for error — and who bears the consequences when the model is wrong.

Previous
← Case 3: The Language Barrier
Next
Case 5: The Governance Gap →
Quick recall — without looking back

Test yourself on this case

Question 1 of 3

What are the three defining properties of foundation models in Storey et al.’s framework — and what risk does each introduce?

1. Scale: trained on data and compute orders of magnitude larger than any previous model; at sufficient scale, emergent capabilities appear that were not predicted. Risk: we cannot fully anticipate what such models can do. 2. Emergence: capabilities appear that were not explicitly trained; GPT-4 can reason about novel ethical dilemmas it was never taught. Risk: surprising failures as well as surprising successes. 3. Homogenisation: a small number of foundation models underpin thousands of downstream applications. Risk: a bias or flaw in the foundation model propagates to every application built on it.
Question 2 of 3

Walk through the Klarna arc: what happened in Month 1, what degraded in Months 2–6, and what does the outcome reveal?

Month 1: AI handled 2.3M conversations (= 700 full-time agents), resolved queries in 2 min vs 11 min average. Commercial win. Months 2–6: satisfaction scores decline; complex queries (disputes, fraud, nuanced refunds) produce inconsistent outputs; escalation rate rises. Partial reversal: Klarna quietly begins rehiring for complex and regulated queries; CEO acknowledges quality issues. Revealed: the foundation model generalised across routine queries (in-distribution) but failed on edge cases requiring human judgment, regulatory knowledge, and contextual empathy (out-of-distribution).
Question 3 of 3

Storey et al. say “generalisation ≠ reliability in deployment.” What determines the deployment boundary of a foundation model?

The deployment boundary is not determined by the model’s capability. It is determined by: (1) the task’s tolerance for error — a 5% error rate on routine queries may be acceptable; the same rate on fraud disputes is not; and (2) who bears the consequences when the model is wrong. In Klarna’s case, the consequences were borne by customers who received inconsistent information about refunds — and ultimately by Klarna’s reputation. The model’s capability didn’t change. The stakes of the deployment domain did.

Sources

Storey et al.
Storey, V.C., Lukyanenko, R., & Evermann, J. (2025). Foundation models and the future of conceptual modelling. Journal of Management Information Systems.
Foundation models
Bommasani, R. et al. (2021). On the opportunities and risks of foundation models. Stanford CRFM. arXiv:2108.07258.
Klarna
Klarna (2024, February). Klarna AI assistant handles two-thirds of customer service chats in its first month. Klarna Newsroom.
GPT-4
OpenAI (2023). GPT-4 technical report. arXiv:2303.08774.
Course material
BUSN9049 Module 2. Flinders University, 2026.