From capability to consequence: AI friction shows up in real systems

Big Picture

What happened: The narrative is shifting again — from capability to consequence.

If Week 20 was about plumbing (infrastructure, incentives, coordination), Week 21 is about friction showing up in real systems : organisational fatigue, uneven productivity gains, and the gap between theoretical efficiency and lived experience. AI and digital transformation are no longer judged by what they can do — but by how they actually perform under pressure.

The Through-line

Across AI and financial systems, the constraint is becoming behavioural, not technical. Tools are improving. Outputs are accelerating. But organisations are discovering that human systems — attention, cognition, incentives — do not scale at the same rate.

The result: productivity gains are real, but inconsistent — and often offset by new forms of friction.

1) AI productivity — gains are visible, but not transformative

What it is: Emerging evidence suggests AI is boosting productivity in measurable ways — faster outputs, improved workflows, reduced manual effort — but not at the level needed to drive meaningful economic transformation.

Why it matters: The expectation was step-change improvement. The reality is incremental gain.

This creates a mismatch between investment, expectation, and realised value — particularly in knowledge work, where output quality, judgment, and context still limit full automation.

First move: Shift from “AI as a multiplier” to AI as a margin optimiser . Focus on compounding small efficiency gains across workflows rather than expecting step-change productivity leaps.

Source: LinkedIn

2) Automation hits organisational limits — not technical ones

What it is: Large firms are signalling that automation has boundaries — not because the tools are insufficient, but because core services still rely on human judgment, trust, and accountability .

Why it matters: This reinforces a recurring pattern: Technology can replace tasks, but not entire roles — especially where responsibility, interpretation, or client interaction is involved.

The constraint is institutional design , not model capability.

First move: Decompose roles into decision layers. Automate execution, but retain human ownership where accountability cannot be outsourced.

Source: LinkedIn

3) Cognitive overload — the hidden cost of AI adoption

What it is: There are growing signals of “AI fatigue” — particularly in knowledge-heavy environments — where constant interaction with AI tools leads to mental overload, reduced clarity, and diminished critical thinking.

Why it matters: AI doesn’t just save time — it changes how work is processed cognitively .

If every task involves prompting, validating, and interpreting AI outputs, the cognitive burden can increase rather than decrease.

First move: Design workflows that minimise interaction overhead.

The goal is not more AI usage — it is less friction per decision .

Source: LinkedIn

4) The illusion of seamless intelligence — integration remains the bottleneck

What it is: Despite rapid advances, AI systems still struggle to integrate cleanly into existing workflows, data systems, and organisational processes.

Why it matters: This creates a familiar pattern (echoing Week 20): Efficiency gains exist in isolation — but break down at the system level .

The result is fragmented productivity: local optimisation, global inefficiency.

First move: Prioritise end-to-end integration , not tool adoption.

workflow

If AI does not reduce total process time, it is adding hidden complexity.

Source: LinkedIn

5) Labour market signals — entry-level pressure is rising

What it is: Early signals suggest a tightening environment for junior talent, as AI absorbs parts of entry-level knowledge work and reduces demand for certain roles.

Why it matters: This is one of the first second-order effects of AI adoption: Not job elimination at scale — but compression at the bottom of the experience ladder .

Over time, this risks creating a structural gap in skill development pipelines.

First move: Redesign entry-level roles around learning, judgment, and supervision of AI systems , rather than execution alone.

Source: LinkedIn

6) The coordination problem, revisited — but now human

What it is: Week 20 framed coordination as a system-level issue (institutions, incentives, infrastructure).

Week 21 reveals a parallel problem: human coordination with AI systems — attention, trust, cognitive load, and workflow alignment.

Why it matters: Even perfectly functioning technology fails if humans cannot integrate it effectively into decision-making processes.

The bottleneck is no longer just organisational — it is cognitive and behavioural .

First move: Treat AI adoption as human systems design .

Optimise for clarity, not capability.

Source: LinkedIn