By now, the story should be simple: give workers large language models (LLMs), and they get more done. Reality is not cooperating.
Across software engineering, customer support, and broader labor markets, we see the same pattern: some users become dramatically more productive with LLMs, while others slow down or barely move. Identical tools, wildly different outcomes.
Independent AI researcher, Francesco Bisardi, who works at the intersection of AI-native productivity and organizational design, has been systematically studying why some teams achieve step-change gains with large language models while others stall or regress. The work focuses on linking empirical results from coding, support, and labor-market studies with concrete design principles for AI-native organizations. In his last paper, “The LLM Productivity Cliff: Threshold Productivity and AI-Native Inequality,” he argues that the problem is not simply model capability, but a threshold phenomenon in how humans and organizations structure work around these systems.
Below a certain capability threshold, extra effort with AI yields modest or even negative returns. Above it, work practices cross a qualitative boundary and productivity jumps discontinuously.
The key predictor is not “prompt skill” or years of domain experience. In the paper, Bisardi argues that the decisive factor is “AI architectural literacy.”
From autocomplete to architecture
The paper starts from several empirical contradictions:
- In a randomized trial of experienced open-source developers, most became slower when using state-of-the-art AI coding assistants. Only one participant, with >50 hours of prior assistant use, saw clear gains.
- Large surveys of U.S. programmers show big self-reported productivity improvements for high-proficiency users, but much smaller gains for others using the same tools.
- In customer support, AI systems significantly boost junior agents but provide little or even negative impact for seniors.
- Benchmarks like SWE-bench show LLMs themselves crossing capability thresholds (e.g., from 4.4% to over 70% solved in two years), while many teams still struggle to harness that capability in production.
Rather than treating these as incompatible results, the paper reframes them as different views of the same mechanism: once LLMs are powerful enough, the bottleneck shifts from model capability to how humans structure work around them.
Bisardi defines architectural literacy as the capacity to:
- Decompose messy goals into model-tractable subtasks.
- Design multi-step workflows and agentic chains instead of one-shot prompts.
- Systematically evaluate and validate outputs against known failure modes.
- Integrate models with data, APIs, and tools so the unit of work is a workflow, not a chat.
- Maintain an adaptive mental model of what LLMs can and cannot reliably do.
This is an engineering mindset, not marginal prompt optimization. Users below this threshold treat LLMs as enhanced autocomplete or search. Users above it treat them as programmable substrates and build around that.
Three levels of practice, one cliff
To make this operational, the paper proposes a three-level model of LLM practice:
- Level 1 – Surface usage
Single prompts, obvious error detection, copy-paste into existing workflows. Effects range from −20% to +15% productivity in complex tasks. Many users and even “early adopters” are stuck here. - Level 2 – Workflow integration
Multi-step prompting, richer context, iterative refinement, basic awareness of limits. Gains become more stable, roughly in the +15–35% range in early surveys. - Level 3 – Architectural design
Task redesign around AI, programmatic APIs, orchestrated agents, automated checks, and integrated data pipelines. Here, the paper cites order-of-magnitude gains: 38% faster issue completion for highly experienced AI developers, firms reporting 70–90% of code authored by models, and experimental work showing multi-agent workflows achieving 88% faster and 90% cheaper completion than human-like prompt flows.
The “cliff” is the transition from Level 2 to Level 3. It is not reached by gradual polishing of prompts, but by reframing the nature of the work itself. Productivity rises slowly across Levels 1 and 2, then inflects sharply upward once architectural practices kick in. Below that bend, more usage can even be counterproductive, especially on complex, ambiguous tasks.
For the researcher, the cliff is not just a usability detail; it is a mechanism for AI-native inequality. Workers and organisations that cross the architectural threshold capture disproportionate gains, while others see flat or negative returns from the same tools. The paper argues that closing this gap will depend less on rolling out ever-stronger models and more on embedding architectural literacy, scaffolding, and AI-native workflow design into tools, training, and institutional practice. Bisardi situates the “LLM Productivity Cliff” as part of a broader research agenda on AI-native inequality and high-velocity development. His work connects architectural literacy in software teams with emerging divides in productivity and opportunity as LLMs diffuse through the economy. By framing the problem as a threshold phenomenon rather than a smooth adoption curve, he offers both a diagnostic lens for policymakers and operators, and a blueprint for organizations that want to move from ad-hoc AI usage to durable, system-level gains.
