Are LLMs Creating Busy Work? How to Avoid AI-Driven Bloat

The rapid adoption of LLMs and "agentic" engineering workflows has brought a hidden cost: a surge in, well, busywork. While AI promises to accelerate development, many organizations find that the ease of generating code and artifacts leads to bloated processes where engineers spend more time managing, verifying, and reviewing AI-generated output than actually solving high-level problems.

The Problem: Token Inflation

The primary concern is that companies are subconsciously equating high token usage with high productivity. Because generating code, test plans, and PRDs is now essentially "cheap" and instantaneous, these artifacts are being churned out at massive, unmonitored scales. The result is "agentic" engineering that looks productive on a dashboard but remains flat on actual business outcomes. The danger is that teams become dependent on tools that burn budget without delivering proportional value, potentially leading to a "budget wall" once the current era of aggressive subsidy or experimental spend subsides.

Optimizing AI Workflows

Despite these challenges, AI agents can be genuinely useful when integrated with rigor. Instead of just letting agents churn out infinite artifacts, consider these strategies to maintain efficiency:

• Establish Automated Guardrails: Don't rely solely on human review. Build deterministic and stochastic tests to run against every AI-generated artifact. Only escalate to human review when an automated gate fails or the task is deemed high-risk.
• Focus on Verification Surfaces: Treat artifacts (design docs, plans) as "checkpoints" rather than just paperwork. This mimics traditional, successful development cycles by ensuring alignment before moving to the execution phase.
• Switch to Cost-Conscious Tooling: If an agentic workflow is too expensive or creates unnecessary, complex loops, look for alternatives. Some lightweight VS Code extensions and integrations prioritize efficiency and lower token usage without sacrificing the ability to handle complex coding tasks.
• Measure Outcomes, Not Activity: Shift organizational focus away from "token spend" or "number of PRs." Instead, develop "artifact-per-dollar" metrics to verify that the AI integration is actually reducing total lead time to production rather than just adding layers of administrative work.

Ultimately, the goal is to utilize AI as a force multiplier that cuts through complexity, not as an engine that creates new, automated bureaucracy. When artifacts become cheap, the true value shifts entirely to the quality of the signal and the clarity of the result produced.