Enterprise modernisation is stalling long before execution begins, new research from Sweep finds, with teams spending most of their time rebuilding system context rather than implementing changes.
The State of Enterprise Systems: Salesforce Edition analysed 12,290 interactions with Sweep’s AI agent and concludes that 80% of systems work occurs in the “understand” phase, leaving only 1.2% of interactions that lead to an executed change.
Sweep frames this waste as a “Velocity Tax”: administrators spend an estimated 620–1,040 hours a year reconstructing context—equivalent to USD $42,000–$70,000 per admin, or up to $700,000 for a 10‑person team—work that rarely appears on roadmaps or sprint metrics. The report shows that most activity concentrates on dependency tracing, automation discovery and permissions analysis, with 89% of work occurring in the Understand and Plan stages for the most active users.
AI is amplifying the problem rather than solving it. While generative and agentic tools make it easier to create flows, fields and automations, they are not improving system understanding; faster change without full dependency context produces fragile metadata and downstream breakages that surface later as hard‑to‑diagnose incidents.
Sweep warns that AI‑generated metadata therefore creates a new form of technical debt: quick wins up front, longer remediation cycles afterwards.
The human cost is visible in working patterns. Planning activity more than doubled after 9pm—rising from 7.2% during daytime hours to 15.7% at night—and 7.1% of interactions referenced legacy labels such as “DEPRECATED” or “DO NOT MODIFY,” informal markers of systems that have become too perilous to touch. These signals underline how complexity forces teams into repetitive forensic work rather than value‑creating change.
Sweep’s CEO, Ido Gaver, argues that AI paired with deep system context can restore velocity: “What used to take 12 months can now be completed end‑to‑end in days. The real issue is complexity. It kills velocity.” Yet he cautions that agentic tools must be grounded in complete dependency graphs to avoid accelerating fragile change.
Third‑party commentary and industry research echo Sweep’s findings. Analysts have flagged similar risks as enterprises adopt AI rapidly but without matching governance: Deloitte’s State of AI research highlights that embedding AI into business processes requires robust data and dependency governance to avoid unintended consequences. Thought leadership on the “quality tax” of AI‑accelerated development similarly warns that short‑term speed gains can translate into higher remediation costs later.
For COOs and CIOs, the practical implications are clear. To reduce the Velocity Tax, organisations should prioritise continuous system modelling that links understanding, planning and action; invest in tools that map metadata and dependencies in real time; and introduce guardrails so agentic workflows only execute when dependency analysis confirms safety.
Doing so converts system forensics into reusable institutional knowledge, shrinking the hidden costs that currently undermine modernisation.
