Sweep says Salesforce work is 80% context rebuilding

Sweep has published research claiming that most Salesforce system work happens before any change is executed, based on 12,290 interactions with its AI agent.

The report found that enterprise teams spend 80% of their time reconstructing system context before they can safely make changes, while just 1.2% of interactions end with an executed change. It described this as a structural problem in enterprise modernisation, with work split across understanding, planning and action rather than handled as one continuous process.

The analysis examined how teams investigate systems before making changes. Users repeatedly trace dependencies, review automations and check permissions to determine what exists, where it is used and what could break if it is changed.

Among the most active users in the dataset, 89% of work took place in the understand and plan stages. This suggests that many projects begin with what Sweep called system forensics rather than immediate execution.

Cost burden

The report estimated that administrators spend between 620 and 1,040 hours a year rebuilding system context, equivalent to USD $42,000 to USD $70,000 annually per administrator, and as much as USD $700,000 for a 10-person team.

Sweep called this cost a “Velocity Tax”, covering work completed before execution begins. The report argued that this effort is often missing from roadmaps and sprint metrics, even though it absorbs a large share of team time.

The findings also pointed to signs of strain in working patterns. Planning activity more than doubled after 9 p.m., rising from 7.2% during the day to 15.7% at night, while 7.1% of all interactions referenced labels such as “DEPRECATED,” “DO NOT MODIFY,” or “DO NOT DELETE”.

According to Sweep, those markers show how teams create informal governance when systems become harder to interpret directly. It argued that years of growth, staff turnover and one-off projects have added layers of complexity that make routine changes slower and less certain.

AI concerns

Sweep said AI tools are accelerating the creation of flows, fields, automations, agents and code, but are not improving system understanding at the same pace. In its view, that imbalance can add new dependencies faster than teams can assess them.

The report warned that agentic tools can generate metadata without a full view of the dependency graph, leading to downstream breakages that are only discovered later. It said this creates a new form of technical debt, in which AI-generated changes save time upfront but add to later diagnosis and repair work.

Sweep said it had also seen that pattern in Agentforce deployments. According to the research, early activity in those projects leaned towards planning and implementation, but investigation rates rose sharply as systems matured.

Ido Gaver, chief executive officer and co-founder of Sweep, linked the findings to broader frustration with long-running system projects. “CIOs are fed up with modernization projects that drag on for years and cost millions,” Gaver said. “AI changes that. What used to take 12 months can now be completed end-to-end in days. The real issue is complexity. It kills velocity. Traditional system integrators have built businesses around that inefficiency. AI, paired with deep system context, removes that friction and restores speed.”

The research focused on Salesforce environments and was published as State of Enterprise Systems: Salesforce Edition. It includes the full dataset analysis, Sweep's cost model for the “Velocity Tax”, early Agentforce findings and a breakdown of where investigative work is concentrated across metadata types.

The company argued that the central issue is not planning work itself, but the lack of a joined-up system linking understanding, planning and execution into a single process.