Salesforce disputes reports it cut 4,000 roles over AI

Salesforce has rejected claims that it replaced thousands of staff with AI agents, after criticism that its Agentforce product proved unreliable in customer deployments and forced a return to hard-coded automation for basic tasks.

The dispute centres on reports that the company planned to replace about 4,000 roles with AI, and then reconsidered after operational issues emerged. Salesforce said it did not execute mass redundancies on that basis. It said it moved roles internally and shifted staffing from support functions into sales.

"Salesforce did not replace 4,000 roles with AI, nor did it reconsider workforce decisions due to operational or reliability issues. Our approach has consistently focused on augmenting teams with AI while evolving operating models," said a Salesforce spokesperson.

The debate has played out publicly as many large companies test generative AI tools in customer service, sales and back-office workflows. These deployments have highlighted concerns about accuracy, repeatability and control. They have also raised questions about how firms talk about workforce change while piloting systems that still require supervision.

In one example described by critics of the approach, a home security company called Vivant used Agentforce in a workflow that included sending customer satisfaction surveys after each interaction. The account said the AI sometimes failed to send surveys despite being told to do so after every engagement.

Salesforce and the customer then used what Salesforce describes as "deterministic triggers" for the survey step. The approach removed the AI's discretion for that task. The system sent the survey automatically when a specified event occurred.

Critics have presented the change as evidence that standard software logic still handles routine workflow steps more reliably than an AI model. They have argued that the need for hard-coded fallbacks undercuts claims that AI agents can run end-to-end processes without close design and testing.

Salesforce pushed back on claims that its approach reflects a step backwards in AI capability, arguing that the design choices underpinning its platform are both deliberate and widely adopted across the enterprise software industry. The company said its architecture balances autonomous intelligence with clear guardrails, particularly in regulated or mission-critical environments where reliability and oversight remain paramount.

"The use of deterministic triggers for specific workflow steps is an intentional and standard enterprise AI design choice. In high-stakes or compliance-sensitive actions, deterministic logic ensures predictability, governance, and trust. This is not a failure of AI or a retreat from agentic systems, but a deliberate hybrid architecture used across the industry. Characterising this as 'forced' or as evidence of flawed technology is inaccurate and misleading," said Salesforce.

Reliability questions

Commentators have attributed the apparent reversal in job replacement to reliability problems in real-world scenarios. They have said these issues undermined confidence among executives after early enthusiasm for substituting roles with AI agents.

They have also pointed to limits in contextual understanding during more complex workflows. In this view, systems may perform adequately for low-stakes tasks such as rewriting internal communications, while struggling in processes where small changes in phrasing can change outcomes.

Another criticism has focused on testing and rollout discipline. The argument holds that some leaders accepted claims that AI could "solve everything" without subjecting tools to routine operational checks, including trialling voice dictation, summarising customer calls, and reviewing failure modes.

Salesforce, meanwhile, pointed to strong commercial traction as evidence of growing customer confidence in its approach to agentic AI, highlighting widespread adoption and real-world deployment at scale across its customer base.

"Salesforce has closed more than 18,500 Agentforce deals, with over 9,500 already in paid production. Ninety percent of Forbes' Top 50 AI companies run on Salesforce. Our confidence in agentic AI and large language models has increased, not declined, as we have incorporated learnings from production deployments into product design, governance, and customer onboarding," a Salesforce spokesperson said.

Workforce claims

Salesforce has disputed the narrative that it fired staff to replace them with AI. The company characterised recent changes as a "strategic rebalancing" rather than a downsizing move tied to Agentforce.

They denied laying off 4,000 employees and clarified that they redeployed these roles internally, primarily shifting headcount from support functions to sales to strengthen distribution capacity.

"Salesforce did not lay off 4,000 employees. Reports suggesting otherwise relate to a restructuring of portions of our customer support organisation over a months-long period. This included internal redeployment of employees to growth areas, decisions not to backfill certain roles, and natural attrition," a Salesforce spokesperson said.

Salesforce also rejected the suggestion that it had retreated from large language models because it had lost confidence. It framed its approach as an adjustment to how AI works in production, with more control mechanisms around model outputs. It said the current approach represents optimisation rather than withdrawal.

Market pressure

Critics have argued that high-profile claims about replacing staff with AI can function as a market signal. They have suggested companies may talk up labour substitution to impress investors and to market products to other executives looking to reduce costs.

Salesforce maintained that its goal differs from that framing. It said the strategy remains focused on augmenting human capability, not replacing it.

The company continues to roll out Agentforce features with an emphasis on guardrails and workflow controls, as customers assess where AI agents fit into operational processes and how much deterministic automation still sits alongside them.