AI agents move from answering questions to running workflows
AI agents are expanding business automation beyond generating responses to completing tasks. They can interpret requests, plan steps, retrieve information, use authorised software tools and check results before deciding what to do next. Unlike fixed automation, agents can adapt to information gathered during a task, while organisations define their objectives, permissions and approval requirements. Applications span customer service, procurement, finance and software development.
Air India uses Salesforce Agentforce for refund processing, email resolution and passenger name changes. Salesforce said the initial refund deployment reduced internal processing time from approximately 14 days to four hours, although bank processing time remained unchanged. At travel platform Engine, Salesforce reported that agents resolved 50% of chat inquiries and reduced customer-support handling time by 15%. Coding agents can also investigate problems, suggest changes and run tests for developer review. Multiple specialised agents can coordinate complex workflows, but that approach can increase costs, complexity and errors.
Greater autonomy introduces risks, including incorrect transactions, confidential-data exposure and prompt injection through malicious external content. Effective deployment requires restricted access, activity logs, testing, interruption mechanisms and human approval for consequential actions. Adoption also depends on reliable data, integration with existing systems and measurable benefits. Businesses remain responsible for agent decisions and must establish clear accountability rather than transferring responsibility to the technology.