As AI Executes More Service Work, Capability Has to Shift Upstream

Air India is expanding Salesforce Agentforce beyond refund processing into broader customer service workflows. The announced uses include multi-intent email resolution, passenger name-change automation and knowledge support for service teams. Salesforce says the system can identify several customer intents, validate information across systems, trigger specialized actions and consolidate a response, with human oversight where required. It also reports large reductions in processing time: refunds from about 14 days to four hours, and name changes from about three days to 30 minutes. Those results come from the vendor and customer, so they should be treated as reported outcomes rather than independently audited proof.

Once the system is doing the drafting, validation, routing and completion for routine cases, the employee is dealing with a different job. Fewer hours go into the standard steps. More of the day is spent on messy cases and on deciding whether the system's answer can be trusted. Sometimes the person will have to stop it and take over. That requires knowledge of the customer problem and the policy, not just the interface. If the employee remains accountable for the outcome, they also need enough authority to change what happens next.

When someone asks for Agentforce training, I'd first check what the employee is now expected to do. If the standard case is largely automated, teaching the old sequence in more detail will not solve much. The employee may instead need to recognize a bad answer, know when policy requires escalation, or understand what they are allowed to change. A course can help with knowledge and practice. It cannot give someone authority or fix an escalation path that does not work.

The evidence here is still limited to selected service workflows. The performance numbers available to us are from Salesforce's customer announcement, so I would treat them as reported results rather than independent validation. Air India says it has more than 30 agentic AI initiatives across more than 140 enterprise systems and more than 100 AI initiatives overall. It is a large program, but the sources do not tell us that jobs, staffing, management structure, career paths or decision rights have changed broadly. We also do not know how often employees override the system, which cases require escalation, or how Air India measures the quality of human intervention. The roles may end up looking mostly familiar even if the work moves faster.

If I were reviewing a similar deployment, I'd ask to see where the system is actually failing and what employees do next. I would want the exception data, the escalation rules and a clear picture of what an employee is allowed to override. That would tell me far more about the learning need than a list of Agentforce features. Some old training may disappear because the system now performs those steps. Other learning may become more important because people are left with the unusual cases and the decisions automation cannot safely finish. L&D needs to prepare people for that work, and the training plan should follow from it.

Update, 17 September 2026

Microsoft has now published a second named example that makes the broader work-design signal harder to dismiss. Its cloud supply-chain team says it simplified workflows before deploying more than 111 agents across planning, sourcing, fulfillment and logistics. Across five monthly planning cycles, Microsoft reports average cycle time falling from about 10 business days to less than 2.5. In a separate nine-person product team, engineers, designers and product managers worked with agents in roles the team described as meta-engineers, meta-designers and meta-PMs.

That does not prove Air India has redesigned its service jobs, and Microsoft's results are still internally reported. VentureBeat's independent review makes the same caution. What it does add is another concrete enterprise example where agents arrived with workflow simplification, new decision boundaries and changed role shape. That strengthens the work-design hypothesis behind this signal without turning it into a universal claim.