How Salesforce Trained AI Agents to Resolve 40% of IT Support Tickets

When something breaks at a large company – a printer isn't working, access to a system is lost, or a corporate app freezes – an employee contacts IT support. In the past, there was always a person on the other end. Today at Salesforce, an AI agent handles almost one in every two of these requests. Let's explore how they got there.

Not a Replacement for People, but an Aid to the System

An important point from the start: the agents didn't push people out of IT support roles. They took on the part of the work that is well-suited for automation – repetitive, similar requests where an algorithm performs just as well as a specialist. This freed up support staff for more complex cases requiring human involvement.

Simply put: the agent answers questions like «how do I reset my password?» or «why can't I get access?», while a human handles what the machine can't yet manage.

It All Started with Data – and an Honest Look at It

Before automating anything, the Salesforce team studied what employees were actually contacting them about. This is a crucial step that is often underestimated: without understanding the structure of these requests, any automation risks completely missing the real needs.

It turned out that a significant portion of the tickets consisted of a small set of recurring scenarios. These became the first candidates for automation. The team didn't try to cover everything at once – they chose the areas where success was most likely and started there.

How the Agents Were Built: Iteratively and Without Rushing

Development proceeded in stages. First, a prototype for a limited set of scenarios. Then, testing with real users. After that, error analysis, refinement, and scope expansion.

One of the key challenges was giving agents access to the right data at the right time. An AI agent is not just a chatbot with canned answers. It must be able to query internal systems, retrieve up-to-date information on a ticket's status, access permissions, or hardware configurations – and based on that, provide a meaningful response or perform the necessary action.

This is where the main difficulty lies: not in making the AI «know how to answer», but in enabling it to act – to integrate with systems, perform operations, and do so securely.

Security and Control – Not as an Option, but as a Foundation

The team paid special attention to what the agents could and could not do. In a corporate environment, an AI with broad access rights is a potential risk. Therefore, each agent's actions were strictly limited: it can reset a password, but it cannot change access rights without confirmation. It can check the status of a ticket, but it cannot close it bypassing established procedures.

This approach is called the «principle of least privilege» – the agent is granted exactly the permissions needed for a specific task, and no more. This reduces risks and simplifies auditing the system's actions.

How the Results Were Measured

That 40% isn't just a pretty number for a presentation. Behind it lies a specific evaluation methodology: the team tracked which tickets the agent resolved completely without human intervention, which ones required escalation to a specialist and why, and how satisfied employees were with the resolution.

Crucially, the metric for success wasn't just the «percentage of automated tickets», but the quality of their resolution. An agent that technically «handled» a request but left an employee with an unresolved problem is not a success – it's a failure in disguise.

What This Means for Everyone Else

The Salesforce story is interesting not just as a corporate case study. It shows what a real-world implementation of AI agents in a corporate environment looks like – with no magic and no quick fixes.

A few observations that apply beyond a single company:

• Start with analysis, not technology. Understanding what to automate is more important than choosing the tool.
• Iteration isn't a weakness, but a necessity. Agents aren't perfect on the first try. You need a cycle: launch, observe, and correct.
• Trust is built through limitations. The more clearly defined an agent's capabilities are, the easier it is for users and specialists to trust it.
• Metrics should reflect value, not activity. The number of tickets processed is a weak indicator. The quality of the problem's resolution is a strong one.

In short: AI agents in IT support do work – but only when they are deeply integrated into actual processes, not just «plugged into» the system.

What's Next

Salesforce isn't stopping here. The next logical step is to expand the agents' capabilities to more complex scenarios that still require a human touch today. The boundary between the «agent's zone» and the «human's zone» will gradually shift as model quality improves and more experience is gained from working with them in real-world conditions.

But it's important to understand: this isn't an autopilot you can turn on and forget about. It's a system that requires constant attention, updates, and tuning. Otherwise, that 40% can quietly turn into 40% dissatisfied users.