After Deflection: Planning for the Requests AI Can’t Resolve

Bright concrete hall with a row of angled gray-green columns receding into soft light.
Read more
Bright concrete hall with a row of angled gray-green columns receding into soft light.

Author

Picture of E7 Solutions
E7 Solutions

Share links

Most conversations about AI in service management eventually come around to deflection. How many requests can AI resolve without involving a person? How much time can we save? How much additional capacity can we create?

Those are reasonable questions, and the answers matter. When AI can resolve routine requests quickly and accurately, employees and customers get help sooner, and service teams have more time for other work.

But as organizations begin planning around those improvements, there’s another question worth asking: What happens to the requests AI can’t resolve?

The answer has implications for how service teams are staffed, how work gets routed, and how performance is measured. Organizations that focus exclusively on reducing ticket volume may find themselves managing a very different workload than the one they planned for.

The Question Beyond Deflection

Ticket deflection generally refers to requests resolved through self-service, automation, or AI before they require human assistance.

It can be as straightforward as an employee finding an answer in a knowledge article or resetting a password through an automated process. More capable AI agents can gather information, answer questions, and complete certain tasks across connected systems.

When these capabilities work well, fewer requests require someone’s attention. That’s a useful measure of how effectively the technology is performing.

The challenge is that deflection rates tell us relatively little about the requests that still need to be handled.

Consider a service desk that introduces an AI agent to support employees. Before implementation, its team handles everything from password questions and routine access requests to complicated application issues and exceptions requiring management approval.

Over time, the agent begins resolving more of the routine questions. The total number of tickets reaching employees decreases, but the mix of requests changes.

The team may now spend a larger share of its time investigating unusual problems, coordinating across departments, and making decisions that require experience or authority.

That changes how we should think about the workload, even if overall ticket volume is declining.

How the Remaining Workload Changes

It’s tempting to assume AI will resolve the easy requests and leave the difficult ones. In practice, the distinction isn’t always that simple.

An AI agent may be able to complete a fairly involved request when the necessary information, permissions, and workflow are available. Meanwhile, a relatively simple question may require human intervention because the documentation is incomplete or someone needs to authorize a decision.

The mix will depend on the service environment, the types of requests being submitted, and what the AI has been configured and authorized to do.

Still, as automation becomes more capable, service leaders should expect the composition of their remaining work to change.

Three types of requests deserve particular attention.

  • Requests that require judgment. These may involve exceptions to established policies, competing priorities, or circumstances where several reasonable options exist. AI can help gather information and identify relevant policies, but someone still needs to be accountable for the decision.
  • Requests that cross teams. An issue may involve IT, HR, finance, facilities, or another department. Getting it resolved depends on having the right people involved, understanding who owns each part, and keeping the work moving between them.
  • Requests that depend on specialized knowledge. Some problems require experience with a particular application, customer, or business process. That knowledge may exist within the organization without being readily available in its documentation.

These requests can require more investigation, coordination, and follow-up than routine service interactions.

If they begin representing a greater share of the queue, traditional assumptions about workload and productivity deserve another look.

A team that resolves fewer tickets may still be doing just as much work, or more, depending on what those remaining tickets involve.

The Handoff Matters

One of the most practical areas to examine is what happens when AI reaches the limit of what it can resolve.

Consider an employee requesting access to a restricted application.

An AI agent might identify the appropriate access policy, determine which permission level is relevant, and ask the employee to provide a business reason for the request.

The request still needs management approval.

In a well-designed process, the approver receives the information already collected, along with the relevant policy and a clear explanation of the decision required.

That person can review the request and take action without starting the conversation over.

Now consider what happens when the request is transferred to a general queue with little context. Someone has to review it, determine what’s missing, contact the employee, and identify the correct approver.

The employee waits longer, and the service team repeats work the AI already performed.

Both situations involve an AI agent that couldn’t complete the request independently. The operational outcomes are quite different.

This is why we think handoff design deserves attention early in an AI implementation.

When a request moves from AI to a person, several things need to be clear: why it was transferred, what has already been established, what remains unresolved, and who has responsibility for the next step.

The capabilities available will vary by platform and configuration. Some tools can pass conversation history and supporting information directly to an agent. Others may require additional workflow design or integration.

Whatever the technology, the objective is to make sure the person receiving the request can continue the work with the information they need.

A good handoff also makes ownership clear. A complicated request shouldn’t spend hours in a general queue simply because nobody has been assigned the decision it requires.

What Service Leaders May Need to Reconsider

As AI takes on more service work, there are several areas where existing operating assumptions may need to change.

Staffing and skills

Most service teams have developed their staffing models around a familiar mix of requests. They know approximately how many tickets arrive, how long common issues take to resolve, and which skills are needed at each support level.

If AI changes that mix, the existing model may become less reliable.

The remaining requests may require greater technical knowledge, stronger troubleshooting skills, more coordination across departments, or the authority to make decisions.

That doesn’t automatically mean organizations need more experienced employees or additional headcount. Some may find that AI creates enough capacity to give their existing teams more time for complex work.

Others may discover that they need different training, clearer escalation responsibilities, or better access to specialized expertise.

The important thing is to understand what work is changing before making staffing decisions based on ticket reductions alone.

Performance measurement

Service organizations rely on metrics such as ticket volume, average handling time, resolution time, and service level attainment.

Those measures remain useful, but they need context.

For example, average handling time could increase after AI adoption because the simpler requests are no longer reaching the team. That doesn’t necessarily indicate a decline in performance.

Similarly, tickets resolved per employee may fall even when the team is spending its time on more valuable or demanding work.

We’d recommend looking at AI-handled and human-handled requests separately, then examining what changes within the human workload.

Useful measures include:

  • Volume by request type: Which categories are being resolved through AI, and which continue to reach employees?
  • Handling and resolution time: How long do the remaining requests take, and is that changing?
  • Escalations and reassignments: How frequently does a request move between people or teams before resolution?
  • Time spent waiting: Where are requests sitting between steps, particularly when approval or specialized knowledge is required?
  • Quality and satisfaction: Are employees and customers getting accurate answers and satisfactory outcomes, regardless of whether AI or a person resolves the request?

These measures can provide a more complete understanding of service performance than an overall deflection percentage.

They can also help reveal whether a higher deflection rate is genuinely improving the service experience or simply shifting work elsewhere.

Process ownership

AI can also expose weaknesses in how service processes are organized.

Requests that previously relied on an experienced employee knowing who to contact may need more explicit routing and ownership rules.

Exceptions that were handled informally may require documented approval paths. Information that existed primarily in someone’s head may need to become accessible to other employees and, where appropriate, to the AI system.

This is an opportunity to improve the underlying operation.

Service leaders should pay attention to requests that repeatedly require the same manual intervention. Sometimes the answer is better documentation or additional automation. In other cases, the business needs to clarify who has authority to make a particular decision.

These are worth addressing whether or not AI is involved, but automation can make them more visible.

Planning for What AI Leaves Behind

We wouldn’t recommend redesigning an entire service organization in anticipation of changes that haven’t happened yet.

A better starting point is to establish a clear picture of the work today, introduce AI to a defined set of requests, and examine what changes.

Before expanding AI across a service operation, we’d work through a few practical steps.

Establish a baseline. Look at incoming request volume by category, the time required to resolve different types of work, escalation patterns, and where requests spend time waiting. Understand the current workload well enough to recognize meaningful changes later.

Examine the requests AI returns. Don’t limit reporting to successful deflections. Look at why requests were transferred, what information was available at the handoff, and what the receiving employee had to do next. Repeated patterns can point to opportunities for better documentation, routing, or workflow improvements.

Review ownership and escalation paths. Identify the decisions that require a person and make sure there’s a clear way to reach someone with the appropriate knowledge and authority. Test that process with real scenarios rather than assuming the existing workflow will be sufficient.

Revisit performance expectations. Compare the workload before and after AI adoption. If the composition of requests changes, determine whether service targets, staffing assumptions, and performance measures still reflect the work being performed.

Use what you learn to improve the AI. Requests that repeatedly fail because of missing information or unclear processes can become candidates for improvements. Better knowledge, clearer policies, and additional integrations may allow AI to handle more work over time.

This should be an ongoing part of managing the service operation. Each improvement in AI capability can change the balance of work between technology and people, creating new opportunities and occasionally new challenges.

What Happens After Deflection Works?

Organizations are right to invest in AI that can resolve service requests more quickly and reduce unnecessary manual effort. Done well, it can improve the experience for employees, customers, and service teams.

But the value of that investment extends beyond how many requests never reach a person.

We also need to consider how effectively the remaining work gets handled, whether employees have the right information and authority, and whether our service processes are designed for the workload we’re creating.

An organization could achieve an impressive deflection rate and still have employees waiting too long for help with the issues that matter most to them.

As organizations plan their next phase of AI adoption, we’d encourage them to spend as much time understanding the remaining workload as they do measuring the requests being automated.

That understanding should influence how teams are organized, how responsibilities are assigned, and how success is measured.

At E7, we help organizations design and improve service management processes using Atlassian technology, including the integration of AI into existing service operations. If you’re evaluating how AI will affect your service teams, we’d welcome the conversation.

Let's start with a conversation.​

We'll begin by listening: to what you're trying to achieve, where things are breaking down, and what success needs to look like. Then we'll figure out the right path forward together.

Start the conversation

Our thinking