Most AI pilots work fine, and that's what makes them so hard to argue with. The demo goes well, the model does its job, and then a year goes by and nobody can point to a number that moved.  

You're not the only one in this. 

 

PwC asked 4,454 CEOs this year whether AI had raised their revenue or cut their costs over the last twelve months. 56% said neither. Only 12% said both.  

 

Here's what I see constantly. The pilot gets aimed at a part of the process that was never costing anybody anything. It runs great, nobody notices, and when someone asks what it saved, the honest answer is nobody knows.  

 

The fix isn't complicated. Look at where people are losing the most time, then put the agents there. The problem is nobody looks first. When my team looked at a global bank's major incidents, the answer was 73 minutes. 

 

The Real Problem  

A large global bank came to our team with a major incident problem. These are the outages that hit critical banking services, where customers lose access to the things they depend on and every minute of downtime carries real cost. Some of these outages ran into ten figures. 

  

Our team started the way they always do, by interviewing the people who live with the problem before proposing a solution up front. What they found: resolving one of these incidents was averaging nearly six hours, and the first 73 minutes had nothing to do with fixing anything. They found those minutes by going through the process one step at a time and asking what each step was really for. You can follow those same steps in a process your team owns. 

 

The Solution  

The team built the Major Incident Agentic Analyzer: a multi-agent architecture that runs together in AI Agent Studio and kicks off the moment an incident is flagged. Each agent takes over a piece of that 73 minutes. 

 

• Problem Analyzer. Engineers used to spend that window debating how to classify the incident and whether it even qualified as major, sometimes with hundreds of people on a bridge call waiting for a verdict. The agent classifies severity immediately, using the same signals a person would check first.  

• Enrichment Agents. Someone had to manually search related incidents and change records, piecing together enough context to find a root cause. These agents run in parallel, pulling incident history, change records, and configuration data.  

• Team Router. Time went to working out who needed to be in the room and who owned which service. The agent maps escalation paths and pulls in the right people.  

• Retrospective Agent. Post-mortems got written after the fact, so the next incident started from scratch. The agent drafts root cause analysis, writes the post-mortem, and captures the knowledge so the next team starts from a stronger baseline.  

• Judge Agent. Nothing checked the work end to end. This agent validates the quality of the whole workflow, the data going in and the conclusions coming out.  

By the time a team sits down to a major incident now, they already have a full readout: what broke, why, and who to call. They spend their time resolving instead of diagnosing, and a person still owns the one piece agents shouldn't: implementing the fix. 

 

The Results  

With this solution deployed, investigation time cut in half, from 79 minutes to 40. MTTR dropped 9.3%, from 7.12 hours down to 6.46. The team's own estimate for annual savings: $4.7M conservatively, and closer to $12–15M once you factor in everything downstream. That's a 28x return, and they had that number locked in before a single agent went live. 

The entire build took twelve weeks, start to finish. 

 

Now the same playbook is showing up somewhere else: a vendor renewal build for a major retailer. Specialized AI agents doing the work, a human still in the loop where it counts, and the business case done before anyone starts building.  

 

[Watch the major incident build.] [Watch the vendor renewal build.] 

 

Build It Yourself  

You probably have a version of this problem. Not major incidents necessarily, but somewhere in your organization there's a process where people are burning time on work that could be handled without them.  

 

The way to choose is to score the candidates on two things. How much is this costing us today, and how hard would it be to build. Start where the cost is high and the build is simple. It's tempting and common for many teams to start with the hardest problem on the list because it's the one that gets talked about most and then spend six months proving nothing. If you want ideas for where to start, [the AI use case library] has more builds like this one. 

 

Where to Start  

You don't need a big program to start this. Pick one process, time the steps, and see what the numbers tell you. Most of what the bank built ran on tooling their team already had access to. And when you want a second opinion, your ServiceNow team can sit down with you and assess where the value is. 

 

Sources:  

PwC's 29th Global CEO Survey (2026) - https://www.pwc.com/gx/en/news-room/press-releases/2026/pwc-2026-global-ceo-survey.html