3 AI CRM roadblocks and how to get past them

It was an incredible moment, one that I’ll never forget. My team was working with a large customer to deploy an agentic AI use case, and we just went live. The AI solution helped service agents triage incoming cases and handle common, time-consuming customer inquiries.

It freed the customer’s team from tedious tasks and reduced the average handle time from 10 minutes to one. The customer was thrilled—so thrilled that the team recorded a music video for us to highlight the success of the project.

That’s only one example of how my team helps businesses unlock the potential of agentic AI. We’ve had many moments of amazement over the last 18 months as customers first see agentic AI handle customer support issues in their production environments.

These projects have generated millions of dollars in cost savings and freed thousands of valuable working hours from monotonous tasks.

However, we’ve also seen organizations contend with the uncertainties of deploying AI. Our customers aren’t alone. A study conducted by IDC and Lenovo showed that 88% of AI projects never make it beyond the pilot stage.

In my experience, it’s not limitations in AI technology that create roadblocks, but rather other challenges that stand in the way of realizing agentic AI’s true potential. Here are the three most common AI customer relationship management (CRM) roadblocks and how you can move past them.

1. Broken data foundation

The challenge: The old technology adage “garbage in, garbage out” also applies to AI. If your data and business processes are a mess, your AI agents won’t have sufficient context to take actions and drive resolutions.

Before considering an AI project, ensure you have the foundation needed to support your use case. A good place to start is your knowledge base. Knowledge base articles are essentially playbooks for resolving common issues, and they give clear direction to AI agents.

You can also set up case types for different customer problems to create strong guardrails for AI. Zero-copy data connectors can be used to unify customer data from different systems within your CRM system and give AI access to the information it needs to provide an accurate response.

A real-world solution: A large restaurant franchise we worked with had more than 19,000 knowledge base articles, many of which were duplicates. As a result, the customer’s first attempt to deploy agentic AI resulted in AI recommendations that were accepted only 20% of the time.

Making matters worse, a company policy allowed service agents to create new knowledge base articles for any case, even though many were similar with only slight nuances. Instead of modifying existing articles, service agents would simply create new ones. This resulted in a hodgepodge of articles that AI couldn’t interpret.

We worked with the customer on a two-step plan to increase the AI system’s success. First, we included instructions for AI to look for knowledge base articles tied to recent cases—configuration item versus simple retrieval-augmented generation (RAG) search to improve data relevancy.

Second, the customer will thoroughly vet its existing knowledge base articles and policies by using AI to pinpoint duplicates and knowledge gaps and then optimize content.

Although we’ve implemented step one only, the AI system’s accuracy has already increased. Thanks to AI-powered assistance, service agents save an average of 10 minutes per case, which is a huge impact considering the restaurant franchise receives more than 100,000 cases per year.

2. The wrong use case

The challenge: Like any tool, agentic AI is well suited for certain tasks and not great for others. It’s important to carefully consider the use case before you decide to go forward with agentic AI.

One of the first things I explain to customers is how workflows and agentic AI fit together to automate business processes. A process is essentially a series of tasks that need to be completed to resolve a customer's request.

Some of these tasks are concrete actions based on structured data and clear logic. These can include routing a case or creating subtasks using predetermined rules or escalating a case after a certain amount of time has elapsed.

Other tasks require reasoning, such as interpreting a refund policy and determining the path to resolution or if an exception is warranted. Deterministic workflows are well suited for handling concrete tasks, while AI agents can manage tasks that require interpreting unstructured data, reasoning, and actions.

Look for common customer requests that consume a lot of the team’s time. These are often repeatable processes with well-defined steps. Process mining and agent mining tools can help you better understand your workflows and identify areas where agentic AI can make an impact.

However, be sure to use agentic AI only for tasks that require reasoning, because AI is more costly to run than deterministic workflows. The probabilistic nature of AI also introduces unneeded risk to tasks that can be completed with basic logic.

A real-world solution: A large drugstore chain used agentic AI to solve all the steps in its case intake, categorization, and resolution process. We worked with the chain to rebuild its process using a blend of deterministic workflow and agentic AI.

Now when store employees submit tickets, AI analyzes the messages, extracts data from images such as product bar codes, and determines the case types based on a search of the knowledge base.

The deterministic workflows still control the overall orchestration, however, and bring in agentic AI only where reasoning is required. This approach not only reduced the overall cost to run the process, but also provided more predictability and compliance while preserving the increases in service agent efficiency gained from the power of agentic AI.

3. Denying AI agency

The challenge: When my team works with business leaders on agentic AI projects, there’s often hesitation to allow AI to complete tasks autonomously. The most common concern is AI hallucinations causing bad customer experiences. As a result, the default approach is humans in the loop with service agents reviewing and approving all AI actions.

It’s important to establish a culture where you encourage service agents to assume AI is right and that they need to double-check its work. Otherwise, service agents will assume they need to redo everything.

The human-in-the-loop design also makes it harder to measure and report on performance improvements. The goal of most AI projects is to reduce the average handle time for cases.

When a human is involved in every case, it can be difficult to distinguish which improvements are AI driven, because many factors influence key performance indicators. Use A/B testing to compare control and test groups to measure the impact of AI. Run AI agents in the background, capture results, and compare against your baseline performance metrics.

Before you start, consider if it’s feasible for AI agents to complete tasks autonomously and without human oversight. Many case types have limited consequences if AI makes a mistake. Establish a tolerance for AI mistakes and conduct thorough acceptance testing to validate that AI is performing to its benchmark.

Keep in mind that service agents make mistakes as well. Don’t demand perfection from AI. Instead, set a benchmark to match or outperform your service agents. In some cases, you may choose to set a lower benchmark if the benefit of freeing customer service capacity outweighs the reduction in quality.

A real-world example: A large airline chose to enable autonomous AI actions but restrict the responses it could provide about baggage fees. AI agents could share information about baggage policies, but they couldn’t reply with any information about fees.

This balance allowed the customer to save time with autonomous replies for certain requests while preventing costly mistakes.

Getting up to AI speed

As you gear up to add agentic AI to your CRM, it can be hard to know where to start. We’ve built a comprehensive playbook that includes guidance and resources to get your project moving.

Find out how ServiceNow can help you put AI agents to work for customer experience.