Agentic AI can only deliver trustworthy action when data, business context and workflows are properly connected
Agentic AI promises to move customer experience (CX) beyond faster answers toward systems that act and resolve requests with varying degrees of autonomy. Many organizations are trying to realize that promise by trying to simply layer AI onto the CRM systems they already have.
But legacy CRM was designed primarily to record customer interactions, rather than to orchestrate the work those interactions trigger across finance, fulfillment, field service and other functions. To deliver meaningful value, agentic AI needs an executable process behind it, rather than just technical integrations between systems.
A smarter interface can’t fix a broken foundation
The push toward agentic systems has revealed a gap between pilot and production. Adding on agents, copilots, or conversational AI promises to transform the interface by allowing customers, sales and service teams to interact with CRM systems through natural language rather than static forms. But while AI might be able to carry on a convincing conversation, that doesn’t mean it can act. Without cross-system execution, the conversation stops before resolution.
This reality highlights the main challenge of trying to layer AI on top of legacy systems: ServiceNow’s AI Maturity Index report finds that only 9% of companies have made significant progress with agentic AI for autonomous multistep workflows.
Moreover, the workflows needed to reach a complete resolution to a customer request often don’t exist at all. In other cases, years of custom code and point-to-point integrations have made it difficult for agentic AI to interpret and execute reliably. Working with a legacy foundation inevitably makes it harder to translate AI adoption into real business value, hence why only 16% of organizations have widely or fully replaced these fragmented legacy systems with a more integrated platform.
Layering agents onto a fragmented environment doesn’t address the underlying workflow gaps, so the priority for CX leaders should therefore be to establish a foundation of connected customer context, documented and repeatable end-to-end workflows, and guardrails that give agents a clear path from customer request to action and resolution.
AI breaks when work crosses system boundaries
Without a solid foundation, agentic AI may still be useful for basic tasks like summarization or surfacing recommendations, but will fall short of the ROI and productivity gains promised without the ability to act. Its ability to act reliably and autonomously becomes constrained when it runs into context blind spots like inconsistent datasets, multiple instances, and different formats spread across multiple systems like ERP and CRM. That means AI might not know what a customer owns, what they’re entitled to, or what has already been done to service their request. Without that connected context, AI will only be able to send out generic answers, rather than taking personalized action on behalf of the customer.
For instance, a customer-facing voice assistant might initially look impressive in a demo. But in production, if it doesn’t understand how the business operates or have clearly defined workflows that let it act across systems, it can converse without resolving the request. As soon as the issue reaches fulfillment, billing or another function that usually sits outside CRM, the voice agent may still have to transfer the call to a live service rep. The work still ends up in the contact center, instead of being resolved through self-service, but with a frustrated customer who had to go through a chatbot first. Without an integrated foundation capable of supporting the complete workflow all the way to resolution, the customer’s request ends up facing multiple manual handoffs, each of which adds friction to CX.
Deploying agentic workflows in a fragmented environment is no easy task. Legacy CRM has built up a mountain of knowledge and customization debt too, and that can lead to weak knowledge bases and point-to-point integrations that make it harder for agents to interpret rules or for people to govern actions. Over years, many companies have hard-wired exceptions and cross-department handoffs into bespoke code, while other processes remain manual and undocumented.
Both conditions are problematic: either the workflow is opaque and difficult to change and govern, or there’s no end-to-end workflow to execute in the first place. Those limitations only become more pronounced at scale.
Fixing the workflow gaps before adding more agents
For agentic AI to deliver real value in CX, enterprises need to start with the desired business outcome. Next, they should map the entire workflow from request to resolution, whether it’s a refund request, warranty claim, order change, or anything else that has a clear starting point and a desired outcome. This step allows teams to define what customer context and data, policies, and safety guardrails are needed to fulfill the customer’s request.
That business definition should then guide the technical work. With the workflows in place and their various dependencies identified, enterprises can fix the foundation by establishing a unified data and context layer, connecting workflows across the required systems, and implementing the appropriate governance and guardrails.
That’s the foundation organizations need before they add more agents. It doesn’t necessarily need to be a full rip-and-replace; you can start small—one team, one geo, or even one use case, like your conversational AI—and work with a vendor that can provide the resolution. That said, the optimal route depends on the business problem and the existing environment. However, once you have a foundation that can reliably connect context, workflow and action, the next question for CX leaders is what it actually means for CRM to become autonomous.
