AI agents in your CRM: When they really pay off.

CRM systems are meant to bring sales, customer operations and campaigns together in one place. In reality, the work is still split across many separate processes: quotes, follow-up questions, approvals, customer requests. This is exactly where AI agents are now supposed to help and "think along" with you. It isn't quite that simple.

Between the promise and the actual benefit lies a fair amount of clarification work, and that is exactly what pays off. In her talk "AI Agents in CRM" at the CRM Experience Conference, Liana Harutyunyan of the media group Axel Springer used four lessons from practice to show what it comes down to.

According to MIT NANDA, 95 percent of all enterprise AI projects fail before they ever go live. This is rarely due to the AI itself. It is because nobody truly understood the process behind it before the technology was introduced. That is the most important lesson before you even think about agents: technology does not solve a problem that has not been clearly defined first.

Before a company builds an agent, it has to be clear what the real objective is. A simple tool for this is a use case canvas. At its core, it is a structured list of questions you answer before you touch any technology:

• What should work better in the end?
• What does the process look like today, step by step?
• Where exactly do effort and frustration arise?

Only then do you get to the actual AI idea: which step can be handed off, and what data does the solution need to work with? That order is no accident. An agent that runs into an unclear process doesn't make it clearer, it just automates the confusion faster.

Many companies reach for an agent almost by reflex as soon as AI is involved.
Salesforce has developed a simple test for this: the
Orchestration Density Framework. Three questions tell you whether you need an agent at all:

• Can you define the path to the goal completely in advance? If yes, fixed rules are enough.

• Does the process need human judgment, or is a clear rule sufficient? Judgment points to an agent.

• Does the process work with clean tables and fixed fields, or with free text and emails? Unstructured content points to an agent.

Your answers lead to a simple classification:

Low means classic automation is enough. Medium means agent and automation work together. High means you need one or more agents that make decisions on their own.

The scale of low, medium and high stays abstract until you apply it to a real case. Four examples from day-to-day CRM work show how the three levels differ in practice.

1. Automatically informing customers about a system outage is a fixed sequence with clear rules. No agent needed.
2. If the customer then comes back with questions about the impact and possible credit notes, things look different: the answer can't be defined in advance, it takes judgment. This is where an agent pays off.
3. A plain price estimate based on fixed rules doesn't need an agent either. But if you first have to understand a customer's requirements in order to build a fitting quote, you need the interplay of agent and automation.
4. At the top end of the scale sits lead qualification: handling objections, booking appointments, running multi-stage email exchanges independently. A fixed rule no longer covers it. That is a clear case for an agent.

Even a use case that has been classified correctly will fail if the underlying data is wrong. An agent can only work with information that is actually there, up-to-date and properly linked. This is where it shows whether the agent delivers at all: an agent built on outdated or duplicate CRM records doesn't get better just because the technology is new. It only exposes faster how big the data problem really is.

Before the next AI agent gets built into the CRM, three questions are worth asking:

• Is the process really understood and written down?

• Does the check against the Salesforce framework actually show a high orchestration density?

• And is the data the agent is meant to work on sound?

Answer those three honestly and you won't end up among the 95% of failed projects. You'll put agents exactly where they change something.

At snapAddy, we've built a solution for this: DataAgents. Our workflow automation platform covers a wide range of sales and marketing use cases and can map any process you need. Human-in-the-loop keeps a person in control.