AI agents in theory work very well for enterprise use cases but even today their deployment and actual usage is quite limited. Why is this the case?

Putting LLMs calls over already existing APIs is similar to the way we built Horseless Carriages during early 1800s.

Lets take a very specific example

Updating opportunity(deal) stage in a CRM

When a Sales Representative works on a sale that typically takes 2-3 months, they are expected to update the stage of the deal in CRM software, such as Salesforce, as the deal progresses. An AI agent should be able to comprehend this information from meeting notes, slack messages, and emails and automatically update the CRM. There are numerous tools available that claim to automate this particular workflow, but none of them have gained widespread adoption. There are two primary reasons:

1. Every single company has different workflow(the deal stages are different)

2. The knowledge is scattered - some of the knowledge of the workflow is inside the SaaS tool and some in the Sales rep’s mind, in this case the opportunity stages are stored inside the CRM and when to move to next stage is usually in the sales team’s mind.

There are hundreds of such workflows in Salesforce alone. Creating an agent for each with different system prompts for every company isn’t a tech problem, it’s an engineering one.

The fundamental design problem

Let’s take a step back and rethink how our SaaS tools, specifically CRMs are engineered?

Carriages - The non AI version

Horseless Carriages - Agents on top of CRMs

In the above architecture the AI agent does not have access to process data and knowledge to act upon the incoming request, thus leading to failed execution in real world use cases.

Automobiles - Rule Engine

How do we go from Horseless Carriages to Automobiles?

Our approach for GTM stack starting with Salesforce

Salesforce in the heart of GTM stack - to automate the workflows in GTM stack, we are starting with Salesforce

1. Audit your system to figure out current inefficiencies and implement best practices

2. Make your process data LLM data - convert process data and knowledge to rule engines

This helps us build the intelligence layer on top of Salesforce enabling any agentic use case.