Salesforce has designed a strategy focused on driving adoption of its Agentforce agentic AI platform in enterpriseswith the aim of facilitating its transition towards an agentic model of Artificial Intelligence. To develop this strategy, which Salesforce calls “the last mile”the company assumes that large language models (LLMs) are not enough by themselves to generate business value. In fact, to be useful, according to the company, they must overcome the problems posed by the lack of context and their non-deterministic nature.
Therefore, the deployment of AI in companies in general, and Salesforce Agentforce in particular, must be supported by additional measures to take advantage of agents, proposed by the strategy. This encompasses them in four pillars: context with client data, workflow orchestration, hallucination barriers, and observability for constant monitoring. Additionally, Salesforce emphasizes its open platform approach, which is manifested through the integration of external data without copies and the use of different AI models.
The first of the pillars, focused on context, seeks to provide AI with the deep knowledge about the customer and the business that it needs. To do this, Salesforce uses shared data, bringing together all the information about what the customer has done and what they need and want. Regarding the orchestration section, the last mile strategy involves the coordination of work not only between various AI agents, but also between agents and humans.
In this sense, Salesforce Agentforce allows you to manage the complex processes of workflows in which different “sub-agents” intervene. Additionally, the company has positioned Slack as a kind of super agent that allows employees to collaborate with AI for HR, IT or purchasing tasks.
As for the barriers, given that the models are probabilistic and can delusion, Salesfore has implemented what it calls hybrid reasoning, which combines the creativity of LLM and an empathetic conversation with the deterministic execution of perfectly structured processes.
The last of the strategy pillars, focused on observability, revolves around the need for companies to know if AI is really working and solving problems after its deployment. With tools like the Test Center you can evaluate agents before they are launched. In addition, the analysis of KPIs and performance metrics allows iterating and improving the agent after its deployment.
In this sense, Salesforce allows full tracking of every decision made by the AI for greater transparency, allowing customers to adjust instructions if the agent is not achieving the results they want.
But for these four pillars to work, the infrastructure must offer sufficient means to achieve this. For this, Salesforce has a platform, which as we have mentioned is open and extensible. Developed with a “zero copy” architecture, it allows the ingestion of data from external systems, whether structured or unstructured, without the need to physically copy them.
The platform also gives companies the possibility of choose which AI model to use (OpenAI, Anthropic, etc.), or even use your own models personalized. And in terms of data quality, Salesforce has the possibilities that the purchase of Informatica has given them to guarantee that the data used by AI is clean, secure, and with a clear lineage.
