Building an Agentic Enterprise with AI Agents in Production

salesforce · Tech · 2025 · source Agentic: yes · promotional bias: high · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Enterprise-wide AI agent deployment

Salesforce wanted to use AI agents to address capacity constraints across service, sales, marketing, and operations without simply adding headcount.

Constraints:

Initial enthusiasm led to too many duplicate agents and unclear results.
Agents need trusted unified data to act accurately across silos.
Agent performance requires continuous testing, tuning, and monitoring.
Humans and agents must be reorganized around impact versus scale.

Architecture

Salesforce deployed Agentforce agents on top of Data 360, starting with a focused set of hero use cases: Salesforce Help, lead engagement, sales support, and web experience. Data 360 unifies structured and unstructured data, Agentforce agents act across channels, and an Agentforce dashboard monitors ROI, speed, relevancy, satisfaction, adoption, and comparative performance.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  customer-or-prospect[Customer or Prospect] --> agentforce-agents[Agentforce Agents]
  agentforce-agents --> data-360[Data 360]
  data-360 --> external-data-platforms[External Data Platforms]
  agentforce-agents --> human-teams[Human Teams]
  agentforce-agents --> agentforce-dashboard[Agentforce Dashboard]
  agentforce-dashboard --> human-teams

Drag to pan · scroll or use +/− to zoom · reconstructed from the source's described component topology

Components

Component	Role	Interacts with
`Customer or Prospect`	Person receiving support, website answers, emails, qualification, or meeting booking.	`Agentforce Agents`
`Agentforce Agents`	Service, sales, engagement, and web agents operating across Salesforce business functions.	`Customer or Prospect`, `Data 360`, `Human Teams`, `Agentforce Dashboard`
`Data 360`	Trusted governed data layer unifying structured and unstructured internal/external data.	`Agentforce Agents`, `External Data Platforms`
`External Data Platforms`	Snowflake, Google Cloud, Amazon, and other external data platforms connected through zero-copy technology.	`Data 360`
`Human Teams`	Customer Success, Sales, Marketing, HR, IT, and other blended teams that shape, build, sustain, and tune agents.	`Agentforce Agents`, `Agentforce Dashboard`
`Agentforce Dashboard`	Real-time observability and business-value dashboard for agent ROI, quality, adoption, satisfaction, speed, and performance.	`Agentforce Agents`, `Human Teams`

Data flow

Data 360 unifies trusted structured and unstructured data from Salesforce and external platforms, Agentforce agents use that context to answer support questions or engage leads, humans tune and manage agents, and a dashboard feeds real-time performance metrics back into product and operating decisions.

“Data 360 unifies trusted, organized data, drawing from both structured and unstructured content across internal Salesforce systems and external platforms”article, stated

“We’ve set up an Agentforce dashboard that monitors in real time the key metrics we’re tracking, such as ROI, speed and performance, relevancy of responses”article, stated

Why this architecture

Salesforce learned not to deploy hundreds of duplicate agents at once; it instead anchored agents on reliable data, focused use cases, and real-time observability so agents could scale work while humans manage impact.

“made the mistake of building hundreds of agents. This led to duplication, lack of adoption, and blurry results.”article, stated

“learned was that we had to focus, earning the right to go fast, by first proving the product with a set of intentional use cases that showed real results. Our advice to you? Don’t try to go from zero to 100 immediately with your agents. Start with a use case that has reliable data, and then scale.”article, stated

Models used: Agentforce · Agentforce 360

Scale

Agentforce Service autonomously handled more than 2 million conversations.
“With Agentforce Service, we’ve autonomously handled more than 2 million conversations, allowing our people to focus on high-priority cases while customers get faster answers.”article, stated
Before Agentforce Sales, 75% of leads went untouched.
“Before Agentforce Sales, 75% of our leads went untouched because we didn’t have the capacity to engage them. Now, Agentforce autonomously reaches out to leads with personalized emails and books meetings for our SDRs, unlocking new revenue potential.”article, stated
Salesforce Help handles 2.2 million agent conversations versus 1.5 million human-engineer conversations.
“Today, Agentforce on Salesforce Help handles more conversations than our human engineers — 2.2 million vs. 1.5 million, offering 24/7 help in seven languages.”article, stated
The service agent offers 24/7 help in seven languages.
“Today, Agentforce on Salesforce Help handles more conversations than our human engineers — 2.2 million vs. 1.5 million, offering 24/7 help in seven languages.”article, stated
Salesforce claims over $100 million in annualized cost savings from Salesforce Help.
“This isn’t just about efficiency; it’s about significant cost savings. The Salesforce Help use case has resulted in over $100 million in annualized cost savings by autonomously answering routine questions and freeing up our staff to handle complex cases.”article, stated

Alternatives considered (roads not taken)

Build hundreds of agents immediately.

Rejected because: That created duplication, low adoption, and blurry results.

“made the mistake of building hundreds of agents. This led to duplication, lack of adoption, and blurry results.”article, stated

Ingest all possible data.

Rejected because: Overlapping data led to less relevant answers, so the team prioritized valuable and relevant data.

“During the process, we discovered we had overlapping data that could lead to less relevant answers. By focusing on quality over quantity, we prioritized the most valuable and relevant data instead of trying to ingest it all.”article, stated

Keep departmental teams siloed.

Rejected because: Agentic enterprise work requires blended technology and business teams to shape, build, and sustain agents.

“It’s not possible to have the kind of impact and create the agentic enterprise if everyone is working independently.”article, stated

Patterns

`agent-memory` (data-foundation)

Data 360 provides the trusted cross-system context agents need to answer accurately at scale.

“This is absolutely crucial as it provides the necessary context and governance for all agents to act accurately and at scale.”article, stated

`eval-driven-development` (operations)

Salesforce continuously tests, tunes, monitors, and iterates agent behavior rather than treating launch as done.

“This continuous testing and tuning applies across all of our agents — it never stops.”article, stated

`observability-tracing` (business-ops)

Agentforce dashboard tracks real-time ROI, speed, relevancy, satisfaction, adoption, and comparative performance.

“monitors in real time the key metrics we’re tracking, such as ROI, speed and performance, relevancy of responses, user satisfaction, and adoption.”article, stated

`human-in-the-loop-gate` (operating-model)

Salesforce frames humans as managing and tuning agent impact while agents scale work.

“The humans drive agents, managing and tuning their impact, and the agents are scaling in ways that are unbelievable.”article, stated

Tradeoffs

Rollout strategy

Chose Focused hero agents over Hundreds of agents at launch.

Rationale: Focused agents produced measurable results and a repeatable path for horizontal scaling.

“hero agents, where we started small and could move fast with low risk”article, stated

Data strategy

Chose Unified trusted Data 360 foundation with quality filtering over Siloed or duplicated data ingestion.

Rationale: Agents need the right governed data at the right time to avoid irrelevant answers.

“agents get access to the right data, at the right time, for the right purpose.”article, stated

Operating model

Chose Blended business and technology teams managing agents over Independent departmental ownership.

Rationale: Agents cut across Customer Success, Sales, Marketing, HR, and IT workflows.

“Our blended technology and business teams work together to shape, build, and sustain agents once they’re prioritized.”article, stated

Gotchas (what bit them)

Agent sprawl produces unclear value.

Root cause: Building hundreds of agents before proving focused use cases caused duplication and weak adoption.

Resolution: Start with reliable-data use cases that have clear business problems, feasibility, and ROI.

“What we quickly learned was that we had to focus, earning the right to go fast”article, stated

“use cases that had clear business problems and ROI.”article, stated

Generic outreach reduced prospect value.

Root cause: Early lead-nurture emails lacked the context needed for personalization.

Resolution: Improve prompts and add more Data 360 context to personalize emails.

“the emails it sent were too generic and not creating enough value for the prospect.”article, stated

“We improved our prompts and leveraged more information in Data 360”article, stated

More data is not always better for agents.

Root cause: Overlapping data can reduce answer relevance.

Resolution: Prioritize high-value relevant data over ingesting everything.

“By focusing on quality over quantity, we prioritized the most valuable and relevant data instead of trying to ingest it all.”article, stated

Evaluation

Salesforce evaluates agents through a real-time Agentforce dashboard, continuous testing/tuning, ROI and performance metrics, response relevance, satisfaction, adoption, and comparative scorecards.

Metrics: ROI · Speed and performance · Response relevance · User satisfaction · Adoption · Cost savings

“ROI, speed and performance, relevancy of responses, user satisfaction, and adoption.”article, stated

“we can compare our agent’s performance to others in the company and understand where to go next.”article, stated

Outcomes

Agentforce Service handled more than 2 million conversations.
Sales agents followed up with leads that previously went untouched.
Salesforce Help agents handled 2.2 million conversations versus 1.5 million by human engineers.
Salesforce reports over $100 million in annualized cost savings from the Help use case.

The article cites more than 2 million conversations, 75% previously untouched leads, 2.2M vs 1.5M service conversations, 24/7 seven-language help, and over $100M annualized savings.

“Today, Agentforce on Salesforce Help handles more conversations than our human engineers — 2.2 million vs. 1.5 million”article, stated

“over $100 million in annualized cost savings”article, stated