February 2025

www.valoir.com
Agentic artificial intelligence (AI) promises to deliver exponential
benefits from AI by automating complex tasks and interactions
without human intervention. However, creating agentic AI that
can handle complex tasks with acceptable performance has been
a challenge. Valoir found using a platform optimized for agentic AI
development like Salesforce Agentforce enables organizations to
deliver autonomous AI agents an average of 16 times faster than
other approaches while increasing accuracy by 75 percent.

The promise of agentic artificial intelligence (AI) is to exponentially

increase the benefits from AI with autonomous agents that can handle
complex tasks and interactions without human intervention. However,
many generative AI projects never moved beyond the pilot stage. Even
organizations with the resources to build models, integrate data,
engineer prompts, and build guardrails and data security found they still
couldn’t achieve an acceptable level of performance and accuracy.
These early failures led many organizations to shift from AI fear of
missing out (FOMO to AI FOMU (fear of messing up), putting generative
AI projects on hold until the technology – and vendor approaches to
delivering accuracy – improved.
At Dreamforce in September 2024, Salesforce announced Agentforce to
help customers build and deploy autonomous AI agents more quickly
and with greater accuracy and performance than DIY approaches, and
many other vendors followed with similar announcements. To
understand the potential for accelerating time to value with a platform
approach, Valoir analyzed the experiences of Salesforce customers.

Salesforce Agentforce is designed to leverage the Salesforce Platform to

deliver agentic artificial intelligence (AI) automation. Components
include:
▪ Agentforce Service Agent is a customer-facing autonomous AI
agent that helps deliver self-service resolution to customers
with pre-built topics and actions for key service use cases.
▪ Agent Builder enables users to customize, set up, test, and
activate agents with an Agent Wizard and low-code tools.
▪ The Atlas Reasoning Engine autonomously analyzes data, makes
decisions, and completes complex multi-step tasks intelligently.

Page 2 Copyright ©2025 Valoir Inc. All Rights Reserved

 ▪ Testing Center enables teams to test Agentforce using
synthetically generated data at scale.
Agentforce is a new layer that is integrated in the Salesforce platform so
customers can leverage other platform assets including:
▪ Customer 360 applications (Sales, Service, Marketing, and
Commerce Cloud) can be leveraged by Agentforce to extend
and augment the work of existing Salesforce users.
▪ Data Cloud connects, unifies, and harmonizes customer data
and metadata for grounding precise and context-aware
responses. Data Cloud retrieval augmented generation (RAG)
capabilities enable users to upload unstructured data for further
grounding. Tableau helps Agentforce see and understand data.
▪ The Einstein Trust Layer protects customer data through
security features and guardrails including zero data retention,
toxicity detection, secure data retrieval, and dynamic
grounding. The Trust Layer Audit Trail enables users to track AI
agent actions and outputs for compliance.
▪ The MuleSoft API management platform enables developers
and admins to bring in third-party data to enable Agentforce to
take action across any system.
▪ Salesforce Flow enables Agentforce to invoke any process flow
created in the low-code process automation tool Flow Builder,
or enable users to create new flows to support specific
Agentforce jobs.
▪ Slack enables Salesforce users to deploy Agentforce within its
conversational interface for employee-facing autonomous
agents.
To assess Agentforce’s comparative time to value Valoir conducted in-
depth interviews with more than 20 Salesforce customers representing
a range of industries, company sizes, and geographies (predominantly in
North America). All had experimented with some form of generative AI
to build their own agentic apps before Agentforce was released with
varying degrees of success, and were piloting Agentforce or using it in
production. For our analysis, we looked at seven phases of agentic
development:
▪ Model setup
▪ Data and application integration
▪ Prompt engineering
▪ AI guardrails and security
▪ User interface and workflow/application development
▪ Tuning
Obviously, the complexity of agentic tasks and volume, sources, and
hygiene of data varied by customer, as did the size and level of data

Page 3 Copyright ©2025 Valoir Inc. All Rights Reserved

science and technical expertise of their teams. Valoir normalized the
data by calculating the number of person-hours each team spent on
each phase of agentic AI development with Agentforce versus the
amount of time spent on the same phases of their DIY agentic projects
to quantify the average time savings for each phase and the overall
average acceleration of time to value with Agentforce.

MODEL SETUP
Valoir found that Salesforce customers were able to leverage
Agentforce’s pre-tuned large language model (LLM) and integrated
retrieval augmented generation (RAG) capabilities rather than investing
in building their own LLM or RAG databases or other pre-built models.
RAG enables more accurate and precise answers by enabling an LLM to
reference a specified set of documents to inform its answers so it can
use domain-specific or current information.
Agentforce’s pre-tuned LLM and RAG capabilities enabled them to
largely eliminate the time DIY projects devoted to initial model
evaluation, prototyping, and tuning. Customers said:
We have a beta of Agentforce that someone developed within
weeks. If we were trying to build all this in our LLM it would have
taken months to do.
We spent the last year building our version of the Salesforce
platform for my research. Four guys working on one team, five
on another. To do RAG, architecture design, custom tables and
schemas, figure out how we index and what vector database
took six months and we weren’t close to testing.
The time for development of a custom LLM could be years. However, in
reality, most organizations taking a DIY approach are using pre-built
models, which typically required three to 12 months to set up. In
contrast, Agentforce’s models are pre-integrated and pre-tuned,
requiring little to no set up time.

DATA INTEGRATION
Valoir found that the investments Salesforce had made in Data Cloud
and Data Cloud Connectors and Service Cloud Knowledge, as well as the
fact that customers’ key customer relationship management (CRM) data

Page 4 Copyright ©2025 Valoir Inc. All Rights Reserved

was already in Salesforce, significantly accelerated their ability to get
value from agentic AI with Agentforce. Customers said:
We’re getting value [from AI] in days or weeks versus months to
years, and the difference is because we have our data in
Salesforce.
Our data lives in Salesforce and for ChatGPT to grab all of that I
wouldn’t even know where to start. If you know how to build a
flow it can grab the data in Salesforce where you have the data
and just use those prompts. To pull that data out of Salesforce
to put in ChatGPT would be way more time and costly than we
could do. That’s why we started down the AI path with
Salesforce because all the data is there.
There’s a big barrier to connecting data – you’re not going to
get a SMB or midsized company to figure out how to pull that
data in. With Salesforce, I haven’t had to figure out where to get
data from.
All our knowledge is in Salesforce. With the help of one of our
SEs we built our RAG in Data Cloud in one of our sandboxes and
it only took us a few hours.
Realistically, organizations not using Agentforce would use some other
data integration tool or customer data platform (CDP) to integrate and
prepare their data for AI. Data Cloud and its pre-built RAG capabilities in
Agentforce are the two accelerators for data integration compared with
a DIY approach.
Using open source or other alternatives, organizations spent at least a
month selecting a RAG approach; integrating document ingestion,
retrieval, and storage tools; and integrating the RAG with their
generative model and an additional two to three months training the
retriever and model with domain specific data. With Agentforce they
were able to complete needed data integration within weeks.

PROMPT ENGINEERING
Valoir found that the investments Salesforce has made in prompt
engineering with Prompt Builder, now part of Agentforce, enables
organizations to easily build and reuse prompt templates that can then
be leveraged by Agentforce agents to complete specific actions.
Customers said:
We were trying to use a prompt in Excel that replaced names
with customers, but it was a huge manual effort, and also not a

Page 5 Copyright ©2025 Valoir Inc. All Rights Reserved

 streamlined process. We spent a year working on it. With
Agentforce, it took a month to operationalize.
They’ve created the structure to plug in your data, plug in your
model, and enable SMEs to do it without prompt engineering
expertise. Companies can use lesser paid resources without
advanced degrees in data science and programming to deploy AI
in a responsible way with Salesforce’s reputation behind it.
Additionally, the pre-built prompts associated with the out-of-the-box
tasks and skills in Agentforce further reduce the burden on teams to
build at the prompt level.

AI GUARDRAILS AND SECURITY

Valoir found that the guardrails and security in the Salesforce Trust
Layer were a key factor in enabling organizations to move beyond
generative AI pilots to agentic AI. Customers said:
The team was trying to train an open AI model by feeding it all
our compliance data and clinical trial data. We were concerned
about putting personal data into ChatGPT, and I don’t think we
could build the Trust Layer.
We’re having Agentforce explain a specific [person’s financial
details] in an authenticated experience. We can't be violating
[privacy laws]. Any potential leak of data would be disastrous.
That's when you really run into issues with exporting
confidential data - you just can't do it with [Microsoft] Copilot.
You could build it, but it would be a 20-30x effort.
Before when a service appointment was completed Zapier would
take the service notes and appointment record and send it to
ChatGPT and chat would summarize. It worked ok but we
decided not to keep going forward with the announcement of
Einstein Trust Layer. We couldn't trust anyone not putting it out
there in the ChatGPT database.
None of the Salesforce customers Valoir interviewed for this analysis
had successfully built the equivalent of a Trust Layer in their DIY
projects. Teams with significant expertise that tried to do so worked on
projects for more than 12 months.

Page 6 Copyright ©2025 Valoir Inc. All Rights Reserved

USER INTERFACE AND APPLICATION DEVELOPMENT
Without Agentforce and the Salesforce platform, Valoir found that
companies seeking to deliver agentic AI would increase their
development burden in two main areas: the development of a user
interface or integration into an existing conversational interface (such as
a bot) and the development of workflows between agents and the LLM.
Agent Builder’s drag and drop interface and conversational instruction
dialogues automated much of that process. Customers said:
The agentic piece is easy. To create a new agent you walk
through a few screens of what topics you want to include, where
to get its knowledge from in Data Cloud, and then you’re testing
and tuning the topic to make sure you’re getting the right
quality of response. You’re not coding and building stuff – it’s
really quite simple. We had a functioning prototype in fewer
than three weeks and now we’re doing a 6-week sprint to
production with a few people.
Copilot didn’t take action – we would have to set up flows to
trigger, and connect it to mail and a conversational UI, and each
org would have to figure it out themselves.
With Microsoft it would be five times longer because you cut out
the UI layer. It's putting the right data in the prompt and passing
the right things to the LLM - because we're all using the same
models.
Organizations already using Salesforce Flow and Einstein had the most
rapid road to agentic AI development. Customers estimated it would
take them, on average, at least six times longer to develop a customer-
facing UI and associated workflows without Agentforce.

TUNING
Valoir found that tuning agentic AI to deliver an acceptable level of
accurate results without hallucinations or toxicity was the biggest
hurdle for DIY agentic AI projects. In fact, the majority of DIY projects
Valoir analyzed never reached acceptable levels of accuracy despite
having been tuned for significant periods of time. In contrast, the
combination of the Atlas Reasoning Engine, Testing Center, and the
ability to tune agents with conversational language instructions enabled

Page 7 Copyright ©2025 Valoir Inc. All Rights Reserved

business users to rapidly tune and test models to ensure they could
deliver accurate results. Customers said:
We took the open AI API from Microsoft and built a connector
from Einstein to Copilot so when it detected certain topics it
would route to a RAG built on Azure and then ChatGPT’s LLM. It
didn’t work. It hallucinated way too much because it had no
logic or reasoning layer for complex tasks. To tune all the
scenarios and eliminate hallucinations it would have been years
of work for our 4000 SKUs.
Topics are how you guide the agent and instructions are on
specific tuning. If the accuracy isn’t there you test it and say,
“where do I need to add more instructions or be clearer?”
Now tuning is just adding instructions to the AI agent. Someone
with a few years in industry and that understands business
language can manage Agentforce.
Clearly, the time needed to tune agentic AI varies broadly based on the
complexity of tasks to be performed. However, Valoir found that tuning
for accuracy was a key area where Agentforce delivered, with
customers finding increased accuracy of 75 percent over DIY with minor
tuning:
With DIY we were able to get to 60 percent accurate. With what
I’m seeing with Agentforce we’ve been able to push it up to 85
percent.
If I were to build it in a custom GPT it would be less accurate
because you don’t have business process and context around it,
and it will guess at the best plausible answer. When I asked
ours, I had a 50-50 chance of it getting it right. Now we’re at 95
percent.
Customers estimated they would need years to achieve the same level
of accuracy that they achieved in just weeks with Agentforce.

A key factor in time to value was time to accuracy; that is, the time
needed to build and train AI agents to deliver acceptable levels of
correct responses. Obviously it is a lot faster to build and train an agent
to respond to simple queries like password change requests than it is to
train an agent to successfully navigate multi-level decision trees and
responses depending on data from multiple sources. When it came to

Page 8 Copyright ©2025 Valoir Inc. All Rights Reserved

the Salesforce customers interviewed for our research, Valoir found
that:
▪ Approximately two-thirds of customers had reached the point with
their DIY project where they were able to test for accuracy using
multiple sets of simulated queries. They found that 52 percent of
the answers or resolutions their agents delivered, on average, had
what they deemed an acceptable level of accuracy. The other
customers who had abandoned their DIY efforts at some point
because of low accuracy, security concerns, or resource constraints
– or because Agentforce became available and it was deemed a
better option than DIY – didn’t believe their DIY efforts, had they
continued, would have delivered more than 50 percent accuracy.
▪ Sixty-two percent of customers had, by late January or early
February, one or multiple Agentforce agents in pilot or production
long enough to test the accuracy of agent responses, and found that
91 percent of answers or resolutions their agents delivered, on
average, had an acceptable level of accuracy – representing a 75
percent increase in accuracy over DIY.

SELECTED CUSTOMER DIY-AGENTFORCE AGENT EXAMPLES

Agent DIY Agentforce %

Description
complexity accuracy accuracy change
Employee-facing HR leave
Simple 50% 95% 90%
request agent
Customer self-service agent
Moderate for complex product catalog 60% 85% 42%
with 1000+ SKUs
Customer self-service agent
Complex with authentication and 40% 80% 100%
personalization based on PII

Complex sales coaching

using conversation insights
Complex 40% 90% 77%
for improvement
suggestions and role playing

Beyond simple use cases, we found that the greatest change in accuracy
in Agentforce agents versus DIY agents was often in more complex
agents that were dependent on multiple data sources for grounding and
often relied on more complex reasoning and workflows to come to the
accurate answer based on the context of the customer query.

Page 9 Copyright ©2025 Valoir Inc. All Rights Reserved

As Agentforce is relatively new, customers are still discovering how
much tuning and support will be needed on an ongoing basis. However,
customers recognize they are achieving agentic AI results that would
likely be too resource-intensive to support without the Agentforce
platform, saying:
We don’t have the skill, time, or money to build guardrails and
then audit them on an ongoing basis to make sure that the
model doesn’t hallucinate.
A big part of AI is the ongoing maintenance. You have to check it
all the time, look at the logs, and it ends up being a product to
manage. Either way you’re doing some work, but do you want
to manage and tune an LLM on an ongoing basis? And build all
the tools to monitor it, because they don’t exist? They exist with
Agentforce.
Colossally difficult to maintain – it’s a much different thing to
maintain the training you’re doing against AI versus these
managed packages.
Like cloud computing, a platform approach to agentic AI like Agentforce
enables organizations to focus their technology resources on fine tuning
and differentiation rather than day-to-day application support.

DIY months Agentforce months

Agentic AI phase
(average) (average)
Model setup 12.0 1.0
Data integration 3.5 0.3
Prompt engineering 12.0 1.0
Guardrails 18.0 0.0
UI and app/workflow development 6.0 1.0
Tuning 24.0 1.6
Total 75.5 4.8

Valoir conducted in-depth interviews with more than 20 Salesforce

Agentforce pilot customers in various industries to determine the
average number of months they spent working on each phase of their
agentic AI projects both on a DIY basis and with Agentforce. We found
the average months spent on DIY projects was 75.5 and the average
time needed to bring an Agentforce project to productive accuracy was
4.8 months – making it 16 times faster.

Page 10 Copyright ©2025 Valoir Inc. All Rights Reserved

Agentic AI has the potential to deliver significant benefits by automating
complex tasks and interactions without human involvement. However,
early experiences have led technology teams from FOMO to fear of
messing up (FOMU), and creating agentic AI systems from scratch has
been prohibitively expensive and often unsuccessful. Even with
extensive testing and tuning, achieving the reliability needed for agentic
AI has remained elusive.
In contrast, a platform approach that enables individual organizations to
leverage vendor investments in model setup and tuning, data
integration, prompt engineering, guardrails and security, UI and
workflow development, and low-code tuning and testing – like
Agentforce – will drive broader adoption of agentic AI by lowering
technical and financial barriers and making the benefits of agentic AI
accessible to all organizations.

Valoir is a technology analyst firm providing research and advisory services with a focus
on the value technology delivers. With deep expertise in CRM, HCM, customer and
employee experience, and enterprise applications, Valoir helps clients understand and
maximize the value of technology. For more information, contact Valoir at
www.Valoir.com or 1-617-515-3699.

Page 11 Copyright ©2025 Valoir Inc. All Rights Reserved