Not Every AI Use Case Needs a Chat Window
Visible Agents vs Invisible AI
If you hear “AI”, the immediate thought that may come to mind is an Agent, a chat window, a user typing prompts and getting answers.
That’s one form AI can take, and a powerful one. However, it’s not the only one, and most of the time it’s not the right one.
The shape that gets less attention and typically gains better adoption is AI in the backend. In the Salesforce ecosystem, this means a Prompt Template within a Flow, or even a piece of Apex that calls an Agentforce agent via the Agents API. The choice between these two shapes matters more than the choice between vendors or models. The end user sometimes doesn’t even know that AI was involved in the process, and this is a big win for some organizations fighting with adoption.
When The Agent Should Be In Front
Any time a conversation is expected (back and forth Q&A) then an Agent approach wins.
For example, a new hire opens the Onboarding Agent and asks how to get reimbursed for expenses. The Agent walks through the steps, and the user can ask follow-up questions. A few Q&As later, your new hire knows what to do.
That’s a real use case. The user benefits from the dialogue, not just the answer. The interface earns its keep because the work is inherently iterative. Sales coaching, internal help desk, complex troubleshooting… anything that needs more than one round of interaction before it can produce a useful response. These are Agent shapes.
When Agentforce Belongs in the Backend
Not everything that makes a good AI use case deserves a front-end agent. It’s not the answer you normally hear these days, but most of the AI value in an organization isn’t conversational. It is classification, summarization, routing, scoring, and extraction. The user doesn’t need to talk to anything.
Three examples from real work:
SOW summarization in Slack: When an Opportunity closes, a Flow pulls the SOW PDF attached to the project, passes it through a Prompt Template, and posts a clean summary into the relevant Slack channels. The Project Manager did not talk to an Agent. They might not even be logged into Salesforce. They did not ask anything. They got named PMs for a new Project, and the summary just appears where the delivery team is already working.
Group routing on a screen Flow: The user describes the kind of support that they need in their own words using a field on the screen flow. A prompt called by the same Flow classifies the request and routes it to the right group. For the user, nothing changed: they see a screen Flow that looks identical to any other one. The Prompt Template is invisible to them → instant adoption.
Approval classification. Before an approval request goes to a human, an Agent reads the record plus its related cases, account activity, and engagement signals, and decides whether the approval needs a human at all (most don’t!). The approver sees fewer rubber-stamp requests in their queue, and they never see the AI step.
Three different problems and objects. But in all three, AI is doing real work. In none of them does the user interact with an agent.
Why Invisible AI Adopts Faster Than Agents
If the technology is there, why do we still see such a high rate of failed/stalled projects? Adoption is the actual bottleneck for Agentforce adoption, not the technology. The model works, the prompt works, but the problem is getting humans to change what they do every day.
The agent in front implies a behavioral change. It adds a new interface to learn, a new place to go, and a new mental model. Even when the agent is great, there is the change management piece.
With Agentforce in the backend, the screen Flow still looks like a screen Flow. The Slack message just appears, or the approval queue just gets shorter. Users don’t have to adopt anything because there’s nothing new in front of them. All the efficiencies are 100% process-related. That’s not a small thing when it comes to change management and why placing agents in the backend can be a strategy for agentforce adoption.
The Trap And How to Decide
These days, all vendors are rushing to add agents to their products, all conferences focus on demoing agents, and all the screenshots in that marketing deck have chat windows in them.
When thinking about the next Agentforce use case, ask yourself: Does the user benefit from a dialogue, or just from a better outcome?
If the answer is dialogue, build the agent. The interface is the point in this case.
If the answer is outcome, put the AI in the backend. The interface is friction, and might be the difference between a successful implementation and a failed one.
Most of your AI use cases are the second one. They’re not the sexy ones, but they work, and they get adopted.
Now, think about the Agentforce use cases your organization tried to implement, but the team didn’t adopt. Or the ones that never made it onto the POC list because nobody could picture them as an Agent interface. How many of them were backend problems wearing the wrong shape?
