Agentforce Q&A: What you need to know about Salesforce’s new AI agent platform

Customer relationship management software giant Salesforce this year went big on AI agents, pitching them as a more sophisticated version of copilots, including Salesforce’s own Einstein Copilot service.

Agentforce, a platform that allows Salesforce customers to quickly spin up AI agents and operate them on an ongoing basis, received its general availability release on October 29 and has been positioned by Salesforce founder Mark Benioff as integral to the future of the company.

New Zealand customers including Fisher & Paykel, Parkable, Media Works, realestate.co.nz, McLeod Cranes and Stats NZ have been using Agentforce

ITP Tech Blog caught up with Rowena Westphalen is Vice President, Innovation, Asia Pacific at Salesforce to find out more about Agentforce,

More details about Agentforce and training resources are available here.

Agentforce has no-code templates and flows to help you build your own AI agents.

What is the difference between Einstein Copilot and Agentforce?

At Dreamforce, we launched Agentforce, which is kind of the next step on top of a copilot. It’s agents that autonomously assist with specific tasks that that you need to do. The obvious use cases are things like automating a marketing campaign, engaging a sales prospect, or deflecting fairly routine customer conversations and troubleshooting queries.

In the service space, it gets into the self service stuff that a customer would do, but in a much more interactive way. There are two particular elements of Agentforce that I think are exciting.

The first is just the speed at which customers are getting value. The level of accuracy and the speed are compelling and exciting. Secondly, when we launched Einstein Copilot it was very much about having a human in the loop. 

But actually, there are areas that can be automated, that don't need a human. With Agentforce we have guardrails built in, that comes out of the box, that's been an incredibly interesting value proposition.

To what extent is an Agent fully automated, and when will a human be brought into the loop?

Instead of a human kind of helicoptering over the agent, the agent is like okay, in this context, you’ve raised something outside of my guardrails. I’m going to contact a human. 

Rowena Westphalen, Salesforce

“You can get that productivity gain and efficiency but still have that kind of intimacy at scale.”

At Salesforce, our support website now has a channel that's entirely powered by Agentforce. One of the guardrails we put on the place at the moment is that as soon as someone asks about a pricing or licensing question, we don’t just automate that. That's when it's actually a useful conversation to have with a human so that agent knows to point it to a human agent and then we've got a seamless connect into the contact center.

How does the Atlas reasoning engine play a role in Agentforce?

Atlas has come out of our research department. It’s basically the planner part of Agentforce. In the early versions of Einstein Copilot and Agentforce, we had a fairly unsophisticated planner, so it was kind of limited what it could do. 

But with Atlas, we're seeing both the speed and complexity to do some really interesting things, particularly in terms of complex queries or sequencing elements together.

What are some of the key use cases New Zealand customers are exploring for Agentforce?

For Fisher & Paykel, 30% of their customer queries are now being handled by agents. So that’s 30% fewer queries that their contact centre has to handle. It's everything from getting my oven serviced to product queries. 

They’ve calculated that 3,000 to 3,500 hours a month are being saved by Agentforce. The early agents were like, this is a service agent, and this is a sales agent. At what point does the business development representative (BDR) become the service agent? That’s where Atlas is incredibly interesting, kind of switching across those different streams.

Agentforce has a consumption-based pricing model. At Dreamforce, CEO Marc Benioff talked about a benchmark of $2 per transaction. What are you seeing in terms of the cost effectiveness of Agentforce?

It feels like definitely the right price metric, $2 a conversation, compared to the time and effort it takes to onboard a person and keep them in place. Atlas isn’t in general availability for all customers yet. We’ll have to see as they start consuming more data how that may change things. 

But so far it certainly feels like it's the right sort of pricing.

How are you seeing Agentforce used to improve productivity in organisations?

As an example, One New Zealand is taking a very agent-first approach to everything that they're doing. They're using Marketing Cloud, Data Cloud and agents to really simplify the end to end journey. 

As that emerges, its going to really deliver personalised product and content recommendations. You can get that productivity gain and efficiency but still have that kind of intimacy at scale.

What has been done to reduce the potential for agents to produce spurious information or even have hallucinations?

Part of the accuracy level is that the hallucination factor is very low. The Atlas engine uses a very advanced form of retrieval-augmented generation (RAG) called Ensemble RAG. Because it’s grounded in such rich data, the more data you give the model, the less likely it is to have a hallucination. 

In Salesforce Data Cloud, we have what's called a vector database with semantic search. As we access data from external data sources and from within Salesforce itself, we're chunking it down into the semantic meaning of what the what the term means. 

There are two ways in which that is very powerful. The first is where there are nuances in similarly defined things. We can cut that short very quickly. But it also works really well with unstructured data where there's lots of noise in the data because it's extracting the actual meaning.

So it means that the agent has this underlying understanding of what you're meaning. It's not having to interpret the English language. It's actually being grounded on the meaning. I believe that's what's causing a lot of the lower hallucination capability.

Data Cloud seems very important to Agentforce, but does all of your data need to be in Data Cloud to use Agentforce?

Data Cloud is an important foundational layer, almost acting like a gateway and an enabler. But it doesn't have to be the place where the data persists. We have strong partnerships with organisations like AWS, Databricks and Snowflake. We have taken a very ecosystem approach.

We've also built in this abstraction layer between the choice of the large language model and the agent. I think this is very powerful because we're not forcing organisations to lock themselves in or make a decision about that. 

Marc Benioff advised attendees at Dreamforce not to take the DIY route to developing AI agents. What’s the real advantage of using a platform like Salesforce to create agents?

One of our customers did a benchmark. They were using Atlas, and they were comparing it to like an Azure OpenAI kind of DIY approach. Agentforce was 39 times faster to value than the DIY approach. It was 33% more accurate and twice as relevant. 

It can exist with copilots but if you have an investment in Salesforce it would be crazy not to let use our agent capabilities because they're so quick to market.

Now that Agentforce is generally available what should Salesforce administrators and others in the ecosystem be doing to get their heads around how it works?

They should go to trialhead.com. There are a few different trails that you can do to familiarise, yourself with Agentforce. There's one that gives you a basic overview, and there's one where you can you spin up a development machine and build your own agent.

This conversation has been abridged for clarity.