Irina Gutman

No, thank you for bringing it up. It is a very, very important topic. When we're designing agents, we have a dedicated and explicit exercise called guardrails and risk management definition. What we're advising our customers to do as they define the role of the agent and what job it needs to do is

No, thank you for bringing it up. It is a very, very important topic. When we're designing agents, we have a dedicated and explicit exercise called guardrails and risk management definition. What we're advising our customers to do as they define the role of the agent and what job it needs to do is

No, thank you for bringing it up. It is a very, very important topic. When we're designing agents, we have a dedicated and explicit exercise called guardrails and risk management definition. What we're advising our customers to do as they define the role of the agent and what job it needs to do is

We asked it to define risk in categories that are very familiar to them, people, process, data, and technology. So we work with the customer to list all the possible risks, all the possible violations that this agent could cause in those four categories. And then we take each of those items and we say, okay, how are we going to mitigate it? Are we going to mitigate it with people?

We asked it to define risk in categories that are very familiar to them, people, process, data, and technology. So we work with the customer to list all the possible risks, all the possible violations that this agent could cause in those four categories. And then we take each of those items and we say, okay, how are we going to mitigate it? Are we going to mitigate it with people?

We asked it to define risk in categories that are very familiar to them, people, process, data, and technology. So we work with the customer to list all the possible risks, all the possible violations that this agent could cause in those four categories. And then we take each of those items and we say, okay, how are we going to mitigate it? Are we going to mitigate it with people?

Are we going to mitigate it with process, with data, and technology? Yes, the risk is inherent. There's some variability, vulnerability inherited in AI. But it does, as you said, it's not an option. This technology is here to stay. So what we're doing is developing framework, just the one that I explained to you, of how companies can go systematically online.

Are we going to mitigate it with process, with data, and technology? Yes, the risk is inherent. There's some variability, vulnerability inherited in AI. But it does, as you said, it's not an option. This technology is here to stay. So what we're doing is developing framework, just the one that I explained to you, of how companies can go systematically online.

Are we going to mitigate it with process, with data, and technology? Yes, the risk is inherent. There's some variability, vulnerability inherited in AI. But it does, as you said, it's not an option. This technology is here to stay. So what we're doing is developing framework, just the one that I explained to you, of how companies can go systematically online.

from defining an agent to defining risks, to defining risk mitigations, and then turning them into what I referred to earlier in our conversation as guardrails, which is instructions that we give to the agent what it's not allowed to do. The other thing that we do to make sure we do stress testing for bias and toxicity.

from defining an agent to defining risks, to defining risk mitigations, and then turning them into what I referred to earlier in our conversation as guardrails, which is instructions that we give to the agent what it's not allowed to do. The other thing that we do to make sure we do stress testing for bias and toxicity.

from defining an agent to defining risks, to defining risk mitigations, and then turning them into what I referred to earlier in our conversation as guardrails, which is instructions that we give to the agent what it's not allowed to do. The other thing that we do to make sure we do stress testing for bias and toxicity.

Salesforce already has a trust layer that is very, very robust and pretty much nothing can get in past us, nothing bad from the outside. However, we put agent through additional stress of bias and toxicity testing. When we overload the system, we talk to the agent in bad language, profanities, using racist slurs on purpose.

Salesforce already has a trust layer that is very, very robust and pretty much nothing can get in past us, nothing bad from the outside. However, we put agent through additional stress of bias and toxicity testing. When we overload the system, we talk to the agent in bad language, profanities, using racist slurs on purpose.

Salesforce already has a trust layer that is very, very robust and pretty much nothing can get in past us, nothing bad from the outside. However, we put agent through additional stress of bias and toxicity testing. When we overload the system, we talk to the agent in bad language, profanities, using racist slurs on purpose.

It's not a pretty testing, but it's absolutely necessary because as you said, nobody wants to be a headline for the wrong reason. or have a lawsuit. So that type of testing is absolutely critical. And my responsible AI team developed a methodology of how to do that kind of testing. And we set thresholds really low, like 1% or less than 1%. And I'm proud to say that our agents passed.

It's not a pretty testing, but it's absolutely necessary because as you said, nobody wants to be a headline for the wrong reason. or have a lawsuit. So that type of testing is absolutely critical. And my responsible AI team developed a methodology of how to do that kind of testing. And we set thresholds really low, like 1% or less than 1%. And I'm proud to say that our agents passed.

It's not a pretty testing, but it's absolutely necessary because as you said, nobody wants to be a headline for the wrong reason. or have a lawsuit. So that type of testing is absolutely critical. And my responsible AI team developed a methodology of how to do that kind of testing. And we set thresholds really low, like 1% or less than 1%. And I'm proud to say that our agents passed.

Amazing. Just for anyone who's thinking about utilizing, which we all should be, how are we utilizing agent technology, agentic AI? We have to have those conversations. What are the risks at play? Just put them all out on the table. Let's make sure we're thinking about the edge cases because they are there.

Amazing. Just for anyone who's thinking about utilizing, which we all should be, how are we utilizing agent technology, agentic AI? We have to have those conversations. What are the risks at play? Just put them all out on the table. Let's make sure we're thinking about the edge cases because they are there.