From chatbots to action: the future of AI agents with Jon Stahl

SPEAKER_1 (0:00)

College football is back, so Hilton called on me, the superstition concierge, to make your fan rituals a reality. Need a room to match your lucky number? We got you. Want to make sure our team doesn't wash your lucky jersey? Oh, that smells lucky. Hilton's unmatched hospitality can keep up with any superstition. Even a marching band wake up call at 555 in 55 seconds. Hit it! When you need a team that will do whatever it takes on game day, it matters where you stay. Hilton, for the stay.

booking.com is the easiest way from a day surrounded by noise.

To a stay surrounded by nature. That's nice. Go on, book it. It's easy. booking.com. Booking.yeah.

Jon Stahl (1:00)

I like thinking of it as the world's smartest but least experienced intern.

Owen Raszkiewicz (1:05)

What does an AI agent need before it can safely make a purchase?

Jon Stahl (1:09)

A verified identity and mandate. Who is this agent? What human are they acting on behalf of? And what has the human actually granted them permission to do? How much money can this agent spend? And that has to be cryptographically signed and verifiable. The last time I wrote any code, it was the last century.

But in the last six months, I have written software to get my job done.

Owen Raszkiewicz (1:34)

Hey there, here's a quick note. This podcast contains general financial information only. That means it's not specific to you, your needs, goals or financial objectives. So don't act on the information until you've spoken to your financial planner. You'll find our full disclosure, disclaimer and link to our financial services guide in the show notes.

Welcome back to the Business Stack of the Future. This is episode 2 brought to you by Stripe. Today we're talking about when AI agents stop just answering questions for you and actually do the work. My guest for today is Jon Stahl, the Director of Product Management at Salesforce, who's deep in the weeds on all the commerce sides of AI. We talk about how and where AI agents can execute decisions or make changes using your business systems, and what needs to happen before these agents start taking or making payments on your behalf. We talk about where AI agents are creating real value today, some of the mental models that you can use to think about their capabilities and limitations. Finally, we discuss just what Jon thinks is going to feel normal in five years time, that today may be just something of our imagination. Let's get into it.

What are people already using AI to do inside a business?

Jon Stahl (2:53)

Oh, yeah. So obviously, they're doing a ton. The most interesting thing that I'm seeing is we're moving from the era of a few, just a few months ago, where the AI agents are getting really good at answering questions. That's really valuable. We're now shifting to the world where not only can they answer questions, they can actually get work done, which is even more exciting. The agents can now execute real multi-step tasks against your real systems of record, like your CRM or other business systems.

And that's when it starts to get pretty exciting. We see that across a bunch of different areas where I sit in the world of commerce. We see it in terms of post-order servicing, where agents can help understand order status, can initiate and even execute returns and exchanges.

If you have a service appointment, they can help you change your appointment or create an appointment. And they know how to escalate to a human when they need to or when they hit an edge case. So moving from systems of information to systems of action is kind of the big theme that we're seeing.

Owen Raszkiewicz (4:03)

It's such an interesting and fascinating time. Is that where you see agents creating the most value today?

Jon Stahl (4:11)

I think that's where I would say the cutting edge is. I think the most value, my sort of personal stack ranking, and I'm sure everyone would have a sort of a different take on that would be. I would say the customer support and service use case is probably number one. That's, I think, where a lot of the early efforts started and where we've kind of gone the farthest in the journey. They also have a bunch of qualities that make those kinds of use cases very amenable to agentic acceleration. They're very high volume. They're very repetitive. They have really clear success criteria. You can really measure. You can measure things like deflection. You can measure cost savings. You can measure increases to customer satisfaction scores, and so all of those together make that support and service use case. There's obviously a ton of use cases inside of that, but that broad area, super high value, super tractable. The very close second and one that's very close to my heart is not going to surprise anybody. It's that coding and software engineering use case. This is where it started. It's not surprising to think that a bunch of nerds could be really good at building tools for a bunch of nerds. We're very good at that.