A calmer week for AI. Though anything would look calm compared to a non-stop barrage of “AI is going to kill us all” headlines.

Ready for another interview with someone building the future of AI?

-MV

What’s in store:

Read Time: 6 minutes

AI THOUGHT LEADERSHIP

10 Questions for AI Leaders: Shawn Goodin

Shawn Goodin's job is to make AI useful for customer success at scale. His biggest concern isn't that AI isn’t capable enough - it's that it's too confident.

The HubSpot principal AI engineer thinks the best agentic systems are honest enough to say, "I might be wrong here." The problem is that most aren't built that way - yet.

Mindstream: When you think about AI's current trajectory, what excites you the most? Does anything concern you?

Shawn Goodin: What excites me is that AI is finally good enough to do the actual work, not just demo well. For years, these tools could write you a tidy paragraph, but you couldn't trust them to take real action on a real customer account. That's changing fast, and it means a tiny team can suddenly do what used to take a whole department.

What worries me is the same power pointed the wrong way. It's easy to build something that quietly boxes people in: you label a customer once, "they're a marketing user," and then every recommendation you ever send reinforces that label, so they never discover the other 80% of what they're paying for. Done badly, AI doesn't open the world up. It shrinks it to the size of its first guess about you.

Explain your AI philosophy in a single sentence.

An AI that can act on its own isn't the same as one that understands what it's doing, and we keep paying for the first while calling it the second.

How do you personally decide what to automate and what not to automate?

My rule of thumb is to automate the reach and keep the judgment. A computer can watch an entire customer account at once and catch the thing no human would ever spot, like the 500 people you're about to email, 127 of whom are already in live conversations with support.

That's reach, and you should hand it to a machine. Whether to pause the campaign, though, or how to handle a customer who's frustrated for a reason that isn't in any database field, that's judgment, and it stays with a person. The mistake I try not to make is automating the decision when I should only be automating the legwork.

Let the agent tee up the call. Let the human make it.

What can a great engineer do today that an AI still can't compete with?

AI is incredible at producing things. Give it a clear problem, and it'll hand you ten solutions before you've finished your coffee. What it can't do is tell you which problem is worth solving in the first place.

The hardest part of my job was never writing the code; it's reading a messy, half-described situation, knowing which thing really matters to the business, and having the taste to throw out the clever solution that answers the wrong question. AI can search the space. It still can't tell you which space is worth searching.

Are you worried we're in an AI bubble?

A little, but I'd split it in two. The froth is everybody slapping "AI" on a product and selling it like it thinks for you, when under the hood it's mostly fancy autocomplete with a big marketing budget.

That part washes out. What's real is the boring stuff underneath. Wiring up good data and giving these systems actual context is what makes them useful, and that's not going anywhere. So the technology isn't a bubble. Some of the money chasing the story around it definitely is.

You're embedded in a customer success org, but report to an AI engineering team. How do you keep both sides happy?

That tension is basically my whole job. I sit between a customer success team that needs results and an engineering team that needs things built right, and on any given day those two don't want the same thing. The way I keep both sides happy is that I won't let them optimize separately.

Right now we have something like eighteen different AI projects running across the org, and almost every one is quietly rebuilding the same thing: its own way of understanding the customer.

My push is to build that understanding once and let everyone share it. When that works, customer success gets smarter outreach and engineering stops reinventing the same wheel eighteen times, so it's the same win for both sides instead of a tug of war.

What's one small decision you've made that had an unexpectedly large impact?

The one that comes to mind wasn't building something; it was deciding whether to keep expanding a perfectly logical idea or stop and ask whether the premise was even right for our customers.

We were about to build a model that sorts every customer into a persona bucket, the obvious thing everyone expects you to do. But I kept watching real customers switch roles three times before lunch, so I pushed to pause and question the whole premise.

That pause is what turned persona into an intent graph: a context layer that's constantly listening and learning. And that's the shift from batch communication to real-time engagement, from treating a customer as a label to treating them as a human.

What does a well-designed agentic system look like to you compared to a badly-designed one?

AI's biggest problem isn't that it's dumb. It's that it's confident. An LLM will give you a totally wrong answer with the same straight face it uses for the right ones. It's the coworker who's never once said: "I'm not sure."

That breaks trust two ways. If I'm building on it, I can't lean on something that fails silently: it doesn't throw an error - it just ships the wrong thing and lets me find out later. And if I'm the customer, the first time I catch it being confidently wrong, I'm done believing the rest.

So a well-designed agent is really just one that's honest about itself. It checks its own work, knows when it's out of its depth, and will actually say, "I might be wrong here." The bad ones are built to sound sure and look finished. The good ones earn confidence instead of faking it.

Which recent AI breakthrough made you rethink something fundamental?

I don't think it's one breakthrough; it's the evolution and integration of reasoning, tools, skills, memory, loops, and harnesses that are resilient and adapt in real time. I think the LLMs are getting incrementally better, but the entire infrastructure built around them is what is transforming them into true autonomous agents you begin to trust.

Finish this sentence: In five years, AI will make people feel ___ about their work. Why?

In five years, AI will make people feel more capable about their work. Not replaced, not obsolete, more capable.

The honest truth of most jobs is that a big chunk of the day goes to busywork and to not knowing what you don't know. If AI takes the busywork and, more importantly, shows you the thing you'd never have found on your own, the opportunity sitting three clicks deep that you didn't know to look for, then the ceiling on what one person can do goes way up.

The best tools don't tell you who you are. They expand what you can reach.

Shawn Goodin is a Principal AI Engineer at HubSpot responsible for the Digital Success Agent program, using multi-agent systems to expand the scope of Customer Success Managers, ensuring every customer gets meaningful personalized value out of their HubSpot instance. He also writes a Substack on transformation strategy. 

You can get our complete guide to AI agents below!

POLL OF THE WEEK

Full results will be at the bottom of tomorrow’s newsletter!

THIS WEEK IN AI

Satellites, rogue agents, and a crocodile that never existed

Google launched a refrigerator into space. An OpenAI agent hacked Australia's Medicare portal and sat on it for months. And a man in Singapore is facing prison time for a fake crocodile photo.

Just another week in AI.

Here's what mattered:

AI Art

Our Image of the Week

Daily Image Prompt

Submit your artwork to Mindstream →