10 Questions for AI Leaders: Paula Goldman

Paula Goldman came to AI via anthropology, which probably explains why her take on it is sharper than most. Salesforce's Chief Ethical and Humane Use Officer and bestselling author of Manage the Machine has been in AI governance since before it was a job title.

In our interview, she covers why accountability is broken, what good oversight actually looks like, and why she thinks AI might make us more connected to each other, not less.

Paula Goldman: Less of a specific belief, but I initially underestimated the velocity and significance of AI improvement. Even though I spent a lot of time projecting forward to what work would be like with powerful AI agents, it was easy to over-anchor on some of the early shortcomings of AI (like an inability to solve basic math questions). 

Many of those obvious gaps closed faster than I expected. There are still real limitations, but they can be less glaring.

On the one hand, that makes many applications more reliable. On the other hand, it makes mistakes harder to catch, because when answers are mostly right, we can stop watching for wrong ones. 

AI is not an oracle, nor is it human intelligence; getting the best from it means actively managing it in service of helping people do bigger and better things. 

I wrote Manage the Machine for precisely this reason. Both extremes make it seem like AI is some inevitable force descending upon us, something over which we have no control. The reality is both more ordinary and also more hopeful. For all its power, AI still has flaws, and how it plays out depends on us…on the choices we make about how to use it.   

I was going to say judgment or relationship-building. But I will pick management. Now, or soon, most knowledge workers will essentially be managers of multiple AI agents, and thinking of ourselves as such is essential.

This means knowing what things to delegate to AI and which to keep for people, giving clear instructions with the right amount of context, preserving critical thinking and expertise to be able to judge AI recommendations, and staying accountable for results. 

We tend to do a bad job preparing leaders to manage people. Many new people managers are promoted because they were strong individual contributors, without training or instruction on the complexities of leading teams.

If most knowledge workers are about to become managers of AI, we shouldn’t repeat that mistake. AI management should be taught and practiced like any other skill. 

I can’t prognosticate the market, nor is that my expertise. I do believe that we’ve barely begun to scratch the surface of what existing AI capability can do to improve business outcomes. And we can count on continued improvement in the raw capability of AI as well.

Markets will wax and wane, but I’m pretty confident that we are in the early stages of a very significant transformation in how work gets done. 

When I first took my role, I had few equivalent peers in the industry. In the last few years, I’ve seen companies (both tech and non-tech) create similar functions. I’ve also seen governments add staff with technical understanding of AI issues, and of course much more widespread public attention to the topic. However, it still feels like early days.  

One of the things that makes AI governance challenging is that AI is everything, everywhere, all at once. In the old days, AI models were used for one specific thing, like predicting customer churn. Now the same models can be used in nearly every domain of knowledge work.

That makes the past approach of reviewing each model for a single intended use much harder to apply. A model may be great for drafting marketing copy but require significant safeguards when employed to sift through mountains of evidence in legal discovery.

So governance increasingly has to happen at the point of use, and responsibility is shared between companies that build these systems and the organizations that use them. 

Perhaps not small, but: deciding to become an anthropologist. Not a choice initially received by my parents as a sure-fire path to success and stability. But it profoundly changed the way I think about technology and business strategy.

I’ve always been inspired by innovation, but when you pay equal attention to the people side of things, it opens up a clearer picture of how tech gets used and the choices that make it succeed or fail. That perspective has shaped my career trajectory, from Omidyar Network to Salesforce (my parents have since come around, too).

Some of it is just common sense. Take the question of accountability. You wouldn’t hire a new employee and not give them someone to report to, or open a retail store and not put someone in charge of it. But when it comes to AI, we’ve been very unclear about how accountability should work.  

The models I’m seeing evolve are matrixed ones: at the tool level, a named individual who is responsible for its effectiveness (that might be someone in IT, or it might be, for example, someone in marketing for campaign-specific tools). And at the output level, individuals who use the tool are responsible for their own work. 

Some of it is about defining in advance where you don’t want AI to make judgment calls. In customer service for example, that might be circumstances that are sensitive, require legal judgment, touch a lot of money, or fall into grey areas of company policy. 

And one of the most important parts is about how we preserve the ability for people managing AI to be able to exercise judgment and creative thinking. I vibe-coded my book website, for example, and I got my settings wrong so that my agent asked me for permission every single time it accessed a website. Naturally, I started immediately clicking yes without thinking.  

And that’s what happens when AI is usually, but not always, right. We stop pausing before hitting approve. Part of the answer is inserting a little bit of what we call “mindful friction” into the tools. When it really matters, you want the person accountable to stop and think.

Another part is about creating spaces for people to think before they delegate to AI in the first place – for example, with innovation tasks, initial brainstorming and ideation is often helpful before asking AI to expand on ideas. 

I’ve been watching self-driving cars with a lot of fascination. In the last few years, the expansion and uptake have been breathtaking. At the same time, new edge cases keep emerging, like the time when San Francisco had a blackout and Waymos stalled at dark traffic signals.  

I used to think that edge cases would make self-driving cars significantly slower to market. But of course, with time, such edge cases turn into known patterns that AI can manage. What was once unusual becomes routine. 

Does this mean that eventually all the edge cases will be figured out and AI will be able to handle all forms of human work, as some would have us believe? I still don’t think so, for many reasons. I’ll share two here.

One is that AI itself is accelerating the pace of change in the world, which keeps producing situations that no one has seen before, from new kinds of fraud to new ways of organizing work. As AI learns how to handle yesterday’s edge cases, it’s helping to create new ones. Another is that interpersonal interactions are a different kind of problem space than traffic, which runs on explicit rules (even if those rules sometimes get messy).

Connected to each other. Admittedly, this is far from inevitable. But some of my favorite AI use cases in the book were the ones that helped people be more present for each other.

The AI coach that helps a manager practice for the hard conversation that opens up someone’s career trajectory, instead of avoiding the conversation all together. The AI agent that allows the salesperson to be fully present in a meeting instead of spending time on lead qualification or research. The employee knowledge agent that doesn’t just answer questions but suggests connecting with teammates who can offer more insight. 

There’s a lot of well-deserved concern right now about AI increasing isolation at work.  But it doesn’t have to be this way. A lot of it comes down to how we design intentionally for connection. 

Paula Goldman is the Chief Ethical and Humane Use Officer for Salesforce, where she leads the company in developing and deploying frameworks to ensure trust keeps pace with technology. She is the bestselling author of Manage the Machine: How to Harness Human-AI Collaboration at Work. 

You can get Paula’s new book here!