Instacart's Ayesha Saleem shares her 5 best conversational AI practices

Conversational AI is constantly changing—but in the creative hands of people like Ayesha Saleem, we can’t help but be excited for the future. In her impressive career, she’s become known as a builder and creative, taking non-existent conversation design (CxD) programs and building highly effective operations from scratch—first at Rocket Mortgage (formerly Quicken Loans) and now at Instacart.

But her tireless creative energy doesn’t stop with her own teams. Ayesha is sought out in CxD to speak at conferences and on podcasts, build comprehensive CxD guides, and mentor emerging leaders. And she’s shown no signs of slowing down.  

With these discussions, our goal is to discover exceptional people in conversational AI and find out how they landed in CxD, how they make sense of the chaos, and how their journeys could inspire your own. Today, Ayesha explains how she proves the impact of her team’s work to leadership (and what happens when you don’t), the five best practices she applies to every design project, and how she manages failure while redefining success. 

SB: Tell me about your conversational AI career journey. 

AS: I majored in computer science at the University of Michigan and did a bunch of software engineering internships—one of which was at GE, building for Alexa back in 2016. I got the chance to help launch their Alexa assistant when everyone was newly excited about the opportunities with conversation design. That really sparked my love of this space.

But I quickly realized engineering was wrong for me. I didn’t enjoy just implementing someone else's vision—it felt like the wrong stage of the creative process. I wanted to find a way to merge all my creativity and excitement around conversation design. That was when I was hired at Rocket Mortgage. 

My work there was unique because they had no clue what they needed. My title started out as “Cognitive AI Engineering Designer” and I remember thinking, “What? That's 10 different skills. What is that?” But it was the perfect role for me—a little bit of everything. I started building their conversation design operation from the ground up. It only took two years to define goals and success criteria, build a team of five designers, and move us out of spreadsheets and into tools like Voiceflow. And that opportunity led me to Instacart, where I’ve had the pleasure to do that all over again as a conversation design manager. 

How do you explain the value of your CxD work to leadership? 

Imagine you're talking to a design manager and you explain it’s taken you 30 hours to iterate on a design. That manager will understand that you’ve been playing around with different designs and thinking through how conversations should flow—you're in the creative process. But when I started my career, I was reporting to an engineering manager who had absolutely no idea what I was doing. He thought I was a graphic designer for a long time. For that engineering manager, he couldn't wrap his mind around that winding, iterative process. He understood progress as lines of code or projects launched. 

Now when I report to managers or leadership who may not understand conversation design workflows or even how my team spends our time, I report on two things to show how we work and why it’s valuable:impact and data. 

1. Impact. For example, while testing our designs we’d find people in the office to play the user. It was as simple as reciting the script verbally to one another. We collected their feedback so when it came time for reporting, we could explain the insights we gleaned from those tests and what we were working on to improve the project. 

Impact can be easy to prove once a product is launched, but if you’re showing that your creative process is already yielding quantifiable results—whether that’s 10 test interviews or a 10% improvement in performance—your leaders will be more likely to trust your process. 

2. Data. You cannot go into a room and say, "Hey, we're doing this because I said so." For people who are new to conversation design, they need some convincing. And you can't really argue with hard numbers. My foolproof way to prove our process is with A/B testing. When I started at both Rocket and Instacart with A/B testing, I managed to prove that our conversational systems were offering us huge cost savings as well as speed and time efficiencies. 

"For people who are new to conversation design, they need some convincing. And you can't really argue with hard numbers."

To this day, whenever I update our leaders on our latest progress or pitch new ideas, I always lead with the impact of what we’re doing right now and include real data. 

What happens if you’re not showing the impact of CxD on the business? 

I’ve worked with people who were more focused on the work itself and unsure how to quantify or explain their output. Sometimes the output of conversational AI doesn’t look like a typical 40-hour work week and that can get you into sticky situations with people who don’t understand our process. In this industry, if you’re not your own advocate, people aren't going to support what you're doing.

I've seen people's careers not move as fast as those who have mastered educating others on CxD or advocating for what they do. Or I've seen people burn out because they’re working hard but their efforts aren’t being acknowledged or supported by their management. 

It’s twofold—your leadership needs to understand what they've hired you to do and give you the agency to do it. And you should learn how to report on your impact, show your value, and educate your leaders and colleagues on conversational AI. 

What are your CxD best practices? 

I have five that come to mind, which can apply to many conversational AI use cases, both big and small. 

1. Design for your user. I am in the delivery space and I know my target demographic. We have to design for both older people and busy millennials who are working all the time. You need to understand who your user is and what they're trying to accomplish.

When I was working in the mortgage industry, we were designing for folks in the U.S. who were in their 50s—those were the people most likely to get a mortgage. So we were designing for people who were not comfortable with technology and who would opt to talk to a human as fast as they possibly could. That posed a challenge, but knowing that fact made us design better tools for them. 

Iterate on your personas, perform UX testing, and do user research to understand how comfortable people are with your designs. Much of that knowledge will affect your internal best practices as well—like never writing anything more complex than a sixth-grade reading level. 

2. Listen to your assistant speak. For voice interfaces, always, always, always listen to your content. Because if you're saying, "Sure, I will order that for you," as opposed to, "Sure, I'll order that for you," you may not realize that including a contraction makes it sound much less robotic. 

Lots of designers will write a conversation, design it in a Voiceflow or Figma, and then they'll publish it. They won’t even realize how weird it sounds. It's really important to hear the tone and how the system is saying each sentence. 

3. Be conscious of space limitations. In conversation design, you never have enough space. If you're designing a chatbot that lives on the corner of a page, you may get three lines. Tops.  

If you're designing an Alexa or Google Home store, you probably have people's attention for about 10 seconds before they've lost interest and gotten annoyed. Become very good at concision.

4. Focus on accessibility. If you're using a screen, make sure your colors are accessible and your font is big enough. Make sure your icons are legible. Accessibility is important. And unfortunately, a lot of people forget about it.

5. Ensure your script is inclusive and translatable. Write for the audience that you're speaking to. I've seen people release assistants in multiple languages and include puns. The second you translate this joke into a different language, it flops. 

Inclusive language is about more than avoiding offensive words, it’s about ensuring your designs don’t leave entire cultures and languages out of the conversation. 

What’s your approach to success and failure?

You should determine success criteria before you even start your project so everyone has a shared understanding of what success means and how we’re going to achieve it. Only from that place can we begin the process. Personally, I define insane success goals for myself and then I rarely meet them (laughs). I'm working on it! 

When you're designing a project, it's all theoretical. So if you push it out and it doesn't do as well as you want it to—which might be considered a failure—you lean into improving it from there. My favorite part is going deep and figuring out why it failed, what we can do right away to fix it, and what we can do better in the future. Honestly, I think it's exciting when things don't go your way 100%. There’s got to be space for iteration and improvement. 

I think the improvement process is gratifying because you can see how the levers you pull move you closer or further from success. By seeing success and failure as fluid, you have the freedom to try new things. You stop being so afraid.  

"Honestly, I think it's exciting when things don't go your way 100%. There’s got to be space for iteration and improvement."

Ayesha Saleem—the builder, the creative

“Failure is just an opportunity,” Ayesha says. At Instacart, she’s exploring how to integrate generative AI into their projects. This new evolution is another opportunity to build, fail, and try—that’s what excites her. “We don’t know all the value AI can bring yet. But we’re building the foundations right now.” 

In addition to the laundry list of projects she has on the go, Ayesha is currently working on her own farm-to-table vegetable garden. “It’s the creative in me, I have to have my hands in a little bit of everything.” 

Connect with Ayesha on LinkedIn


Conversational AI has a massive, UX-shaped hole

No items found.