3 expert insights on building AI features successfully 

To help you incorporate AI agents—and AI more broadly—into your product, we asked product and business leaders for advice based on their own experience.

You can read on to see what they shared.

Ignore the flashy AI ideas

Matt Schmidt, the CEO and Co-Founder of Peoplelogic, a leader in AI for HR, naturally felt pressured to build AI features that could garner press mentions, accolades from business leaders, and impress customers and prospects. 

Giving into this pressure, however, wasn’t in the best interest of his users and, by extension, his company. 

Matt and his team took some time to understand the most impactful ways that AI could impact the business. This involved analyzing their product usage data, talking to customers about potential ideas, and testing and validating agentic use cases with design partners.

This led them to launch Nova, a suite of people ops AI agents that can perform all kinds of HR and recruiting tasks, from sharing timely insights with interviewers on candidates to analyzing employee survey responses and sharing actionable feedback.

Matt’s top takeaway is that users aren’t looking to use AI agents in your product as soon as possible. They’re just looking for frictionless, valuable experiences. 

According to Matt:

"We live in a hype-filled echo chamber that our customers likely aren’t part of. That disconnect creates pressure to build flashy AI features that customers aren’t asking for. Ignore this pressure and build AI features that deliver real value.”

https://www.merge.dev/blog/ai-product-strategy?blog-related=image

Robust fallbacks are as important as innovation

As the team at ExpenseOnDemand, which offers a full-suite expense management solution, began incorporating AI into their receipt scanner functionality, they realized that the AI wasn’t consistently interpreting the receipt data correctly.

This led them to realize that data quality, variety, and volume were all table stakes for training this AI feature effectively, whether that’s providing handwritten notes, foreign formats, low-resolution images, and so on.

Also, even with the best training data, scans could still be inaccurate (albeit at a much lower rate). As a result, their team invested in a powerful error-handling workflow for any scan that wasn’t 100% accurate. 

After going through this journey, Sidd Nigam, the Director of Product and Strategy at ExpenseOnDemand, has the following takeaway:

“Every AI feature will inevitably fail. Designing robust fallback mechanisms to ensure graceful failure is just as critical as the feature’s capabilities.”‍

AI needs to fit organically with your customers’ workflows

Tushar Makhija, the CEO and co-founder of TeamOhana, a headcount management platform for enterprise companies, had a clear idea of what an AI agent is and what it needs to do for their business from the get-go:

“AI agents need to be embedded in and tightly aligned to enterprise workflows. If your AI agent isn’t saving your team time or taking work off their plates, it’s not an agent—it’s a widget.”

This led them to build AI agents that fit into and up-level existing use cases organically. 

For example, as users review different headcount scenarios in TeamOhana, an AI agent can notify them if and when certain roles need to be pushed back and by how many days (the AI agent uses historical context on the time to fill those roles).

https://www.merge.dev/blog/ai-agent-integrations?blog-related=image

Final thoughts

There’s no right way to build AI agents—and AI-powered product features more broadly. 

But if you follow the advice above—prioritizing your customers’ wants and needs, implementing robust fallbacks, and incorporating AI into key existing use cases—you’ll be well on your way to building valuable, differentiated, and durable AI features and capabilities.

{{this-blog-only-cta}}