Home AI Prog IO: After Years of Learning, Microsoft’s AI Assistant Is Ready For The Copilot Chair

Prog IO: After Years of Learning, Microsoft’s AI Assistant Is Ready For The Copilot Chair

SHARE:
Chat bot, using and chatting artificial intelligence chat bot developed by tech company. Digital chat bot, robot application, conversation assistant concept. Optimizing language models for dialogue.

Microsoft will be the first to tell you that successfully implementing new technology means learning from your mistakes.

That’s what Microsoft spent the last several years doing since the launch of Bing Chat, their first AI assistant in 2023. Their most recent iteration of Copilot, described as an “AI companion,” is a far cry from Bing Chat – and definitely for the better.

What’s wrong, what’s right

Microsoft’s early-stage AI assistants weren’t nearly as complex or effective as today’s iteration. Bing Chat functioned similarly to a search engine, according to Paul Longo, Microsoft’s general manager of AI in Ads, who elaborated on the topic during a session on Wednesday at AdExchanger’s Programmatic IO in Las Vegas.

Users would pop in, look up their search query, then close out of the tab. Over time, Longo said, Microsoft developed a more personalized search companion, Copilot, which, as of April, now features stored memory. And with memory comes massive improvements – namely, a learned awareness of users’ likes and interests and the ability to refer back to prior conversations.

Initially, ads were only incorporated into Copilot during “commercially relevant” conversations, Longo explained. And in many ways, this proved successful. Microsoft saw higher click-through and engagement rates along with better conversions.

All of that sounds great. So what went wrong?

Consumers thought “the entire experience was an ad,” said Longo. They weren’t seeing a clear distinction between organic versus paid results.

But Microsoft didn’t view these blips as setbacks or failures. Rather, they paved the way for a better tool.

“We stubbed our toe,” said Longo. But that uncomfortable experience led to new standards for future products.

“We learned a lot,” he said.

The principle of the thing

Microsoft came up with three guiding principles to determine the best way to incorporate ads into conversational AI experiences.

One is that sponsored placements should never affect Copilot’s organic results, ensuring that the AI companion remains a “trusted, unbiased advisor,” as Longo put it.

Copilot ads are also clearly labeled, so the user is clear on what part of the display is an advertisement as opposed to an organic result (in direct contrast to the aforementioned toe stub).

Lastly, ads are only delivered in what Longo described as “highly relevant conversational scenarios,” where an “ad voice” can explain to the user why they are receiving that particular ad.

Microsoft has seen performance improvements since introducing these principles. But improving the ad experience is just part of a larger goal for Microsoft. It wants its products to gain adoption. And Longo said these changes help the company improve its consumer trust levels.

Consumer confidence in the ad experience actually matches confidence levels in an ad-free experience, he said.

And that’s because the user is part of the process, and the human is the one who generates results.

That necessary level of human control and knowledge is why the best AI solution is implemented as a copilot, Longo said. There’s a reason the product isn’t called Autopilot.

Must Read

Apple Has Far-Reaching Plans To Block Hundreds Of Programmatic Data Companies From iOS

Apple’s WebKit crackdown appears to extend well beyond The Trade Desk, putting hundreds of ad tech, data and identity vendors on a mysterious, dynamically updated block list.

Josh Reed, Zoom's VP of brand and content, speaking at AdExchanger's Programmatic IO event in New York City (September 28, 2006)

Zoom’s Marketing Challenge Is That It’s Too Well Known For Its Own Good

Zoom has 99% unaided brand awareness, which sounds great on paper. But there’s a catch: Most people still think it’s just a video-call app.

Why Agencies Think They Shouldn’t Own Agentic AI Tools Or The Data Used To Build Them

Agencies are differentiating their tech stacks by building custom agentic AI tools for their clients. And they’re rethinking owning those AI tools – particularly since licensing them creates new revenue streams.

Privacy! Commerce! Connected TV! Read all about it. Subscribe to AdExchanger Newsletters

Programmatic IO: Insurers Are Building Ad Tech’s AI Accountability Layer

Agencies and marketers discussed the future of AI governance at AdExchanger’s Programmatic IO NYC this week. The main takeaway? Expect insurers to play an increasingly important role in managing AI compliance.

Apple’s Latest Operating System Blocks The Trade Desk From Serving Ads On Safari

The Trade Desk is unable to serve ads to the Safari browser for Apple device owners that have downloaded iOS 27. Apple has been investigating the issue since last week.

Who Will Stand Up For The Open Web?

The open web is done, stick a fork in it. Banner blindness is near universal, search traffic has run dry and publishers are struggling for oxygen. But what if that’s … not true?