Suggested reply

Designing AI-powered replies that help sellers communicate with confidence and convert more buyers.

ROLE

UX/UI, Prototyping, Ux research

THE WORK

A trust-building AI co-pilot for Fiverr sellers

01 / CHALLENGE

Sellers struggle to write clear, professional replies that convert buyers, especially across language barriers, while concerns about AI limit their trust and adoption.

02 / APPROACH

I worked on shaping an Inbox experience that made AI useful without taking over the seller’s role. The suggestions needed to be relevant to the conversation, sound natural, and stay within the seller’s actual service offering. Just as importantly, sellers needed to be able to review and edit every suggestion before sending it.

03 / OUTCOME

Suggested Replies brings editable AI-generated responses into Fiverr’s Inbox. The feature helps sellers communicate with buyers more clearly and professionally, while giving them full control over what they send. It was designed to improve buyer conversion and introduce sellers to AI assistance in a lower-risk way.

The challenge

A seller’s reply can shape whether a buyer decides to place an order. But replies may be unclear, poorly structured, or mismatched in tone—especially when sellers are communicating in a language that isn’t their own. Some sellers already use external AI tools to help write messages, but bringing AI into Fiverr’s Inbox raised a separate challenge: sellers were concerned about trusting AI with buyer communication and losing control of the conversation.

The opportunity was to test whether better-written replies could improve buyer conversion, while giving sellers a safe, hands-on way to experience AI.

The SOLUTION

Suggested Replies appears directly in the message composer, where sellers are already writing to buyers. It generates one concise reply based on the conversation and relevant service details. Sellers can insert it into the composer, edit it, refresh it, or dismiss it.

The experience balances visibility with control. Suggestions can open automatically after a buyer’s message burst, but only up to three times in a conversation. After that, the feature remains available on demand. Sellers can also close the feature, and a first-time onboarding tooltip explains that suggestions are optional and editable.

To keep replies useful and grounded, the system considers relevant context such as the conversation history, gig details, custom offers, attachments, and whether a call has already been proposed or completed. It is designed to ask clarifying questions when information is missing and avoid promising anything outside the seller’s service scope.

Measuring success

The primary success measure was the difference in buyer conversion rate between test and control buyers. We also defined measures for suggestion views, use, edits, and repeat usage, alongside observational measures of PA activation and usage. Cancellation rates and negative feedback served as guardrails, helping identify cases where a suggestion might mislead buyers or create a poor experience.