Agent Copilot

My Role:

I own the design of Sprinklr's Agent Copilot end-to-end. This is an ongoing project that I will keep updated here as more capabilities are released. In my ongoing work on this, I consistently collaborate with designers on other Copilots within Sprinklr to maintain platform-wide consistency.

Customer service agents are overwhelmed by scattered information

Customer service agents juggle too many things at once: Knowledge Base articles, past cases, customer history, all scattered across their dashboard and external links. This slows them down and makes it harder to help customers effectively.

Copilot brings it all into one place

Instead of switching between tabs and tools, agents ask copilot and get answers in one chat interface. Interact with the prototype below to see how it works.

Foundational patterns and interactions

Input via text or out-of-box prompts

Performance of customer service agents includes a metric called Average Handling Time (AHT). So it is important for them to reduce time spent on each case. For that reason, we decided to give them repeatedly used prompts upfront in the prompts carousel.

I think a nice addition to this would be giving them a toggle to show or hide the prompts so the users can decide for themselves if the screen space taken by the prompts is of value to them or not.

Streaming of copilot responses

Streaming shows the response as it's being generated, so agents aren't staring at a blank screen. It reduces perceived wait time and lets them start reading sooner.

Sources and citation

Agents need to trust what copilot tells them before passing it to customers. Showing sources lets them verify information quickly and builds confidence in the tool.

Clearly showing response types

Even though brands were asking for AI-powered features to make customer service more efficient, we learned that they had low trust in Ai generated responses. They did not want their agents to use AI-generated responses.

This insight shaped a core design decision: clearly distinguishing between canned responses and AI-generated responses. Canned responses are pre-defined standard responses by brands. Copilot would never generate a canned response, it would just surface them based on context.

Context addition for effective search

Large enterprises have huge databases of brands, product lines, and product names. To provide the copilot with relevant information, agents can add context tags that scope the search.

Customer cases usually involve a specific context that stays consistent throughout the conversation. We placed context tags in their own section above the input so agents don't have to re-add them for each message and are less likely to remove them by accident.

Brand specific information (ex: sub-brands, product lines)

These are detected automatically when a case is assigned. The AI identifies what brand or product line the customer is asking about and adds the relevant tags. Agents can override if needed.

Messages from customer conversation

For long cases or cases where the topic shifts, agents can manually add specific messages from the conversation as context. This was added based on customer requests for more granular control.

Prototype to experience agent copilot in context

I constantly explore new patterns for AI interfaces and document them in my AI patterns note. I would love it if you took some time to check it out and play with the prototypes there!