Pathfinder AI
Project Overview
Agent Resolution Console, the internal case management product, was built to be an AI driven support experience. However, the AI model used was underperforming and lost the trust of its users. As a result, the CX org wrote their own AI agent to help resolve customer cases that showed results. Our goal was to replace Agent Assist with this new agent, then build a user experience around it that could help it scale for the product and used by 100% of agents.
Client
Coinbase
Industry
AI & Customer Experience
The business wanted to replace CX’s broken case management AI with a new but rigid model to improve overall CSAT & restore agent trust. Our goal was to integrate it in the product and launch to 100% of users
The original model, Agent Assist, wasn't producing results and lost user trust. It was too noisy, lacked context, and gave inaccurate responses resulting in the CX org build their own model Pathfinder. It resulted in better CSAT with a small cohort, and the business wanted to scale it to 100% of agents. However, it was rigid and required necessary manual effort; it worked but didn't meet baseline AI expectations, risking specialist trust if shipped unchanged.
CX needed confident resolutions and control of AI’s output. Instead of lifting & shifting the new model, we designed the experience to naturally fit into CX troubleshooting workflows while restoring the trust lost from the old model
Stakeholder interviews & shadow sessions helped us define how CX needed AI for their jobs, and I designed the Northstar to show how the foundational model could be used more intentionally to meet their needs. It also served as the core mechanism for aligning stakeholders on where we were heading, why this was the right way, and how we get there. Everyone knew what they wanted to do, but there was no shared vision for how to get there until the Northstar.
At the start of a case, Pathfinder AI provides specialists with context and an easy path to diagnose the problem, saving them time & effort while rebuilding their trust with confident AI responses
Case Context gives an overview of the customer’s issue & profile removing the need for manual searching or context inputs
AI investigates to diagnose the issue and find the best solution. AI does the heavy lifting while CX uses their judgement to determine the best course of action
If a resolution or step is not correct, Specialists can provide feedback and use alternative options. This trains the model to improve accuracy and build toward the long term goal of automating less complex resolution tasks
Specialists got alternative solutions if the first one was incorrect. This prevented dead end scenarios while teaching the model what solutions were right for a given situation
Agents can provide feedback on individual steps as needed, helping the model learn what steps are correct to improve accuracy
We rebuilt the agent’s prompts and MCP connections to support searching for account related data in the product. This reduced context switching for Specialists while making contextual information easier to find
I designed an orchestration experience to guide the foundations for how we scale the product to support multi-agent workflows and enable the business to add new AI capabilities
With teams building specialized AI capabilities to use in the product, we had to figure out how to enable specialists to use them in a seamless way. I designed an AI orchestration experience to define how multiple agents would automatically run based on the case context. This helped engineering leaders define the architectural foundation used to scale our product from a single AI agent to a system of interconnected agents anyone in the business could contribute to.
We worked backward from the Northstar vision to iteratively ship features, using each release to guide the product toward that vision while also refining the vision based on new learnings
To move us toward the Northstar vision, we incrementally shipped features, watched them in action, then made adjustments, with success being measured by CSAT improvements & CX adoption. The process required us to be ship fast and be flexible to meet the immediate user and business needs, while also requiring me to hold the team accountable for ensuring each feature was aligned with our end goal.
