Fellow Projects
SurfOps: Building AI Agents for Customer Success
An Overclock fellow project by Angel Ancheta, Roberto Lagos, and AK ALMoumen.
A customer success manager opens Salesforce before a call. Somewhere in the account history is the context they need: a usage decline, an unanswered email, a missing contact, a task someone meant to follow up on. Finding it is work. Deciding what to do with it is another job altogether.
That was the problem behind SurfOps, a customer success workspace built by three Overclock fellows. In their capstone presentation, Angel Ancheta, Roberto Lagos, and AK ALMoumen walked through a system they called Operator Surface: specialized AI agents gather account context and propose actions, while a CSM reviews the work and approves what happens next.
Their starting point was the morning workload. Which accounts need attention? What changed? What needs a decision?
The workspace brings those questions into a daily brief, an approval queue, and a record of agent decisions. In the demo, the brief surfaces 14 accounts. One has five proposed actions waiting for review, including an outbound check-in and a task to create. The CSM can move from the account summary into the approval queue without reconstructing the situation across several screens. See the walkthrough at 2:16.
The agents also have access to operating procedures. The team describes giving Galileo, its account-facing assistant, context from roughly 23 SOPs, including documents and presentations informed by customer success work. Those materials give the system a basis for recommending how to handle an account.
Angel demonstrates another use: preparing a junior CSM for a difficult customer conversation. Galileo works with specialist agents to examine the account data, check its quality, and score it against a rubric. The resulting brief includes context and a suggested action plan. A CSM can ask follow-up questions through the chat. See the preparation example at 6:21.
The permission boundaries are a particularly useful part of the design. Some agents can read Salesforce data. A smaller set can carry out changes or send communications. Angel explains that these restrictions are enforced through Salesforce permissions, so a read-only agent cannot acquire write access simply because someone asks it to make an update.
Human approval connects the two sides. As the team describes it, agents put proposed work into Supabase, the database behind the application. The web app displays it for review. When a person approves an action, the application records that decision and triggers the agent responsible for executing it in Salesforce or Gmail. The team explains the flow at 14:16.
That makes the approval queue a working part of the system. A recommendation has somewhere to go, a person responsible for reviewing it, and a defined path to execution.
The presentation also shows how much work sits behind that experience. Angel describes spending hours wiring the agents into Salesforce. Email actions were taking longer than expected during the demo; she reported a successful send before class but did not demonstrate the full send completing live. Other ideas, including broader CRM support and open-source distribution, were plans for the project’s future.
AK approached the interface with continued changes in mind. He describes building a shared design system and reusable components so his teammates could adapt the workspace as their ideas developed. The team reports assembling the project in roughly ten days, with each contributor working in a separate development branch.
SurfOps is a capstone build, with no measured retention gains or team-wide productivity results presented in the demo. What it does offer is a concrete example of the decisions involved in applying AI to customer success: which information an agent needs, how it should assess that information, which actions it may take, and where a person enters the process.
For an operator considering a similar project, one account workflow is a useful place to begin. Map the information you normally gather, the procedure you follow, and the decision you need to make. Then define what the system should prepare for you and what requires your approval. That gives you something specific to build and test.
In the AI Ops Accelerator, fellows work toward a capstone grounded in a problem they want to solve. SurfOps shows how much there is to learn by taking one familiar workflow all the way from account data to an approved action.