Gepetto Skills - AI
Semi-automated customer service using generative AI, with human review and data governance.
ENTERPRISE
Cora Bank
YEAR
2023
PAPER
Product Designer
PRODUCT
Gepetto Skills
.png)
OVERVIEW
With the advancement of generative AI, a clear business need arose: to reduce operating costs and gain efficiency through semi-automation of processes and activities in customer service. It was in this context that the Gepetto Skills app was born.
The operation had a high volume of email requests and relied on manual responses, which resulted in time consumption, scalability difficulties, and higher operational costs. The opportunity lay in using AI to generate responses based on customer content and account data, while maintaining human review to ensure quality.
THEN...
We needed to use AI to generate real value for the operation, creating a workflow where internal teams could:
Generate quick and consistent responses to requests received via email.
Test and evolve skills (prompts) safely and with traceability.
To measure performance (accuracy and time) and continuously improve, with human review (human-in-the-loop).
INVESTIGATION
STAGE 1
Discovery (quick and practical)
Benchmarking: how the market uses AI and what trends were already being applied.
Desk research: prompt engineering courses + study of concepts such as Human-in-the-loop and data structures.
Brainstorming with multiple areas (Design, Dev, Cyber, TechOps, CX, Data, etc.) to map real opportunities.
STAGE 2
Validation and prioritization
In-depth interviews with 5 people, testing solution hypotheses and everyday needs.
Impact vs. effort with multidisciplinary teams to choose what would generate the fastest return with technical feasibility.
TRADE-OFFS
Before designing the interface and workflows, it was necessary to train the teams in the use of a recent technology; therefore, an internal hackathon (Corathon) was created to accelerate knowledge and alignment.
The solution was structured with human review (semi-automation), prioritizing safety and response quality before any ambition for full automation.
SOLUTION
Build an internal application that generates automatic responses based on customer text and account data, for the CX analyst to review and send, significantly reducing response time.
Prioritization strategy:
Map the main pain points by tags/service topics.
Identify the requests that were occupying the most space in the email inbox.
Distribute tags to managers to create prompts ("skills") and track performance.
IMPACT - 3 MONTHS
100% data governance, since we now use an internal platform to generate GPT Chat responses, without directly depending on the official OpenAI website (data protection);
92% of all operational demands are semi-automated;
Cost reduction in BPO contracts;
Increased operational efficiency, allowing agents to dedicate themselves to tasks beyond responding to tickets, and thus work on their development through our learning paths.
"Gepetto transformed emails into ready-made responses in seconds, with human review, traceability, and scalability."
LEARNINGS
In AI applied to customer service, it's not enough to simply "have a model responding." What generates value is an operational workflow: useful responses with human review, accuracy/time metrics, and traceability for safe evolution.
The most efficient shortcut was to start with operations (customer service/sales) and then validate with clients. Early alignment with data/development avoids promising what cannot be measured/demonstrated within the timeframe.
The most efficient approach was to structure the foundation (tags, assignees, metrics, and testing playground) before attempting to refine the experience or "automate everything."
Aligning early with CX, data, and engineering prevented a "pretty" solution without sustainability: without defined accuracy, without time metrics, and without usage tracking, there is no consistent evolution.