Charles AI Squad
Role: Senior Product Designer
Duration: Jan 2025 - present
This case study highlights my work at Charles in bringing AI-driven automation to our WhatsApp platform. I focus on reducing manual support workflows and designing intelligent, scalable agent experiences.
Background
At Charles, we help our clients manage their WhatsApp communication by offering them tools to build automation flows, run marketing campaigns, and respond to end-users queries directly through our platform.
To support a seamless eCommerce experience, we also offer integrations with a wide range of tools including CDPs, CRMs, shop systems, service desks, and analytics platforms which allow our clients to connect their existing tech stack and fully control their customer journey on WhatsApp.
For more information visit Charles
Why We Introduced AI Agents
Manual customer support is costly and inefficient.
Businesses usually spend up to $1–5 per support interaction, and that number scales quickly for growing eCommerce brands using platforms like WhatsApp to engage with thousands of customers daily.
Instant 24/7 Responses
Manually answering repetitive queries often leads to delays, especially outside business hours. AI agents can respond instantly 24/7, improving customer satisfaction and reducing drop-off.
Helping Users Find the Right Product
When customers can't easily find the right product or get their questions answered quickly, they abandon the conversation.
Inconsistent Information
When multiple support agents reply manually, there's a risk of giving outdated or inconsistent answers. The FAQ Agent ensures consistent, verified responses every time.
Solution
As we began exploring AI automation, Charles acquired Spectrm.io, an existing solution with pre-built AI agents.
- As a product designer, my challenge was to seamlessly integrate Spectrm’s AI capabilities into our platform while preserving the clarity and consistency of our existing information architecture.
- This meant finding the right place for AI Agents within the interface without disrupting user workflows or adding unnecessary complexity.
Business Impact (Beta Phase)
We launched AI Agents as a beta to a selected group of 5 customers. Early results are promising:
- 100% of interviewed customers expressed interest in purchasing AI Agents.
- All 5 beta customers are actively using the feature.
- We met our T1 OKRs by integrating 2 AI Agents (FAQ Automation Agent & Product Finder) and enabling usage across 5 client accounts.
- Early results show that AI Agents helped reduce onboarding time by approximately 23% for beta users.
What were the biggest challenges?
Rethinking Research in B2B 🧠
Conducting user research in a B2B SaaS setup is never easy, especially when the decision makers aren’t the daily users. In our case, the core users were CRM managers who juggled multiple platforms (CDPs, CRMs, etc.) and were often hard to reach or too busy to participate in research.
To navigate this, we combined two valuable sources of insight:
- Our Customer Success Managers, who work closely with clients and deeply understand their needs.
- Spectrm's prior experience and user feedback, which offered an additional layer of depth from a product already focused on AI agents.
We took a lean MVP approach: Building a simple, functional version, testing it with our CSMs, gathering feedback, and improving quickly. Just before the OMR , we deployed the MVP for the event, and tested it with real users, collected direct input, and used all insights to refine the experience.
Ensuring Clean and Compatible FAQ Uploads 🗃️
To train the FAQ Agent, users are required to upload a CSV file containing questions and answers. This file is converted to a structured JSON format, which Magpie(our underlying AI engine) uses to process and retrieve relevant answers.
However, this upload process came with critical challenges:
- Parsing errors like incorrect delimiter detection, could lead to a malformed dataset
- Special characters and line breaks within cells that could corrupt the conversion.
- CSV files with extra columns, where only two were relevant for Q&A mapping.
Our 3-Step Validation Flow
Step 1: File Cleanup
After uploading, we immediately scan for special characters or formatting issues. If found, users are prompted to fix and re-upload before moving forward.
step 2: Delimiter Detection
We use PapaParse to auto-detect the file's delimiter and show a preview. If fewer than two columns are detected, an error prompts the user to manually select the correct delimiter.
Step 3: Column Mapping
For files with more than two columns, users map which ones correspond to the question and answer fields.
Safeguarding Brand Voice in an AI-Powered World 🛡️
Ensuring brand safety was a key challenge. Our potential clients, along with Spectrm's existing clients, were concerned about AI agents engaging with sensitive topics like scandals, politics, or competitors, where even a single bad response could damage trust and credibility.
To solve this, we introduced SafeMatch, a feature that lets businesses flag sensitive keywords. When matched, the AI triggers a fallback flow (using our already existing flowbuilder) instead of replying freely, helping brands stay in control of sensitive conversations.
