AI oversight and feedback loops 🔄 (with practical examples).
AI without human oversight is similar to fully autonomous cars without human operators; it may be possible, but you’ll be cautious when hopping in. Add a human for oversight; as the technology improves, the human does less, and less... and less.
Here’s how to do it-
👁️🗨️ Confidence Scoring.
If the AI's confident, it automates.
If it’s not, it loops in a human.
From this feedback we can train the model to get sharper with every cycle.
✅ Faster workflows
✅ Human-aligned decisions
✅ An optimising system that gets better over time
Attached is an example of how we do this in practise using Make for Social Media Comment Management: 1. A comment is left on the social media channel (in this example we use Instagram)
2. Gather information from the website or knowledge base to assist the replying Agent
3. Analyse the sentiment of the comment
4. Determine the 'type' of comment- question or response
5. Route accordingly based on the type of response.
6. If it’s FAQ, check whether the answer already exists in the vector store. If it does, respond with that response..
7. If the answer doesn’t exist, loop in a human to provide the answer via the dashboard or email.
A human manager can view all comments and responses, and provide information to the agent via a dashboard built on Airtable. Once the response is added, the AI will add the answer to its knowledge base, so that in the future no human will be needed in the loop.
This Retrieval-Augmented Generation (RAG) approach pulls data as context for the large language models (LLMs) to improve relevancy.
With this approach, we’re able to reply to customers in 30 seconds 💨 with responses that are both accurate and contextually aligned. We manage 1000s of comments a day, in different languages, across multiple social media channels 🥳.
Feel free to ask me questions about this below 👇