Voiceflow

The order returns process for retailers is manual and inefficient. We built an AI agent to automate this customer support workflow for New Balance. The chat experience goes beyond simple L1 tickets and can complete actions for users: 👟 When you’re looking to get a refund, you can scan a QR code, look up by your order number, or show all your recent orders. 👟 If you give a return reason in your chat messages, the agent can grab that context and input it automatically in your return form 🤯 👟 If the refund reason is determined to be a manufacturing default, the agent will generate a gift card via the Shopify API Human agents spend their time on more complex tickets, while CSAT and resolution rate jump. This type of AI automation can make level 2 customer support inquiries easier — even enjoyable — experiences for your customers. Check out Peter’s flyby demo to see it in action.

Here's how to pick your first AI agent use case ⬇️ When planning AI projects, we often see teams assess priorities based on 5 categories: ➕ Highest ROI ➕ Familiarity ➕ Technical simplicity ➕ Lowest risk ➕ Organizational simplicity Where you start with AI will depend on how your company handles these priorities. Should you build something that will yield the highest ROI but will open the team to potential risks? Or should you play it safe, building an internal automation that can be iterated upon slowly? Balancing these priorities can be challenging and will differ by organization and opportunity. Reflect on the makeup of your team, the technical capabilities available to you, the industry you work in, and the goals of the company. Once you have one AI product launched and out in the world, the path to scaling and adopting more complex use cases becomes easier. Here's an example from a team working in Voiceflow: ✅ Trilogy had an AI-first ethos as a company, making them a good candidate to start with a high-ROI use case. They started with an ambitious program to automate L1 customer support for 90 individual products. ✅ The team started by developing Atlas Core, a foundational agent integrated with Trilogy’s help center, a knowledge base, and core set of user support flows and functions. ✅ They then used Atlas Core to build the other 80% of standalone projects. This helps them scale AI agents quickly and add features to specific products while maintaining a complex core that they’re continually improving. If you need help getting started, here’s a simple flowchart to see where you might land for your first project. Then, take a look at the suggested scaling pathways to see where your team could end up after a few successful launches.

🛠️ Introducing: Secrets Manager You can now securely store and manage sensitive information within your AI agents. We know impactful agents are built on business-critical data via 3rd party integrations to resolve tickets, convert leads or drive orders. Secrets Manager makes it easy to keep these integration credentials secure with industry-standard encryption, and simpler to use by inserting secrets directly into Function and API steps. Some highlights: ⭐ When duplicating or exporting projects, secret values are excluded to maintain security across teams and projects ⭐ Utilizes AES-256 GCM encryption for top-notch security ⭐ Specify different secret values for Development and Production environments See it working in action below, and try it out for yourself (available on all plans).

🛠️ Video Recap: Everything we released in September Our releases and updates from September simplify the building process, improve response quality, and increase LLM efficiency — making it easier than ever to create amazing customer experiences. Dive into our recap video to learn more about the updates: 🗣️ Multimodal Projects (Voice + Chat) - build and test both voice or chat agents from one unified project type with Michael 📈 Better Token Efficiency - token multipliers for all models have been cut in half, making it less expensive to build complex prompts with Daniel ⚡ Updated Condition Step - visually configure enhanced logic and add custom Javascript expressions with Evgeny 🔎 Knowledge Base Search Step - build RAG pipelines for improved AI Agent response quality with Amanda You’ll also get a sneak preview from Tyler of LLM Entity Extraction, designed to bring more reliability and accuracy to user interactions by introducing rules, exit scenarios, model customization, and automatic reprompts. Happy building! 🚢 #aiagents #aiautomation #buildingwithai #generativeai #conversationdesign

A basic AI agent may be able to answer FAQs or direct customers to additional resources. But for more complex inquiries? They’ll frequently transfer to a human. And at high volume ticket times, those transfers can be overwhelming — even requiring hiring more staff. Complex agents connected to key data sources and tools, designed to take action and not just answer questions — that’s where support organizations start to see real scale with AI. This Thursday, our AI Blueprint series continues with a special guest: Parkfield Commerce, experts in customer experience who have built agents that reduce tickets by 60% and more, while using extensions to recommend products and increase basket size. We’ll cover: ➡ Prioritizing AI agent development for order support and triage ➡ Handling high support volumes with better profitability ➡ Reducing customer friction by running automated returns through Loop and discount codes via Klaviyo ➡ Real world examples from e-commerce retailers using AI agents to drive sales and resolve support tickets

Today's under-appreciated use cases for LLMs? Brainstorming. For our in-house conversation design expert Peter Isaacs, he says brainstorming with an LLM can "turbocharge the creative process" -- from designing out workflows, to writing prompts, to troubleshooting why something might not be working. In this episode of Pete's Prompts, Pete shows how he uses Anthropic's Claude Sonnet 3.5 as a co-pilot, helping him build a global fallback framework. Watch Pete use Claude to: 🔧 Build a mermaid diagram for guardrails 🔧 Write a prompt, referencing the diagram 🔧 Run the agent itself inside Voiceflow #generativeai #conversationdesign #agentdesign #aiagents

We're so excited to show you what we have up our sleeves for this month's #MakingBots! 👟 ⚡ Here's a preview of one of the chat extensions Nicolas Arcay Bermejo built for yesterday's Developer Lab, which Peter Isaacs has incorporated into the build for New Balance. How beautiful is this returns flow? 😮 Talk about a smooth customer experience! See it live: voiceflow .com/events

Can your out of the box chat widget pull in custom Webflow components? 🤔 Custom interfaces are changing the way customers engage with AI experiences. Originally focused on Webflow development, Zac Santer and his team have been using Voiceflow for the past year and a half to bring the power of AI automation to their web clients. This is a great example of thinking outside the [chat] box — how can AI interact with other apps through custom interfaces and provide more value? As Zac put it, “[web visitors] can get what they need inside that chat experience, and be directed to areas on the site if need be”, rather than navigating aimlessly around the website. We love to see how creative our builders can be in developing complex AI customer experiences that deliver value for customers and businesses. Check it out! ⬇️

🛠️ Get better AI responses with the Knowledge Base Search Step You can now build RAG pipelines for improved AI Agent response quality - with far less hallucinations ⚡ The KB Search Step simplifies querying data from your Knowledge Base, giving teams more control over how they fetch data chunks in their conversational workflows. This step will make answers from your Knowledge Base 10x more accurate, without the added effort of manual API configurations. With the KB Search Step, you can: 🔎 Query your Knowledge Base without the need for manual API configurations via the KB Query API. 🔎 Control how many data chunks are returned and set similarity thresholds to draw the most relevant information. 🔎 Automatically convert and save data into variables for use in future steps in the conversation. Using the KB Search Step in Voiceflow creates answers that are highly accurate, very flexible, and in tune with what the user is doing. Along with this new step, we’re releasing a template that will be automatically applied in any new chat projects. Check out Daniel’s video below for a walkthrough of the details, then click over to creator[dot]voiceflow[dot]com to try it out for yourself.

For the retail world, the last weekend in November can be a make-or-break sales window. The four-plus days of doorbusting deals and record-breaking revenues are known as Black Friday Cyber Monday, or BFCM in the industry. For some retailers, BFCM sales can amount to over 50% of their annual sales. In 2023, e-commerce stores built on Shopify alone generated $9.3 billion in this period — up 24% from the previous year. Online retailers can be inundated by order volumes and customer support requests at this time, with many increasing their customer support teams to handle the influx — which can reduce profitability. If you’re looking for alternative tactics to drive better conversion rates, repeat buyers, and better profitability, you won’t want to miss the next AI Blueprint event. On October 3rd, we’ll be joined by Shopify Plus partner Parkfield Commerce to share best practices for preparing for BFCM traffic — and getting a little help from AI automation. We’ll cover: 🛍 Prioritizing AI agent development for order support and triage during BFCM 🛍 Handling high support volumes without harming profitability 🛍 Reducing customer friction by running automated returns through Loop and discount codes via Klaviyo 🛍 Real world examples from e-commerce retailers using AI agents to drive sales and resolve support tickets Register at voiceflow .com/events, or follow the link in the comments.