Why I built Deflekt.ai

| Published

Today I launched the public beta version of Deflekt.ai. Deflekt.ai is a customer support ticket deflection tool. It basically tries to reply the customer support emails, before an actual support ticket in created in your shared inbox app like Zendesk, Hubspot or Crisp.

Background

I’ve been working in eCommerce and SaaS for a decade and despite all these fancy tools I’ve got to experience I thin customer support is one of the toughest things to scale. Twice the amount of orders? Twice the amount of email tickets.

You can add different support channels (chat, WhatsApp, etc.) and in these channels it’s already a lot easier to deflect questions. You can suggest knowledge base articles, create some simple bots and even fetch some data from your ERP or CRM to answer questions about a shipment status.

However, some people just like the asynchronous aspect of email. Type, hit send and wait for a reply because you are busy right now. Triaging customers or users through a knowledge base, through a chat and only then show an email address or phone number has also become a standard, but personally I’m not a big fan.

AI Chatbots and LLMs

Lately tens or maybe a hundred AI chatbot tools popped up. I think I tried at least 30. In theory they are all the same (although some are much better then others). You ingest some documents and/or webpages, the tool creates a vector embeddings database and when a question comes in, relevant chunks of text are inserted into a prompt for an LLM to generate an answer.

With the right content this works extremely well. Especially for those 20-50% of questions to which the answer is clearly answered in a knowledge base or set of FAQs.

AI for email

That made me think. Why not do this for email (or why doesn’t anyone do this with email)? I started prototyping this and realized it needed a ton of backup mechanism. What if the AI service is down? What if the reply doesn’t answer the user’s question? One by one I tried to solve this issues and by the end I was convinced I could morph this into a useful tool for others.

Deflekt.ai does exactly this. You create an inbox (alias) and set up a forwarder for an email address in your company. I’m a big fan of doing this in a transparent way, meaning you should tell people that when they email you they will get an instant reply by your trained AI bot. Your ‘human’ support inbox can be hidden for now.

For each inbox you create you configure an existing AI chatbot (Mendable, Chatbase and AskAI are the three services that are supported right now, let me know when you come across a decent service with a documented API). If you’re not using an AI chatbot you can also use our built-in bot feature (it’s still in alpha but works fairly well).

The email pipeline will take care of all incoming mails. You’ll be able to track the status for each email. If a reply is sent, you’ll be able to read it so you can monitor quantity and write new knowledge base articles.

When a user reply to a deflection mail, Deflekt.ai routes that mail through the Deflekt.ai server. The message in Deflekt.ai is then marked as undeflected. This way you’ll be able to monitor the deflection rate of your inbox. I’ll add more advanced stats soon.

If for whatever reason a reply can’t be generated, we mark the email as skipped. Deflekt.ai will also skip the reply if you hit your plan’s limit. The original message is relayed to your human inbox - that way you never miss a message from your users.

Would love to hear from you

I’m extremely curious to hear what you think about the product, so go ahead and give it a try. The Free plan lets you send 50 deflections per month, all messages after that will end up in your ticketing system. If you like it you can upgrade to a paid plan. Plans don’t have any feature caps, just the number of deflections and inboxes.