KI-Workflows machen aus Wissen erledigte Arbeit (Full Transcript)

[00:00:02] Speaker 1: An answer isn't an outcome. If the AI tells you what the project plan is and you still have to write the recap and update the CRM and chase two people, then we've saved you the searching and left you the chores. That's really not good enough.

[00:00:18] Speaker 2: So third tier is action, turning retrieved content into completed work.

[00:00:22] Speaker 1: And the important design decision is that these are not an agent does it. A workflow is a graph of nodes and the nodes are different kinds deliberately. There's trigger nodes, there's deterministic nodes, we have AI reasoning nodes, and connected system action nodes. You mix them together on purpose. The parts that should be reliable are deterministic, things like routing, field mapping, and who gets notified. The parts that need judgment get a reasoning node. The whole thing is one model called and you can't debug it and you can't predict it.

[00:01:00] Speaker 2: Triggers come in two shapes, event triggers, a meeting ends and you can scope it down to a specific meetings rather than, you know, every meeting on the tenant, a scheduled and scheduled triggers on reoccurrence, which is a separate model. And for anything sensitive, you put a human review step in the graph before the write. That's not a limitation we're apologizing for. It's the correct design for an automated system that touches a system of record, which brings up the thing I said in the last segment. And I will say it again, because it's the one that people get wrong. The action tier authorizes separate from the retrieval tier. Connected actions execute with the permissions in their connectors, all scopes, finding something never implies the ability to change it. And every step is audible, which is what you should actually care about. If something automated touches your CRM, you need to know what it did and when at no granularity, not the AI updated it. Automation without an audit trail is just a rumor about your data. It reaches into the tools you already use.

[00:02:08] Speaker 1: So email, calendar, CRM, service cases, development tickets, and documentation. Plus all the AI productivity suite on our side. So we have slides, sheets, canvas, and paper. So finished work can mean a deck or a spreadsheet, it can be a doc, or it can be a CRM update or a message in a channel.

[00:02:31] Speaker 2: So Matt's going to build it and show you what he did.

[00:02:34] Speaker 3: So here we can see the workflows tab within the client. We see that we have a variety of templates provided by Zoom for simple things like meeting participation reports or pre-meeting reminders, or more complicated things used by IT or engineering teams such as ticket escalations or incident reviews that use a combination of deterministic flows with AI reasoning capabilities. So those deterministic flows give you the confidence that it's going to work the right way every time. And the AI reasoning capabilities give it the ability to do more than a traditional workflow would be able to by reading through information and making determinations and decisions from there, or generating something in a human readable format that would otherwise be technobabble. All of these were created by Zoom. There's quite a few in here. Many of them can be used out of the box. But you can also create your own workflows either for yourself or for your colleagues. So your organizational templates can be distributed to the entire organization or to subsets with specific needs. So say you need something for the IT department and something else, marketing department, you can distribute those to just those departments if you like, and be able to have members of those teams take advantage of them without having to build their own. Let's see what building looks like. So I'm going to go click on Create. And the first thing I see here is that I can create a workflow just by describing it. So I can type whatever I want in this box to describe a workflow that I would like AI to help me create. We give a few examples here for after meetings, other things that we can accomplish. We can also choose from one of the templates that we just saw. We can import a workflow from a file. They can also be shared directly without having to import or export. Or we can start with a node. So let's go there first. Here we can see that I have a variety of different types of nodes. But I always start with a trigger. This is what's going to kick off the workflow in the first place. I can kick off the workflow manually with a button, maybe a form where I fill in a couple of pieces of information needed for the workflow to take action. There are webhooks to do things programmatically or when a link is clicked. But in this case, I'm going to choose a meeting event. That's probably the most popular. So meeting events can be of a variety of types. These can be things that happen when a meeting ends, can be happening when the meeting starts, or when certain specific things happen. So a summary has been created, or if someone joins late or early, or if someone joins and we want a specific action to happen with them. So let's say I want this one to happen when someone joins my meeting. And in this case, it's set for all my meetings. But I may want specific meetings only or ones that meet criteria. But I'll leave it with all meetings at this point. The next thing we do after the trigger is we choose some things in order to take action. So I can choose from deterministic nodes like Zoom chat. So here I can choose to send a direct message. Maybe I want to send a direct message to myself after when this is created. So I'll put my name in here, grab myself from the list, and I can either use markdown or plain text and just say something like, someone has joined water meeting. So if I was distracted doing something else because I joined early, I'll get this direct message letting me know without them having to DM me and tell me that I'm ignoring them. We can also use other tools such as AI reasoning nodes. And AI reasoning nodes are able to do more complicated things. I highly recommend taking advantage of the conversational AI to generate these. It can create instructions, choose which tools to use, which can include your own MCP servers or tools that we have available through our connectors. All of these can be put together to do simple things or extremely complicated ones that involve skills, third-party tools, or advanced searches. Let's look at a couple that I've already built. So here I've got a simple one for a text follow-up. So this is after a meeting event occurs, a summary has been created. I'm able to use the AI reasoning node and this send SMS message in order to send the summary to my phone. This is great when I'm on the road and I'm going to have stuff at my fingertips for follow-up, but I don't necessarily want to jump into the client to have to find that information. We can see another one that I made that takes advantage of third-party tools is able to work on a schedule. So this one, rather than being based on a meeting event, happens every day at nine in the morning. It can reach us into a third-party tool to find things that are blocked. This query language was built for me by AI. And then a reasoning node is able to turn the results of this into natural language and send me a chat message so that I get a quick summary every day without having to look at all the items that are blocked and addressed to me. All these together can take some of the tedious parts of your day and automate them.