From Chatting with AI to Changing Work: Finding Practical AI and Agent Opportunities

Hi, everyone. My name is Eric Mistry. I work for Zapier. For those of you who knew me and other roles, yes, I am the kid who got into the candy store.

Before we begin, just because I know AI is a hot topic, and I got some takes on it. So I think AI is both overhyped and underhyped at the same time. Both can be true. The other thing that's very important, I've seen both at Zapier, but also for all of us, is that AI fluency is becoming really important.

It doesn't really matter what your field is. You need to have AI fluency to protect your role. And customer success, customer education, the fields we work in, these are all very ripe for disruption, and having the skill set, having the tools, having the vocabulary is going to help you talk to leadership about how to work with these tools and not be replaced by them. So let's begin with an activity that's been kind of fun.

We've done this internally, and I'm very excited to see what you all come up with. So take about a minute and write down what is impossible today. What can AI not do? And because this is a fancy conference, there's a timer right in front of me, so I don't even have to set one.

And also, I should know, but it looks like all you figured out, there are worksheets in front of you. We went a little analog for this AI session. And there's also slide links and everything with QR codes. None of them for this session are rickrolls, so you can trust all those links.

Great, so that's about a minute. At your tables, talk to someone next to you, some of you are in tables of three. Just share with people near you what did you write down and why. All right, so...

You all might be asking, "Why are we talking about impossibilities? "Why are we talking about things that are impossible when this is AI?" This is like what you're supposed to be doing with these new tools. There's something called the Bannister Effect, and it's about running-- it's an old, old, old story. 1954.

Before May 9th, 1954, a four-minute mile was impossible. It could not be done. Medical science, you know, doctors, all the experts. It cannot be done.

It is physically impossible. Roger Bannister broke that record. Okay, that happens. People break records.

What happened 46 days later-- 46 days later, it got broken again. 16 people that year broke that record that was impossible. Same thing is happening with AI right now. So things that were impossible.

Once we see it done once, suddenly the impossible becomes possible again. So we've all got these lists. We're gonna come back to your items at the end of the session. So look at what you wrote and come back to it, even if it's not possible today, come back to it in 18 months and say, "Oh, turns out that was possible." So another little bit of interactivity, because this is a workshop.

Just by show of hands or shout-out the answers, question one, what percentage of the world-- so 8.3 billion people on Earth-- what percentage is actively using AI in any way in their lives? 5%. 10%. 5%.

Okay, I'm hearing 5% during 10%. You are a very informed crowd relatively. Question number two, what percentage of the world is paying for access to AI? Oh, you're very good.

Someone stole my notes. So because you are the quickest crowd I've got, this chart really puts it a lot in perspective. We're hearing about AI constantly. Like, my LinkedIn feed is either, "Please buy my Claude skill," or, "Here's a new AI product." This is the real part of the world.

I'm in Duluth, Minnesota. For those of you who don't know, when I say, "Oh, I work with AI," people are like, "I've heard of that." And I said, "Like, Claude," and they're like, "Who's Claude?" So most of where I live is either in those green dots of people who have used AI, or a lot of them are just in that gray. They've heard of it. They've never used it.

Those little lines up at the top, which I can barely see, and I'm very close to the screen, those are the people paying for AI using things like codecs, using things like cursor, using things like Claude code, is a very, very, very small percentage. So even though it seems like it's everywhere in our lives, it's really just a small fraction. So everyone is using AI is actually mostly a vibe. Like, just like vibe coding, it's a vibe.

It is growing. I looked at this same graph three or four months ago. It's about doubled since then. So it is growing, but the number of people meaningfully using AI is still minuscule.

And what that means for the rest of us is-- this is a good day to start. I see a question in the front. [indistinct talking] Yeah, I think-- Totally, yeah, good question from the front. So he's like, "If you're on AI, you're really interacting with AI models." It's a good point.

Methodology on this is, are you interacting with a chat GPT, a Gemini, or a Claude, like the free plan? But really good question and point there. But with this, if you're just getting started, like sometimes people show up to this and it's like, "Oh, I feel really behind. I'm in these tools every single day.

I feel behind." It's kind of the no good parent worries-- no bad parent worries about being a bad parent. If you're worrying about being behind with AI, that's probably a good thing. Like, that means you're thinking about the right way. Scrambled eggs.

Why is Eric talking about scrambled eggs on an AI session? This is the number one mistake I see with people getting started with AI, automation, all these things. They try to make a seven-layer soufflé, the fanciest, fanciest dishes. They've never made scrambled eggs.

So if you're learning to cook-- I started with scrambled eggs. My dad taught me when I was five or six. I broke a lot of eggs. I ate a lot of eggshells, but I learned to cook.

I did not start with that seven-layer soufflé. It is fine to learn by making scrambled eggs. And today, if you've not done anything before, we're going to help you find your scrambled eggs. So this is the problem I see.

I go into businesses and companies basically every single day. This is the question that people struggle with the most. It's not, "Oh, which model should we use?" or "We've got this complex thing we want to do." It's, "What should we automate or augment first with AI?" The three places these fall are individual productivity. So, "Hey, I'm going to manage my email to my calendar," those kind of things, teamwork flows.

"Hey, we have to pass information between members of our individual team," maybe the five or six people you work most closely with. The last one, this is the more complicated one, that's more the soufflé land, is cross-functional life cycles. "Hey, how do we hand off from sales to customer success?" These three areas are where people should be building, but they really struggle to choose. So before we get into what we're going to be building and how we find those things, there's three shifts I've noticed over the past year.

And I've been using AI chat GPT basically since GPT-3, a little bit of GPT-2. It's like, "I've been doing it a while, but a lot has shifted in the last year." We're going to go over each of these individually, but first, context beats structure every single time with AI. Two, we're moving to skills, not tasks. And three, and most important, the craft of what we do has moved upstream.

So, what does context beats structure actually mean? About a year ago, I gave a session on, "Hey, here's the best possible way to use AI." And it's like, "Hey, you're going to have this role. This is your task. This is very specific by the book.

You must follow this structure." My prompting now, I literally just yap at my computer for five minutes. It is stream of consciousness. It is not clear, but it is context-loaded, very similar to if you're just going to write a note of what to do. You're not going to write every detail, but if you're telling someone to walk through a process, you're going to give a lot more context.

So, hack number one, if you have a voice mode on your computer-- I use WhisperFlow because it just makes it really easy, but every one of these models has the little microphone icon. Just click it and yap, and give as much context as possible. The new models will sort through. It'll do all that fancy prompting we had to do just natively.

So, yap intelligently. Don't just go on tangents too far, but talking to your models is going to help a lot. Second thing, we're thinking a lot more in skills, not necessarily tasks. So, the old mental model is, "Hey, my job in customer education operations.

I do the data analysis. That is my role." That's a one-time kind of thing. The new mental model is actually, "I write the skill that lets the data analysis happen. I'm not the one doing the analysis.

AI is doing the analysis. I'm that human that does the results from it. So, I am building skills. I am not performing tasks." Third, and I think the most important, is that the craft has moved upstream.

And this is both a practical piece, but also very philosophical. I'm a maker. That is what I like to do. I do woodworking.

I do leatherworking. I do 3D printing. I like to make things. My work is the same way.

I like to build the workflows. I like to build slide decks. All of that is pretty much removed. 60, 70% of my work now is designing prompts, designing systems.

I am now just a manager of a bunch of AI agents versus the person who actually makes the thing. And that means you have to have a different set of skills. It means you have a different mindset. I have to approve the work that my AI team does, even though I am not the one making it.

And I really like, I heard someone earlier say that Albert Einstein, like, "Hey, you've got a bunch of time to think about a problem. Spend most of it thinking about the problem." I like to steal from Abraham Lincoln, who's like, "I got six hours to cut down the tree. I spent four sharpening the axe." Sharpening the axe is what you need to do now. The more you set up your tools to do successful things for you, the better your work is going to be.

And how do we know what to actually do? We've got to make the process visible. This is something that I've noticed again and again and again, and I am equally guilty of this. I am not like, "Oh, Eric does this really well." No, I'm terrible at this.

You cannot automate or bring AI into a process that you can't fully articulate. So we're going to work on that today. The four things we really need to capture whenever we're looking at any sort of process. What are the steps?

So step by step, what actually happens? Not the dream path, what actually happens. Step two, what decisions happen? So where does a human make a decision?

Where could an AI make a decision? Third, inputs and outputs. This is very much, again, from the automation side of the brain, but equally important when you're doing AI. So what's going to go into the system?

And then what do you need to come out and in what kind of format? You have to have these answers before you start building. And then finally, handoffs. So any time work moves from one person to another person, that is the most dangerous time for that work to get done.

Anyone who works in customer success and has sales to success handoff is called dangerous. So with that, we're going to move into activity two. Got about three minutes, and there's some parts in your paper there. Feel free to go off script if you want, but just map a single workflow that you own from end to end.

It does not have to be anything fancy, does not have to be game-changing. Like, just something you do daily would be great. So in order, list the steps exactly as they happen. Then I'm going to have you mark every decision point, just mark it with a D.

And then any time you're handing off information, circle that. And again, do not try and make this perfect. We're mapping reality, not like the magic dream land that we wish happened. So we'll take about three minutes starting now.

So if you're trying to make your way through, don't worry, you can keep working on this while I'm talking about other things too. But also I can tell you all we're very good in school because no one is like being class clown out there, probably because I'm standing up here. Awesome. All right, so in the interest of time, because I do want to get you all out to the after parties and fun vendor-sponsored events after this, we're going to move on to how do we identify the rest of the things we actually have to do.

For those of you who have seen similar talks before, this acronym should look familiar. I use this basically every single day. Far. So frequent, annoying, repetitive.

When we say frequent, is it something that happens every single day, multiple times a day? Is it annoying? Is it something like you don't want to have to do? Say writing follow-up emails with action items.

For me that's annoying. I'm also really not good at it. Very good thing to automate. And then third for repetitive, the way I like to describe this is it's not exactly the same thing every time.

It just has to be close enough. So I kind of call it like taking off my glasses. I take off my glasses, everything gets kind of blurry. It's the shape of the problem starts to look the same.

It's, oh, whenever I have a meeting, I take notes and I need to write a follow-up. The meetings might change, the audience might change, but the shape stays the same. So that's a repetitive. So we're going to be using far fairly often later on during this, but if you take nothing else from this session, just use this for finding what options you should be doing.

So that's for finding a large list of options. If you look at your week, if you start to look at what you do, you're going to find a ton. And that's how you choose what to go with. Impact.

So for the cycles, is it going to save you an hour? Is it going to save you five minutes? And with time, it's one of those funny things. If it's something that happens multiple times a day and it saves 10 minutes, that's more important to target than something you do once a year that saves you three hours.

That might seem like, oh, I should do the three-hour task. That 10 minutes is going to compound rapidly. Quality and consistency. Again, coming back to meeting note follow-ups.

If I have that automated and they know within five minutes of the meeting ending, they're going to get follow-up notes from Eric rather than, oh, did Eric remember to send out meeting notes? Quality and consistency vastly improved. And then capacity created. Does this give you back real time?

So if you're saving 10 minutes but that 10 minutes was going to not be useful anyways or you have to read through the meeting notes anyways while you're creating them, that's where you want to look at, like, what is going to actually happen with that freed up time, that freed up effort, that freed up brain power. And then most important, and this is especially when you're talking to C-suite, is there an outcome, especially a customer-facing outcome, that moves the needle? Are you going to be able to improve adoption? Are you going to reduce time to value?

Are you going to have people retain more? Those are big impacts and those are really good things to start with. When you're getting started, though, those small convenience wins, really, really great places to learn. So if you're like, I don't know how to get started, I want to do these big impact things but I don't know how to do this, start with the easy things, start with that low-hanging fruit and then you build up those building blocks.

So go ahead and take a look at your process, your workflow that you mapped, and we're going to do a few things with it. You're going to tag each step. Is it a frequent step? Is it an annoying step?

Is it repetitive? And does it have one? Does it have two? Does it have three?

And then for one of those steps, circle the one that looks like it's the highest impact and that's the one that we're going to use for later parts of the session. So look at what you've built, figure out if I was starting today, if this was the process that I was going to be automating, bringing AI into, this is the first step, this is the highest impact step I can take care of. And as you're wrapping up, if you're already done, feel free to get started with this next activity. I want you to take some time to share with the folks near you.

So share what you circled and why, and then we're going to switch. And what I want the person listening to listen for is, is this first step something that's convenience, something that's structural, like is it going to solve a long-term problem? Are we just kind of solving like, oh, that's an inconvenient thing we want to take care of. So share what you picked and why, and then for the listener, try and hear and identify structural or convenience.

And go ahead and take the time now as folks are wrapping up. All right, I hate to end good conversations, but... And actually real quick, does anyone have something that popped up as like a universal thing in the group? I do want to take a little bit of time for that.

I'm seeing we got a little bit of time for that. Anything pop up for your groups? I know putting you on the spot is always dangerous. All right, you'll keep your secrets.

So for all these things, what I see as patterns that emerge is these four patterns or some combination of them. It is, hey, we have a draft, it needs to be reviewed, and then it gets published. That's documentation, that's courses, that's any sort of communications that need to go out. Websites, all those things.

Second kind of shape, we have to intake something. Then we got to triage it, then we got to route it to the right person, and they're going to take care of it. That's tickets, that's support calls, that's even just like onboarding. Hey, we got to make sure onboarding gets done right.

That's the second pattern. Third, we got to summarize something, then we have to get some recommendations, then we got to escalate. Calls, slack threads, which get long, documents, meeting notes, all those kind of things. Fourth, and a lot of what my life is these days is we have to monitor something.

We have to alert if something trips that monitoring, and then we have to act on that. So, hey, some customer has stopped using a certain feature. Okay, I'm monitoring it, I got to take an alert on it, what do I do then? These four patterns, this is again taking off the glasses, a lot of things look like these shapes or some combination of them.

So, save this slide for yourself too. Another thing we think about a lot, and this is another trap I see a lot of folks walk into is you do not need to build an agent for everything. Put the minimal amount of brain power and inference and creativity into your workflows as possible. So, I kind of think about these as a spectrum of like assistant or workflow.

It's just going to help me in the moment, right? It's going to solve something, it's not going to do any sort of creative aspect. Workflows, that's going to be very consistent, that's in Zapier parlance zaps, but other tools like Workado and it ends, workflows, that's where you can get a lot done without having to bring in AI. Fourth, third, is agents.

So, what happens when we start to combine the chat GPT assistant and workflows, we get agents. They are trickier to set up, high impact, but also you don't need it for everything. Like meeting notes, that's basically a workflow with an AI step. It does not need to be a super agent that does those kind of things.

So, think critically and pick the lightest weight thing. And the main reason I'm saying that is guess who gets to maintain everything you build? It's you. And when you build a lot of things, it's kind of like writing code.

It's like who wrote this terrible code? And you go back and it's, that was me. I'm seeing a lot of the same thing happen with agents and as you build agents, as you rescappled things, it's a lot easier to replace a workflow than to replace an agent that's touching a million things. So, kind of a fourth activity.

You've got your tasks you're working with. I'm very inspired from this activity by my four-year-old because she asks, "And then what?" "And then what?" "And then what?" And I do love it. It's like one of my favorite things because I'm like, "Oh, you just have so much joy and curiosity about everything." Sometimes at bedtime, we do want that question to temporarily cease. But when we're thinking about building with AI, release your inner four-year-old.

Keep asking, "And then what?" So, start with that first AI system move. Okay, AI takes meeting notes. And then what? Okay, the notes go to the team that had the meeting.

And then what? They have to decide to take action. And then what? The action gets tracked.

Just keep asking, and then what? We're going to take some time to go through this, and feel free to do this as a table on your own, but I see a question in front here first before we get into that. [inaudible] Yeah, so I think this is one of those slides I really struggle with because there's two ways to think about this, is the alternate slide I use for this, which might be a better description, is if you think about the left as deterministic, like very much step by step by step. On the far right, you've got inference.

I'm going to have very creative agent-led things. Workflows typically live in that far left thing. It's very much, "If something happens, then this happens." And it's very step by step. An agent is going to take that same approach, but it's going to act more creatively on that data.

So, it'll say, "Do an analysis of the data, and then take one of these five or six paths." You don't necessarily know what you're going to get out of an agent. You can shape it a little bit more, but a workflow is going to be much more deterministic. And I have assistant in there because it's kind of like the, "I'm not going to build a full workflow because it's not repeatable, but I'm not going to build a full agent either." I'm just going to say, "Hey, chat GPT, how should I write this email?" So, it's kind of the piece you do for one-off work. Once you find yourself asking an assistant the question multiple times in a row, it's time to probably start building a workflow or agent.

All right, so back to our four-year-olds. Let's take a few minutes here to look at your tasks, look at that first thing that you identified, and then start building that, and then what chain. So, look at that action you selected, and then answer and then at least three or four times. All right, in the interest of time, because I'm seeing this countdown and I know I am the last session of the day, and you all have been very patient and very good audience so far.

So, let's keep it going before we get into rebellion. There is, frankly, not enough time to do this justice. This is a 45-minute session itself. So, what I'm going to tell you how to do today is how to use this.

In the QR codes on the front of your paper, you'll be able to grab this. The full worksheets are in there. We're going to do the lightest abbreviated version of this, but understanding this framework is going to help you build, build your agents, build your workflows, build your automations, and thinking about them in a good way. So, you know, walk through it a little bit.

We're going to give you a little bit of time to do it as table, and I want to leave a little bit of time for Q&A, because that's always my favorite part. It's like, what did I forget to put in this that you're actually curious about? So, there is an agent blueprint in your worksheets, but the first thing we always start with is the goal. What is the outcome of what we are building?

So, we've mapped the process, we've identified it. It's got good impact. Okay, what is the goal of what we are building? You've got to get that written down.

What are the inputs? What does it need to actually succeed in doing this goal? If it needs data analysis, if it needs meeting notes, if it needs a recording, what do we need and where does it come from? Tools.

What parts and systems does it need to read with? Does data need to go in and out of Gainsight, for instance, and how does that happen? Really, really important, especially as you get more into that inference side of the house. Guardrails.

What must it never do? I don't know if any of you have tried to have an agent write a customer success email. It loved to suggest, we should do a meeting or a workshop. Yeah, there's not enough time for every email to suggest a meeting workshop.

So, put in those guardrails, and before you send anything, test that the guardrails are working, that's the part I probably spend the most time actually fixing and refining is the guardrails. And then, really important, success metrics. How do you know this thing worked? For the meeting notes example I keep coming back to, we know that's successful when the action items from the meeting are actually done.

So, the first few times you run through this, you still want to run through it manually in a way and check, did I get the same result as the agent, or did it miss something, or did I miss something? That's a really good way to evaluate the success metrics, and then you can actually start to trust what you've built. So, we'll take the world's shortest time to draft your blueprint, just do a little bit of scribbling down, like how would I do this while it's still fresh in your brain? And then we'll do one last quick slide and bring it to you.

All right, take about 10 seconds to scribble down. Flashbacks for the ACT for anyone? All right, the last ask of the day is actually pretty easy, but it's important. A lot of these sessions happen, you come to this conference, it's like, I got grandiose ideas.

I get grandiose ideas every day, but conferences especially. We're going to make some little commitments, little commitment devices. Take a little bit of time, maybe during Q&A, book 30 minutes on your calendar and say, I'm going to build this thing, or I'm going to do something. If it's a process your team's going to do, message a stakeholder and say, hey, I was just in the session, this is what I want to build, let's work on it next week.

If you have team notes that you're taking, I know a lot of you are here with a group, throw in your team notes and be like, we're going to try this agent activity as a group. Those QR codes, you have these handouts, you have these notes, if you need anything, steal it. Go ahead, you have my blessing. I want to see people build cool things.

So do that. And last kind of thing, take a look at what you wrote on activity 1, come back to it, save it. The things that are impossible are only impossible today, and the people that are going to make the difference are the ones who are going to change and edit that list of impossible. So QR code there, if you want to say hi on LinkedIn.

I post a lot of things, most of them are useful. Some of them are memes. But yeah, you'll find me there most of all. So thank you and any questions?

[applause] [inaudible] Yes, so the question is, are there any good to know so we can just use them between agents and types? The best advice I give them that is you just have to try them out because also the answer changes week to week to week. Two weeks ago, Gemini had the best image model, and then suddenly Chat GPT had the best image model, and now it's back to Gemini again. It's like one of those things for that.

Generally, I find myself using Claude for anything that's like speaking, like thinking and copy and like planning, and I find Chat GPT and Gemini for code or image creation. So that's my shorthand on it. Yeah, so the question was, if you've built a workflow and you're trying to share it with your team, but they're running to blockers, they're not building it right, they're changing it, like how do you actually make it work consistent with the team? First, you try and build for that.

It's like what can I make as foolproof as possible? What things always stay the same? If it's going to be like a shared task list, not everyone needs the same, like their own version of the task list. You can point at shared resources as often as possible, and also there's certain pieces where it's like, oh, can I do this on your behalf as a team?

Like if everyone's team tasks from the team meeting just go to their shared spaces, that's a really good way to handle it. So like you own the automation, but it's still helping your team. And then just audit it consistently. Like I spend probably one day a month just auditing things I've built, because then they don't get broken as often.

Well, thank you. Oh, we've got one more. What's your big on having the LLM tool of choice helping you build this? Oh, yeah, yeah, so the question is, what about having like Chat, GPT, or Claude help you design and build this?

Feed it this worksheet and be like, I want to go through this process, it will do a great job of it. And putting it in something like plan mode for Claude, where it's like it's going to force you to think and answer questions, is a really great approach. And like, what a good use of AI, like force the robots to do the robot work and like make you a better human. All right, I'm not going to keep you all from the fun stuff.

So thank you so much. Appreciate it. Thank you. Have a good night.