Making AI Work for You: Airtable’s Framework for Finding and Building Real AI Workflows

Dude, I saw the ice cream in the hallway, and I was like, should I? Imagine I'm up here like chocolate around my mouth. I'm like, okay. I also asked if we could start the cocktail service early, and that was declined immediately.

So I was like, let's, here we go. So here to talk about AI, first person at the conference to do that. Just super quick intro about me. Oh, sorry.

Could you make the cheat screen a little bigger over here? Sorry. But just really fast. Born and raised Massachusetts.

I live outside of Boston. I have two kids. My wife and I, we love to travel. We try and do local breweries or whatever the best craft owner it is.

PMP certified. In a previous life, I hosted the Agency Unfiltered podcast. It interviewed owners, founders, executives of agencies and managed services providers. Kind of a controversial fun fact here.

I'm a big Dave Matthews band fan. The crowd goes wild. All right. Let's go.

Yeah, dude. All right. Let's go. That's the best reaction I've ever had.

Usually it's like, yeah, it's marching is great. What's your favorite live version? Recently named to the top 100 customer led growth leaders from base. I lead the customer led growth engine at Airtable.

Honestly, the big hit last year was my dog Sully. He's a 10-year-old black lab. So you got to get him in the mix. Then those are my two kids.

That's my wife for Halloween. We were the Barbie movie. So the bottom right is my son. He was a mojo dojo Casa house can.

And he's a stud. Okay. So I've been at Airtable for about two and a half years. Prior to that, again, I spent eight years at HubSpot Academy.

I was incubating a lot of enablement education programs within that team. And this is always like, I just like it. It's an interesting kind of look back. I find there's like three key phases of growth at my time at HubSpot.

And on the Academy team, it's obviously customer education focused. But you can just see over eight years, that's a lifetime in tech and in software. And so, you know, it was fun to help navigate that maturation. And so today there's three core pillars to my team at Airtable.

So I lead the Airtable Academy team. So that's self-paced on-demand courses, our certification program. That's hosted in Skill Jar. We have the Airtable community team who's here today.

That's our forums, our user groups, our virtual events, and our Airtable MVPs program. That's like our champions ambassadors. And that's on the customer communities product. And then the third pillar here is like our events and like AI experiences.

So this is like our in-person intensives. So like cohort style boot camps, like our build-a-thon, which was like our master class, or like hackathon style event, and some other things there. So that's kind of the gamut. That's the remit here at Airtable.

Cool. Folks are always interested to see, as I am of other programs and of other teams, of like our indicators, our KPIs, the data that we use to measure. And so probably like many of you, we have leading, and then we have lagging indicators. So here's kind of like the cumulative, like the tally numbers that we track and report on, you know, monthly, quarterly, et cetera, across those three pillars.

8.95 CSAT's pretty good, guys, OK? It's pretty good. And then here's like a quick look at the lagging indicators. I tried to like give a quick refresh and update to these based on like the previous quarter we just had.

So, you know, weekly active users of Airtable is kind of our North Star metric. It's specifically like the weekly active users of the native AI capabilities in Airtable. So Airtable is a no-code app platform. It's AI native, and so there's a lot of like agentic capabilities.

And so we want to inflect the AI adoption curve within our users and builders within Airtable for context. So you can see that, hey, after a course completion, we want to see an uplift in weekly active users. 30 plus, 60 plus days after they learned. We want to bring learners into the product.

We want to see uplifts in daily, durable, long-tailed AI activity and usage. But then for our own processes, like internally, we're trying to measure like the impact and influence of AI for our own capacity, efficiency, et cetera. And so, you know, we'll tease this out in slides to come. We rolled out this big production assistant agentic workflow, and we were able to add 125% output to our normal content sprint.

So we were super psyched about that. Right? Cool. Do folks here follow similar metrics?

OK, cool. All right. Who here has an AI mandate from leadership? Like we got to-- yeah, dude, yeah.

We have to AI it. We have to AI agentic. You know, dude, totally. I'm right there with you.

But it's just-- I don't mean that for you. I truly am. That sounded like apathetic, but like, no, I'm all in. I promise.

And it's really important to note that like the technology is there. It's already exceeding where we thought it was, where it would go. And so I know-- I don't know how-- I don't know what the resolution is for folks in the back. But basically, this is like the performance of AI, like chat GPT or like the models.

And so you can see just over like two, two and a half years or so, you know, it started at this baseline of like, OK, it is operating at the level of a human with a PhD, you're tasking them to answer a question or to learn something with the help of Google. So like a pretty baseline, obviously intelligent, scholarly, educated person. And now it's already starting to exceed the bar of asking that same human with their PhD to research a topic within their field. Like we can't stress how powerful, how smart, how intelligent these models are today.

And so just to like kind of bring it into a practical example of like, well, all right, well, what does that actually mean? And what does that look like? And so let's just like go through this idea that, you know, we are all working at Nike and we're trying to determine if we want to reshore production of one of our shoe lines. And hey, you know, should we do it to a company that they're sourcing from China?

And we, you know, there's a lot of like dependencies and intricacies and like, there's just a lot to think about there and so naturally, you know, we're gonna ask AI for their opinion. And so you can see the growth from like chat GPT 3.5, even out of Claude and now we're even seeing, you know, frontier models beyond this is that AI was really great at just like, it says simple summarization, but it's like regurgitation, you know, as anyone else feel that like the older models of chat GPT just kind of like validating your question. And again, actually it is pretty advisable to consider moving, you know, like, yeah, you asked me and I think that's a good question. Okay, great.

But now it's actually like becoming opinionated, data informed, research informed opinionated. And then it's actually producing original thought off of that, right? So it's like, yeah, no, that's important to consider. It's actually like, no, relocating is actually counterproductive and here specifically are the reasons as to why, right?

So again, just like example of how these models are advancing. And so, okay, great. It's really powerful, it's really smart. The intelligence is here, but it's also not slowing down.

It's actually accelerating at a more rapid pace than it was just in the graph we just saw, right? So even in just the past quarter, we saw some of these frontier models from all the big model, like the LLMs and the contributors, the open AI, the anthropics, the Googles, and it's not slowing down. And, you know, this was a slide I used maybe like twice and like all of the examples are already, it's already blown past it. Like we cannot keep up on these slide decks.

My PowerPoint downloads that I give the staff, like last minute, is anyone else like tinker with slides? Like to the last minute, it's like such a bad habit. Like we download them, but, you know, cloud opens 4.7 came out. But I also amongst all of this too, it's just like important to note, this is like the most overused sentiment and statement with AI.

It's like, okay, AI is not taking our jobs yet. No, I'm just kidding. It's not taking our jobs, but the people that effectively use AI or are harnessing AI and keeping up with the frontier, like they are the ones that will take our jobs. And so it's not gonna just come for us and replace us from a customer education perspective, not from a community perspective.

All it's doing is raising the bar for us, the practitioners, the builders, right? Cool. Kind of another angle here with AI that I think is like really important. And again, this is just like a lot of kind of contextualizing before we get into what this looks like for Airtable.

But not only are the models changing and evolving, kind of like the way in which you leverage and engage AI is also changing, right? And so I think obviously our entry point was just like chat-based AI, right? Chat GPT rolled out, you gave it a question, now we call them prompts, and it gave you an output directly related to that input. It lived in isolation, like within that chat window, that thread, but it was yours to copy, export, move, and then go do something about, right?

Like the humans still required action, right? And so then we started to move, and this is, I think, I don't know, a lot of folks might be here today, but it's actually going a little bit beyond just like the input output, singular isolated chat thread. Like, no, we're tapping AI at multiple touch points to assist kind of a complete workflow or a process, right? And so kind of the goal here is finding things that are repeatable, they're structured, they're monotonous potentially, AI is now becoming more and more embedded in the tools that we rely on for these workflows.

It's still human triggered, it's still human supervised, it's still probably predominantly human to like leverage an action and do, but it's clearly more comprehensively AI assisted, if that makes sense, right? Anyways, the third kind of, this next frontier where we're going today, I think it came up a bunch today, it probably comes up against tomorrow, but it's like actual agents, right? Agenic systems. And so this is actually pulling away from the human necessity of task ownership, the agent kind of knows if and when to take action, maybe beyond you even asking it to or invoking it to, it operates on dynamic environments and live data and it can produce actual outcomes, it can produce completed tasks, it can do it in the tools that you would normally operate, right, so you're actually getting hands off.

Does anyone else have, does anyone else have this agentic system level? Is anyone starting to think about it? Yeah, I'll tell you this. So I don't know, I like ramble sometimes, but when I did the dry run for the Gainsite team, this was like the one slide, they were like, the one note was like, keep like the roadmap slide, and now I know it's all like jungle maps themes, and so I get that, I get that.

That's, this took me a lot of time. That's not like a singular image, I had to find like, flat icon.com or something, and like little, and I tried to, I spent, anyways, I put a lot, that was a lot of work. Okay, I seriously needed to hear that. How's the ice cream?

Is anyone still working on their ice cream? Dude, two thumbs up over here, sick, all right, cool. All right, next few slides I'll go through quickly, it's just kind of like reinforcing this point. Yes, the world, the practitioners, our companies, we're moving towards agents, we want them to do the work for us.

However, we're very early innings with it, and very few organizations are set up to do this well, right? Software vendors are increasingly adding agentic capabilities into their tools. Many companies are trying to adopt and procure agentic systems in softwares, but they don't have like the telemetry or the data setup yet to like actually maximize it. A majority of companies are like, listen, like we're facing this like tech sprawl, like everything has AI right now, when do I use cloud coworker?

Wait, my, my gain site has, or Airtable has a field agent capability, like there's just a lot of competing functionality right now, and companies are having a hard time figuring out how to operationalize it, right? Does anyone else feel that way? Yeah, dude. And it's a very small percentage of companies with this opportunity are taking advantage of it, right?

6% of companies report using AI for regular, like program management and production today, but it's pretty clear like that there's value in doing so, right, so it's like, oh man, it's clear as day that there's time savings and efficiency gains in ROI, however, we can't get our house in order to make it happen, right? And so anyways, all that to be said, I don't think mandates and strategy, we need more AI, like that's not gonna scale and that's not gonna find the success, it's us the practitioners putting it into the right use cases, right? The gap isn't the model, the AI providers, it's those that are feeling empowered and enabled to experiment, and then those that are pulling ahead in this like AI arms race are the ones not doing the most AI, it's doing the right type of AI, right? Dude, who's with me on that?

You guys with me on that, all right, cool. Yes, listen, you're at the top of the list right now. Whoever did that, who is the number one guest right now? Okay, so a couple of dimensions when trying to experiment and like find the right place to like truly experiment with AI, right?

And so you kind of first start with this idea that like there are just naturally some things that AI is good at, and there's obviously gonna be some stuff that AI is bad at. And so here's just like a short list. There's more probably, but these are the ones that at least like our team at Airtable kind of like lean into, right? So you have like summarization and synthesis.

Oh, sorry, I can click one. I have some, I don't know how small that is, but like some EDU or some community examples. All of these examples are things that our team does today, right? So like these are real, these are practical.

But yeah, it's good at classification and like prioritization. It can personalize at scale. Obviously the long tail and the jungle was the big story this morning. It can automate and orchestrate workflows.

It can enrich your data. It can assist you with research. It's really effective at prediction and forecasting. And then obviously, as we all know, kind of bread and butter content creation, content generation, right?

But that's not the only dimension I think that's important. So yes, we need to know where AI can plug in really effectively. But we also have some guardrails to understand what makes up a good experiment or like test case or like first use case for AI. And so here's some, again, some of the dimensions that we rely on when trying to map the opportunities.

Sorry, I can click ahead one more there. There you go, some examples. But listen, you don't want like super complex data integrations, like yeah, maybe eventually we wanna overlay our Academy data to our product usage, to the renewal dates from our CRM. And like that's gonna be deeply powerful, but that takes a lot of time and energy and buy in.

And so that shouldn't be a blocker to get started, right? You shouldn't rely on anything that's gonna take like a super hefty change management flow, right? Like it's gonna be painful and laborious to get everyone enrolled. It shouldn't rely a ton on like high risk, like high precision, right?

Like, hey, having AI generate your earnings call, you know, transcript is gonna be different than like, hey man, can you just like summarize our meetings and like kind of recommend next steps? Like there's degrees to precision requirements, right? Is the data available to you? Is it owned?

Is it trustworthy and reliable? Will it be quick to see the aha moment, right? It's not gonna be like super lagging or delayed of like the value realization exercise. And then is it something that you can own in house, right?

It's like departmental, it's not gonna be, you know, dependent on effective cross-functional collaboration. XFN is cross-functional collaboration. That's come up a couple of times. Does anyone else use that acronym?

I don't know, dude, yeah. We don't have to woo for the acronyms. I'm just kidding, you can woo whatever you want. Yeah, thank you.

All right, all right, chill, chill, chill, chill, chill. Okay, so those are the dimensions, right? It's like, okay, what's AI good at? Where do we know it fits well?

And so now, you know, you can run a list, you can ideate, you can do a team brainstorm, like, dude, let's list everything that could be a place for AI, right? But you can't do it all at once, right? You can't peanut butter spread. Okay, peanut butter got one, gotcha.

And so there's a couple different prioritization framework. Dude, it's like at the end of the day, you know? All right, it's like hot. I was supposed to, so I had like a blazer.

I was gonna be like, all right, I'm like, you know, before my session, I'm gonna change. And then I was like, nevermind. So you're just getting what you get today. Okay, it's in my suitcases down the hall.

So there's prioritization frameworks to help you here. You don't have to do them all. They're not sequential and stackable necessarily. Kinda pick one that works for you.

They're pretty straightforward. And then like chart the AI opportunities on there. And so the first one here is just a simple two by two, love a good two by two. Okay, what's the implementation effort?

Low to high. And then what sort of like impact will we feel from that? And as you can imagine, the best place to start would be any of those AI opportunities that fit in the high impact, low implementation effort quadrant. Best place to start, obviously.

And so another way to think about it too is that like, yeah, those are the immediate ROI, right? Like that's where you will feel and see ROI. However, if you were to start to like, you know, phase these out or horizon them out, the high impact, high implementation effort, like, nah, I wouldn't ignore it. I would just say, well, it's not so much quick ROI, but we should be putting the investment in there.

Maybe there's a little more research and development. And so that would be like a good, you know, phase two or horizon for you, but just map to the two by two, right? Cool. I think I can see, just so you know, I can see kind of the Slido questions as they come in.

Nobody's asked me about Dave Matthews Band yet, and that's offensive. Okay. So another framework here is just REV repetition, repetition effort value. I find that like everybody thinking about AI is starting to converge roughly into the same prioritization frameworks.

It's just a matter of like, well, how are you gonna economize it? Or like, what are the labels you use? So we use REV. It's like, yeah, how often does a task occur?

How much effort does it take to do that task? And then what's like the value or importance or return on delegating that task, right? And so, yeah, just generally, you know, obviously you'll probably have something a little less generalized, right? Like you'll have kind of the measurements to include here, but these aren't so like staged.

It's just a spectrum, right? And you don't have to just prioritize something that's only strong fit across the board. You're kind of just like loosely scoring, you know what I mean? And so maybe this is a good fit, whatever this use case is to prioritize and give it a whirl.

Okay. I think there's a few other things at play here that are important. I kind of just like put a subtitle slide in here because I don't really know where else to put it in the deck. But, and let me know if you all agree with this, like AI clearly in my mind has transformed the way in which we approach program development, right?

I think for a really long time, we just anchored to this idea of like the traditional waterfall, right? And it's like, okay, first things first. We establish our business requirements. We run thorough analysis.

We're gonna design the solution or they're gonna, you know, go through the development and creation process. We're gonna test it, QA it, and then we're gonna let it rip. And then we're on to the next one, right? Pretty standard.

However, AI has kind of pushed us into this new era where it's like, actually not, you know, forget that. It's all about rapid experimentation. It's about rapid iteration. It's how fast can you get an MVP out, right?

And so we've kind of crunched this middle section where it's like, nah man, design quick, get a V zero out. And then just once it's out, iterate, iterate upon it, optimize, move fast. And so you start to get into a little bit of like a, a cyclone in this middle area, right? It's not so like monolithic and sequential and like chapter start, chapter end, right?

Has anyone else felt that? Like felt that? Yeah, no woos for rapid iteration, just a validation. Okay, woo.

You know, this, it's interesting. I think like the way, we just talked about the intelligence of AI and like where it's going. I think this, the art of the prompt is slowly going away. You know, like there was a moment in time where there was a ton of discourse.

It's like, okay, it's actually prompt engineering and it's a very, you know, there's a whole thing to it. And it's actually an art form and a science. But like, I think there's still some best practices that are important to know, right? Like, yeah, dude, calling Waymos is awesome, but I do want to know the rules of the road, right?

And like, I want to know, so anyways, I say that just, you know, it's, you still should understand kind of the principles behind a prompt. However, some of the most effective prompt thing we do at Airtable is relying on AI to help us craft the initial prompt, right? Pretty meta on that. So three elements of a prompt.

It's the medium, what am I looking for? What's the output? It's the data. What am I putting in?

What can it access? What can it reference the context? And then it's the direction, right? It's like the prompt 101 is like, okay, start by telling AI what it does.

They're like, what its job is, right? And so do that, it's the direction, and then like, what are the tasks? What are the things that it has to deliver, right? Tell us more about the Matthew's bid, thank you.

Okay, great. More ice cream for that person. And so just like really quickly, like, let's just say, you know, we have kind of this prompt going, and we're gonna ask for like executive summaries of like, I don't know, some deep analysis that we want to bring to our leaders or our bosses' bosses. And so it's like, yeah, all right, you're clear about the medium here.

You are, you know, an analyst, and we want comprehensive executive summaries. It's gonna be structured. It's gonna be plain text, because when you do a bunch of bullets, and like structured AI loves that too, where it's like titles, subtitle, bullets, emoji in the titles, and like, when you paste that into a document, it's a nightmare. So I always, like, I'm literally super pointed about, all right, don't do that.

I want paragraph style plain text, but the medium, right? I'm giving it the data. This could be like an attachment and upload. You can just point it in the right direction.

You know, if it's like doing web research, you can point it to a URL or something like that. In tools like Airtable, you can have it, like reference a certain field in your table or your database. And then obviously like the direction, right? Compile the findings, make it structured.

I want their professional background, the recent news about them, any other relevant context, and then, you know, clarity when there's like a, if it's not working, outlet, right? Cannot find the information to do this well, generate insufficient data. Like don't hallucinate and make it a nightmare to share. Right?

Okay. A few other just quick notes on prompting patterns. These are super easy to incorporate. Maybe you all already do, but it's just worth like quickly ripping through them.

The first is like chain of thought. So you're actually asking it not just to deliver the output, but actually show its work down the sequential task, right? So like, hey, I don't want just that structured executive brief. I want you to compile a bulleted list of references that you pulled from.

I want you to, you know, put a tape, I don't know, whatever, I'm thinking of an example on the fly here. It's the end of day, but you get what I'm saying. Does everyone get what I'm saying? Okay.

That's such a cop out. I'd be like, you guys get it. All right. All right, the next one here is thought.

This one's super easy. I don't know, like honestly, first off, we don't have to go down this rabbit hole, but the biggest concern I have with AI is like, I call it cognitive degradation, or like, dude, I just can't think strategically anymore. I'm like relying on it too much for simple tasks. And so like, anyways, to say that is like an example of, I don't know, man, I gotta send an email.

We're launching a new certification. Help me with like a subject line. And rather than just say, help me with the subject line, it's like, nah, man, give me like 15 subject lines. And then I end up like kind of building a blend, but truth of thought, ask for optionality to whatever the output is, right?

If you shot, you're actually, if something you know is tried and true, there's a template, there's an output that is, tends to work really well. Literally give it an example of what you want to see, right? Hey, man, like, I wanna do a recap post to LinkedIn. I just spoke at polls, it was awesome.

Here's like an example post of last year, just so you got like my tone, my voice, the structure of these posts, like literally just paste in an example and Zolk can emulate it. And then the last one, which is really interesting, is like self critique. So, hey, man, this is how I talk to chat. Hey, man.

Hey, dude, listen, it's Kev. But it's, hey, before you consider this case closed on your output, actually like self critique, like, hey, that executive summary as part of the output, incorporate a few places in which you think it could be a little clearer or a little more concise. I'm always asking for like more concision or brevity, but like have it self evaluate. And then you can see that and then usually it's like, all right, great, apply those recommendations, right?

Are these familiar with folks? Do folks use these today? Is anybody like ripping through all four of these? Same, same, okay.

The other kind of component here with prompting, and this is where you start going from like prompt to assisted workflow to agentic workflows, is it's just like this idea of chaining, right? And this is like, I think, you know, it's not super complex, but it's deeply powerful in that, okay, we have one initial task for AI. And it's like, hey, man, hey, guy, customer feedback comes in, a survey, CSAT, whatever course or something that comes in. I want you to run some sentiment analysis, audit it and then just categorize it as a positive or negative sentiment, right?

Great, then that could be a workflow. However, that sentiment rating is actually the input to trigger the next step of this AI workflow. And great, now that we have a sentiment for these surveys, start drafting an email that is wrapped around that sentiment, right? So you could be have an empathetic tone or a grateful tone, or whatever, right?

But it's like, again, the output of task one is now the input to task two from an AI perspective. And great, that gets delivered to the person that has the action or review or approve the email drafts. You know, they can click the thumbs up and maybe you have an automation now built off of that to trigger and send all the emails, right? And then boom, so that's just prompt chaining, right?

And so like, you can just do this, sequentially, sequentially, sequentially, across multi-dimensions, right? So maybe one output triggers three next steps. And so this is kind of where it all starts to string together for like agentic AI workflows, really, right? Because I know I really sold you on the map.

Slide. This is a sick slide. Because like, so first off, I'm like, we're gonna run out of time, I'm so sorry. But just so you know, we have to, I had like the gradient, look at that, the gradient of the things.

And then I was like, all right, this is an example. And then I used the same, dude, that's sick. All right, all right. Put that in the surveys, be like, that one slide though.

Four stars, no. So again, going back to like this push towards agentic workflows. Again, we've talked about prompting, one in, one out, conversing, right? It's kind of isolated in, we talked about medium and data and direction.

Then you start to go into this workflow, maybe you're chaining, the model executes against the steps that we've determined. It's repeatable, it's structured. And again, it goes back to that chaining slide. But now, again, where we're trying to go is agentic, right?

And so it's actually like, I'm gonna brief my colleague or the person that's gonna own this, the operator, right? And so they're actually maybe gonna decide the steps to take, how they're gonna structure it, what tools they're gonna use to make this happen. And it's not so wired and rigid. We're just briefing the operator and we're gonna see how this pans out on the other side, right?

And so this is kind of like the reliable structure that we've landed on. I bet many folks doing this also do it similarly. But it's kind of this weird like, well, first off, the first agent you build is like the lead agent, right? They're the judgment, they're gonna kind of orchestrate, they're gonna delegate out to specialists sequentially.

And so you have a lead agent that like, hey man, that's where the input goes. And then now you have a team, a fleet of sub-agents underneath that, which are like specialists for very specific pieces of this workflow. And we'll get into an example in a moment. But this is a super reliable setup for us when we start talking about like agents, right?

You have the lead, our production assistant, and then you delegate and manage the specialized operators underneath, right? And then ultimately, it all gets bundled together and circled back to you, the person that put in the original request to review or action or approve, right, whatever it is. And so from the Airtable Academy side, again, we have a production lead bot. Sometimes we like to humanize them and give them names.

We'll just, this one's production lead. And so we invoke the production lead or the production assistant over Slack. And it could be us, it could be an automation that funnels it into a Slack channel. And it could just be, you know, the next topic that we wanna build kind of a draft content bundle for for Airtable Academy, right?

And so now what used to take three to five collaborators, we have our team, we have a contractor, we have subject matter experts cross-functionally, and this is gonna take two to four to six weeks. It's a bundle of assets back in literally 10 minutes, right? And so this could be like, hey, I want to build a lesson for an Academy course on how to mitigate, I don't know, context rot when, you know, over prompting AI, field agent, or whatever it is, right? And so you put that in, great.

It delegates the script review, the script creation to our script reviewer bot, and we built the skill around like, these are the really effective use cases and we always wanna ground them and in real world practical use cases of Airtable, right? We have another specialized operator for deck design. We've taught it, we've uploaded the skill for our brand design, our brand kit, colors, dimensions, iconography, all that good stuff, right? We have an e-learning designer.

So we are probably like many of you are trying to build kind of our e-learning assets around AEO best principles. There's kind of a point of view on that now, right? Like, hey, statement, embed video, key takeaways, term definition, FAQs, right? And it's like, you know, we want this rinse and repeat for every e-learning asset that we create and we've built the skill for that.

Quiz builder, and we also have a skill for like psychometric evaluation. I'm terrible at quiz question creation. And so I'm like so bad, like, hey man, the distractors aren't the same kind of like sentence structure, it's like an easy tell. And so it audits and does all that for us.

And then oftentimes for an Academy course when it's Airtable product specific, you know, we're gonna have like a screen recording or like a demo, like a walkthrough of what this looks like in Airtable. And so now an agent will actually build us in Airtable base based on the script, based on the concepts, based on the fictional business, the departmental use case, et cetera. And so again, these are all just specialized agents. We've trained them on the skills to do this consistently and repeatedly.

And the production lead at the top is the one managing and delegating the work, right? And this, we have this going today, right? Cool. On the community side, we have found that managing and organizing events is like just incredibly time intensive.

And we have user groups and we're trying to work with our user group leaders for virtual events. We're getting pulled more so into in-person events, diversity of event types, we have master classes, we have meetups, we have hackathons. And so it's like, something's got to give with capacity and effort and task ownership. And so again, similarly, we can have a lead agent kind of act when a new event gets submitted by our user group leader.

And so we'll have like the webinar coordinator integrated into Zoom, or sorry, webinar platform, no free ads. I did that on purpose. I was like, I'm not gonna name any companies on this. But anyway, so it connects to Zoom and it builds the Zoom webinar link.

Gives us like a recap synthesizer flow after the event. And so it can post a recap, right? From the transcript, clips, information into our messaging app. It rhymes with flack.

Hey, find the most juicy video clip, right? Wistia, or a company that rhymes with Flistia, has the ability to like find and kind of auto generate impactful clips. And so again, like, hey, drive that and deliver it for us. We also want to kind of remix and leverage content we capture from these events in other places.

We have a like a builder spotlight series we call change makers. And so this is just like ripe for that, right? Builders are the heroes of the air table story. And we want to get that blog cadence going on the community.

And so, you know, we can do that. And then, you know, based on who attended, right? We can integrate our CRM data. We know accounts are there.

We know who their account executives are. And we can help them with some follow-ups and contextualize, you know, email templates, et cetera. And our CRM, which is Flails Norse. Okay, that was tough.

I'm out of Salesforce. Now all that to be said, we don't just like, yeah, go. And now we're like kicking it up and just relaxing. There are some checkpoints and like human in the loop is not going to go away and we have to monitor and maintain.

So just like a couple of best practices here before you just like completely, you know, build all these agent workflows and you're chilling. First, we do, you know, we're very thoughtful about the scope of what an agent can do, what tools it can touch. It's not just like, hey, integrate into air table. It's like, okay, wait, all of air table, our entire company's data, or are there very specific kind of pockets of air table, right?

So it's platforms, but also specificity within the platforms. We have checkpoints, right? So that Academy production lead, it doesn't just like publish right into Parda and Skilljar and then publish and then it's out there in the wild. Like we do want to look at it quickly and validate and approve.

We tried running, literally we ran like eight of those. Production agent flows like at once and like it just completely fell off. It turned into a nightmare at the end. So we're like, okay, one at a time and we are gonna look at it before it gets published.

We're trying to, like basically like, you want to avoid irreversible actions, you know? Like, hey, it just deleted our entire Academy learner data set, like sick, okay, no. So we want to avoid that destructive behaviors. And then lastly, evidence, right?

Like there has to be traceability, audibility. How do we know when they're working well? How often they're working? I think a big conversation right now are like costs associated to running these agentic workflows.

We just talked about these like super front edge, frontier models, but they don't have to be doing the like super basic sentiment categorization. And so it's like, all right, what's the sweet spot of model to task intensity? And so yeah, there's some of that there too. So anyways, there's checkpoints and there's parameters here.

And so ultimately not everything is great for AI as we know from the top. Therefore, not everything is great to just completely move into an agentic workflow. And so just like super kind of quick dimensions here of, hey, when should we delegate this workflow to an agent operator or fleet of agents versus like, when should we kind of keep it on the AI assisted workflow slot side, right? And again, it's repeatability, you know, does it happen enough where the juice is worth the squeeze to systemize?

Is it verifiable and are there natural places in which we would interject a human? Does it feel like it's super clear when it is done or is it ambiguous? Is it always going, right? You kind of want like definitive guardrails.

And then obviously errors are recoverable. Again, you're avoiding the destructive behaviors. Anyways, you know, just some things to check off. Okay.

We're bringing it home here. Key takeaways. So that's Sully. Last time it was like the feedback, the surveys and stuff, they're like, could you use a little more Sully?

You know, and so I was like, let me throw them on the last slide here. But I did legitimately like build takeaways around them. This was literally the slide that made me be like, can you re-upload my slide deck? Because I kind of like did another Sully bit.

Anyways, brainstorm your own AI use cases, right? We know what AI is good at. We know the dimensions to build around those opportunities, right? Pick a framework and prioritize those opportunities.

So you have a punch list in order of impact versus effort intensity. Optimize experiments around good enough, right? Rapid iteration, rapid experimentation is the name of the game. You're trying to get an MVP out so you can get early signal and continue to optimize, right, against traditional waterfall.

You start with this like prompt to AI assisted workflows to agent, right? This is kind of the progression. So you should always keep that in mind. You want to delegate when you get to the agentic layer of this whole thing, delegate out to a specialist fleet of agents, starting with kind of a lead conductor at the top.

And then you're earning trust and you're building defendability and you're building durability as long as you keep in mind, right? The scope of the agentic workflow, you have checkpoints, and then there's evidence of its performance, right? Sweet, okay. So we have a few minutes here.

So I have some questions. If we could flip over to the slide, oh. Do you want me to take that? If you want me to read the questions out or you want to take the questions out?

Yeah, read them. Yeah, that sounds great. Does anybody else, sorry. Are there any volunteers to read the questions?

All right, you're good, you're good, you're good, you're good, you're good. You can get rid of it. You can get rid of it. Yeah, come on.

I'm sorry. The Dave Matthews one did get voted down. That one is not on the screen. We lost that one.

Yeah, so, but anyway, first things first, ladies and gentlemen, huge round of applause for Kevin. Thank you. I mean, that was truly as tactical as you could get. You could probably not go any further.

He literally gave you everything that you need to know. So thank you, Kevin. I gave you everything I had today. There you go, there you go.

Yeah, it's the dog, you know, yeah, yeah. That's probably a Dave Matthews thing, I don't know. It should be. It should be, okay.

So the first question here, which got a few votes, is that a dedicated team for AI work at Airtable? Second part, it seems like every team is building on their own. Do we all like run towards one goal or should some larger strategy be applied? Yeah, so there's no like dedicated like AI transformation team, right?

I think what we're finding obviously is that, you know, AI champions are emerging across departments. And so, you know, you kind of have this like sub community of our AI champions and, hey, how do you continue to have them kind of lead from the front and help empower and enable and continue to upskill their team? There's like formal channels to do that. There's also informal channels.

So it does kind of happen, distribute it as decentralized. However, Airtable does have some company-wide programming in play to really encourage experimentation. So just really quickly as an example, we're kind of in this kind of like experimentation phase slash competition called Air Loop. And so it's like literally shut it down, focus time and we want everybody to try and build at least one agent based on all of these dimensions and the criteria.

And then, you know, there's like some competition and there's some prizes based on, you know, agent leverage and impact. So it's company-wide priority and they've created the space to do that, but it's very much the actual agent building and the AI work is departmental and decentralized, if that helps. That's a great answer. But we have been seeing some patterns across the past few sessions in that there are different levels of, you know, the way that organizations have implemented it.

And so this is about as valid as any of the others as well. So thank you for that. The next question is, where are you building most of your agents? Is it within Airtable?

Yeah, so again, we kind of have the benefit to dog food Airtable. So Airtable has this emerging product. It might be a made or something. It's just called, it's Hyper-Agent is the name of it.

And I think like the closest kind of, like the closest comparison software is something like OpenClaw. If folks are familiar with OpenClaw, we're trying to be like, hey man, that's a little Wild West. That's a little cowboy out there. And so there's probably a little more secure, stable way to do it.

So we oftentimes use Airtable's agentic capabilities. I think Clod Co-Work is obviously another really effective way to do that. And we have some workflows going through Clod as well. I would vote for Clod Co-Work as well.

I would vote for Airtable. I'm just kidding. (laughing) Well, by the way, Clod doesn't pay me anything to talk about. Airtable does pay me.

Yeah, they do. They do. Not enough. No, I'm just kidding.

So there is another one that is related. So I'm just skipping over to the next question. What AI tool should I be asking my leadership for access to in order to get a production lead agentic model going? Yeah, honestly, I think it's kind of the similar answer as the previous question.

I think the best way to do it is combination of Clod and Clod Co-Work and Airtable. But Clod Co-Work is probably the best place to start, for sure. And probably we could end with this one, which is to do with learning, but it has to do with AI certification. So somebody wants to know, at least a few people want to know, have you taken any AI certification, and what would you recommend?

So there's a completely free, high-impact, high-leverage, self-paced, on-demand course catalog and certification program called Airtable Academy. I highly recommend. We are launching a certification on agentic systems design. This is like shameless plug-sees, and so thanks for having me through this in.

We're launching a certification, actually like this or next week, agentic systems design. And so we're trying to teach some of this but outside of Airtable, I'm a hands-on learner. So just getting into the tools, throwing spaghetti against the wall, that's just how I've learned, and so that's what I recommend. But obviously we know Anthropic, another Skill Drawer customer, is probably a really great catalog.

I think Google has some stuff too. So it's definitely around, but I don't really have a great answer for it, unfortunately. I highly recommend the Anthropic Clod 101, if anybody is interested in that, that is literally kind of creates your foundation. If anybody does learn with Airtable Academy, make sure you become a weekly active user of our AI capabilities.

It's gonna really do wonders for our North Star KPIs. Do what he's had now. There was one other question, but the gentleman who asked me that question didn't know, and this is like an aside. So in all your transition slides, there was a dog on the bottom right side.

Yeah, all right. The question was, why were there bugs flipping around its butt? It's a really important question. So I have a couple of thoughts on that.

How much time do we have? We have enough time. First off, it's also like a, it appeared to be a yellow lab, or a yellow lab adjacent, or breed. And I'm firmly a black Labrador retriever family.

So those were the graphics and iconography that they gave us in the slide design. So I'm not happy about it. So yeah, that dog probably should have some fleas. Maybe that's a question for our events.

Yeah, put it in the feedback. Yeah, but thank you so much, Kevin. This was absolutely amazing. Thank you.

(audience applauding) So I will not be at the party, only in spirit, but I'm hanging around, I'm here as a, yeah, yeah. Come find me. Absolutely, but thank you so much. Of course, thank you.

All right, with that, everybody, we'll see you at the Pulse party. And tomorrow morning, do not miss the keynote. It is probably as exciting or even more exciting than what you heard today. So you'll actually get to see everything that you heard about today.

You do need your ID. We will ID you. Whatever you look like, you will get ID. So definitely come with your ID and your badge.

They're gonna lock out anybody that didn't give a five-star review to this session. So good luck getting in. Yes, there you go. Unless, yeah.

Thanks, guys.