51黑料不打烊

Transcript

Hi and welcome to Building an AI-Powered Lifecycle Engine for Marketo. My name is Josh Arrington. I鈥檓 Chief Marketing Technology Officer for Capturall. We are a Marketo-focused agency based in Madrid, Spain.

Just to start off, I鈥檒l tell you a little bit about myself. I鈥檝e been working with Marketo for over 16 years now. I鈥檓 a certified expert and architect, and I鈥檓 honored to be part of the champion program this year for the second time. I am a full stack developer and I have built probably hundreds of custom integrations and services for Marketo, but please don鈥檛 let that scare you. I assure you that everything we will discuss today about building an AI agent to manage your read lifecycle in Marketo can be done without writing a single line of code.

Just a quick overview of what we鈥檒l cover today. We鈥檒l start with a quick introduction, then we鈥檒l dive into the differences between traditional MQLs and what we鈥檙e calling AI-qualified leads or AIQLs. From there, we鈥檒l look at the concept of agentic marketing automation, what it means to move beyond static rules and put reasoning at the center of your process. We鈥檒l walk through how to set up an AI marketing agent and how to equip it with everything it needs. We鈥檒l also touch on how to add human oversight when it鈥檚 needed. Then I鈥檒l share some of the top tools available right now for building these kinds of agents. We鈥檒l wrap up with a quick look at what鈥檚 next, some practical rollout tips and a few key takeaways before we jump into the Q&A. So let鈥檚 get started.

Imagine you are an international corporate real estate firm. You work worldwide and cater to clients from startups to the Fortune 100. Your sales team is divided into regions and by company or property size. You generate hundreds, maybe thousands of leads and you try to qualify them before sending them to sales using traditional methods like scoring, but sales continuously complains that the leads are poor quality, sent to the wrong person or that the good leads are sent too late. They say that leads are sent to them without enough information to go on and now they鈥檙e requesting that you hire people to review each record manually, but that鈥檚 just not feasible in terms of time or budget. So what can you do? The problem is that traditional lifecycle management with lead scoring and thresholds for reaching MQLs is good, but it has its limitations.

Static models lack adaptability. Point-based scoring and fixed lifecycle stages can鈥檛 adjust when lead behavior changes or market conditions shift. You need to create smart campaigns with specific rules and they can quickly become outdated. The process can be quite fragmented as well. The system qualification and lifecycle progression are usually handled in different places inside Marketo and although they should work together, they often get disconnected from each other and that leads to inconsistent handoffs and poor coordination between marketing and sales. Traditional systems also lack context awareness. Typically, data points are considered in a vacuum, evaluated once and not seen holistically. Things like company size, intent signals or recent behavior aren鈥檛 considered together. The result is that leads are mis-scored or acted on at the wrong moment.

Traditional systems are also high maintenance. Every time your strategy changes, new markets, new products, you need someone to manually rewrite scores, the scoring rules and adjust lifecycle definitions. It鈥檚 slow, it鈥檚 tedious and it鈥檚 prone to delay. And finally, there鈥檚 also the lack of transparency. Sales teams often don鈥檛 know why a lead has been marked as an MQL or its history in the lifecycle and that lack of visibility damages trust and hurts alignment. With AI, we have an opportunity to look at each lead holistically and make a common sense judgment. We can equip our AI with our ideal customer profile, our ICP, and detailed guidelines for not just how to score the leads but how to evaluate them. We can then give our AI the data we have for the lead, that鈥檚 demographic data, firmographic data and the full activity history. If that鈥檚 not enough to make a decision, we can give the AI tools like LinkedIn, so it can look up additional information from the web and get the full picture. From there, we just ask the AI to apply the guidelines we鈥檝e given it to evaluate the lead. We can then give the AI a handful of actions it can take and ask it to take the best one based on its evaluation. Even better, since this isn鈥檛 a black box of scoring and transition rules, instead of sifting through activity logs to figure out how a lead got to the current score or to its specific stage, we can have an AI explain its actions. So an AI qualified lead doesn鈥檛 just appear in a salesperson鈥檚 queue, it arrives with a clear explanation of why it鈥檚 qualified and even which products or services are most relevant. It sounds a bit like science fiction, but it鈥檚 real and you can implement this into your Marketo instance today. So what does an AI powered lifecycle look like? Well, it looks basically like this. Instead of transition rules and smart campaigns, we have an AI agent making real-time evaluations and orchestrating the movement between stages. It鈥檚 as if we have a person carefully reviewing all the records in our database one by one and taking the correct actions in real time. And even better, it鈥檚 AI. So it works super fast. It never gets tired, never takes breaks.

So what is this magic AI agent and how do we get it working in our instance? Well, let鈥檚 start with what an AI agent is. AI agent and agentic are buzzwords and they鈥檙e getting thrown around a lot lately. But what are they really? Let鈥檚 start from a place we are all familiar with and that鈥檚 workflow automation. This is like a Marketo smart campaign where we set triggers and filters and specific actions. We might have some if-then logic in there, but for the most part, it鈥檚 a rigid system with predefined rules and actions. An AI agent on the other hand has a list of actions it can take and is given some instructions and when it鈥檚 prompted or triggered, it will use its reasoning skills to decide which actions to take. Agentic AI is the next evolution of this, where the AI is continuously thinking and evaluating. It takes actions proactively, learns from the outcomes, and retains memory and context over time. In essence, it鈥檚 an autonomous digital coworker. For our purposes, traditional MQLs fall squarely in the workflow automation camp, while AI qualified leads can be between AI agent or agentic, depending on how far you want to take it. Which leads us to our next point. What goes into an AI agent? I like to think of AI agents as AI interns. They鈥檙e smart, they鈥檙e capable, and eager, but just like human interns, they can鈥檛 succeed on their own. They need structure. When we build an AI agent, we have to set it up for success the same way we would onboard a new team member. That means giving it three things. First, we need to provide it with clear instructions. AI doesn鈥檛 know what qualify lead means in your business unless you spell it out. We need to break down the steps, the logic, and the decision points. Just like training someone on your team. Second, the agent needs knowledge. And in this case, that鈥檚 your ideal customer profile. Your lead categories and any documents or examples that show what a good lead looks like. This is the business context it uses to evaluate and reason. And third, we need to equip it with tools. Ways to interact with the data and take action. That could be enriching leads using LinkedIn, reviewing activity logs, or triggering actions in Marketo using something like Request Campaign. The good news is once we give the agent these things, instructions, knowledge, and tools, it can operate at scale, making consistent decisions and freeing up your team to focus on strategy instead of manual lead review. Now, I said there are three things your agent needs, instructions, knowledge, and tools, but there is a fourth thing here at the center. And that is a brain. That鈥檚 hopefully something your human intern arrives with, but in our case for our AI agent, we need to give it a brain.

And that鈥檚 in the form of an AI model or LLM. Like a human brain, it brings in with it a vast amount of general knowledge. Some of it useful, some of it irrelevant, and some of it just plain wrong or outdated. It can reason, but it doesn鈥檛 automatically know how your business works. That鈥檚 why the real performance comes not from the model alone, but the clarity of our instructions, the quality of the knowledge, and the tools that we allow it to use. And here鈥檚 the best part. Unlike a human intern, we can actually swap out the brain anytime we want. So as new models are released, smarter, faster, or more efficient models, we can upgrade the agent鈥檚 capabilities without rewriting everything else.

Let鈥檚 look at how the full picture looks when we build the agent itself, and we鈥檒l unpack each of these. Here you see the four components you鈥檒l need to put in place. First, we give it a brain, and that, as I said, is the underlying model, OpenAI, Gemini, Llama, whichever foundational model fits your needs. Then we load it with knowledge. In our case, that means our ideal customer profile and other business context. Next, we provide it with tools. These are the things the AI can use to do its job, like calling LinkedIn enrichment, querying lead activity from Marketo, or triggering a smart campaign in Marketo鈥檚 API. And finally, we define instructions. This is the logic and reasoning process, step-by-step guidance for how to analyze a lead. So just like a human intern, our AI agent has a brain, is trained with business knowledge, is equipped with tools, and is given clear instructions on how to do their job.

When selecting a model for your agent, you want to match the model to the task. In our case, that鈥檚 classification, reasoning, and summarization. You also need a balance of speed versus accuracy. Sometimes a slighter, faster model is good enough, and cost matters too. You will be calling this agent a lot, so be aware of token costs, and always check that your full ICP and lead data fits within the model鈥檚 token limit. And the great thing is, again, once your agent is built, it鈥檚 easy to swap it to a newer model. Better models as they鈥檙e released without having to rebuild your entire structure.

When setting up your ICP, we want to give the AI agent a very clear profile of the leads that matter to us. For our real estate company example, we would prioritize industries with reoccurring real estate needs. Financial services, tech, life science. We also might focus on companies that are big enough to act. So 50 plus employees, five million in revenue or more. And we鈥檒l also look at growth stage and geographic presence. For the contact fit, we target decision makers. So titles like VP of facilities, asset manager, or key execs in the real estate financial legal area. We most likely are looking for director level above as well. Here, we map the data to specific fields so that we are on the same page as our AI agent about what data we are evaluating. We include behavior signals so the agent can evaluate intent, not just demographics. That鈥檚 high intent form fills, repeated property search activity, and Marketo interesting moments like downloading a report or engaging with a key sales email. And finally, we define lead categories that match our sales playbook. Things like potential buyer, property owner, or potential seller. And for maybe job seekers, we鈥檇 have like HR leads. And of course, we give the agent clear rules for when the lead should be disqualified. If possible, giving the AI examples of records in each stage or each action will help it make better decisions. And you can even build in a real data feedback loop. So you give the agent a list of real examples over time. With this structure, we enable the AI to evaluate leads with nuance and common sense and take consistent explainable actions at scale.

So once we build our ideal profile, our ICP, how do we actually give that knowledge to the agent? Well, the good news is with modern tools, this part is actually easier than ever. Most of today鈥檚 AI agent platforms like Microsoft Copilot Studio or Azure AI Foundry or OpenAI Assistance now allow you to simply upload a document, paste structural content or point to a knowledge source URL. And the AI can reference this in real time. You don鈥檛 need to code it in or build some complex model. You just provide the ICP document. This is a huge shift in the way we work. We鈥檙e not hard coding logic anymore. We鈥檙e giving the agents business context and letting them apply that context intelligently across every lead. Here you can see how simple it can be in Microsoft Copilot Studio, for example, but it鈥檚 similar in all platforms. You upload the ICP, give it a short name like lead qualification guidelines, and now it鈥檚 available anytime the agent needs to reference it. And unlike a human that may read it once or twice, get the gist and then act, our agent will reference this guide word for word for each record it processes. The agent will read from it, reason with it and apply it consistently. Next, to make our AI agent truly effective, we need to give it tools. These are ways that it can interact with data and take action. At a basic level, tools like web search and calculators help the AI reason and fill in gaps. For lead qualification, a key tool is lead enrichment, allowing the AI to pull additional details from LinkedIn or APIs when data isn鈥檛 complete so they can get a full picture. And within Marketo, this gets even more powerful. We can give the AI access to Marketo鈥檚 API to update lead records, and most interestingly, to trigger smart campaigns. This lets the AI choose from a list of smart campaigns that we鈥檝e defined, sending leads to sales, adding them to nurtures or firing off alerts. And AI can even pass in program tokens to the AI agent to give them a summary or to explain its decisions. This allows us to create flexible, expandable tool sets for our AI agent directly inside Marketo. Let鈥檚 take a look at how that works. So inside Marketo, we can create a series of smart campaigns for each action that we want our AI to be able to take. Each smart campaign would have a single trigger and that鈥檚 campaign is requested. And then we would select web service API for the source. This allows the smart campaign to be triggered by our AI agent. From here, we can set up our flow steps however we want. So for example, for an AI qualifying lead flow, we would change the revenue stage, maybe push it to the CRM and change the owner to sales. And here we can send an alert to sales as well. When we add these smart campaigns as tools for our agent, we can explain in just plain text what they鈥檙e for and when they should be used. This gives us additional control because the future, in the future, if we change our minds on the specific actions that should be taken, we can just update them directly in Marketo and we can also specify which program tokens, we can also specify which program tokens should be passed along with when triggering them. And this is super powerful because generally program tokens are set at the program level and they鈥檙e static. However, when triggering a smart campaign from the API, as our agent will, these tokens become dynamic, meaning our agent can pass in a different value for each lead that it鈥檚 triggering in the smart campaign. So for example, if the agent decides to qualify a lead, it can pass in an AI summary explaining its logic and why the lead is qualified. We can use this token to change a data value, like description, or even to populate the body of our alert email to sales. We can also have it pass a lead score, a suggested stage, or any other information we want to share. With this, we are not just allowing the AI to take action, but also giving it a voice.

Finally, once we鈥檝e built our ICP and given the AI tools, the next step is giving the AI instructions. And this is where a lot of the magic happens. Think of this like training a new team member. You can鈥檛 just say, go qualify leads. You need to be clear about how you want them to think through the process. The good news is today鈥檚 AI agents can follow fairly complex instructions, but you do need to be specific and objective in what you tell them. This diagram shows a simple example of what this looks like. We tell the AI first, perform lead enrichment. If the data is missing, go get it. Then calculate the demographic score. Does this person match the roles that we care about? And next, review activity. Pull in the behavior history and interesting moments. And based on that, calculate a behavior score. Is this person demonstrating real intent? Then analyze the ICP fit. Does this lead match our ideal customer profile on the full picture, based on the full picture? Finally, based on all of that, take an action. And again, we give the AI specific menu, a specific menu of allowed actions like triggering a campaign or creating an alert. The key here is that while this reads like a flow chart, the AI isn鈥檛 just executing rigid steps. It鈥檚 using reasoning at each of these stages. But the more clear and objective your instructions are, the more consistent and reliable its decisions will be. In short, write instructions for the AI. The way you鈥檇 coach a really smart, fast learning intern. Be clear about the steps, what matters, what signals to look for, and how to decide what to do next.

Now with our AI agent builds, we have several ways to use it from Marketo. We can pass the leads to the agent with a web hook or self-service flow step. We could also have the agent regularly do a query and get leads to process. But personally, my favorite way is to create a static list in Marketo as a watch list. We can add leads to this list that we want evaluated, and then we can have the agent check this list on a regular basis and evaluate all of the leads in it. So with that in place, we can see here the big picture from leads being pulled in by the agent, the tools and knowledge and instructions connected to the agent, and finally the ultimate action the agent will take as a requestable smart campaign in Marketo.

One of the best ways to roll out agents, especially for something as important as lead qualification is to start with what鈥檚 called a person in the loop workflow. This means the AI still does the heavy lifting, enrichment, evaluation, scoring and decision making, but it pauses before taking actions and asks for human approval. We can start out where all the records go through this approval process and then maybe loosen up the rules. So for example, we might set a rule like if a lead is less than 10 days old, always send an approval request before routing it to sales. These approvals can go through Teams, Microsoft Teams, Slack, or even email, wherever your team already lives and works. You review the AI鈥檚 recommendations, see its reasoning and click approve or reject. It鈥檚 just like when you鈥檙e training a new intern. In the beginning, you don鈥檛 just give them free reign to do whatever they want. You look at things together, give them feedback, and slowly build trust. With the AI agents, you can follow this same exact model just a little faster. And the best part is the AI never complains. It always learns from every correction. It gives you a chance to revise your guidance and it keeps getting better. So this approach gives us confidence and oversight early on and a smooth path to gradually hand over more autonomy as the agent earns it.

You might be wondering, okay, how do I actually build one of these AI agents? Well, the good news is we have a growing ecosystem of platforms available that make this not only possible, but practical. On this slide, you鈥檒l see a range of popular tools for building agents. They all take a slightly different approach, but most follow the same core structure. Give your agent a brain, the model, give it access to knowledge, define the instructions, and equip it with tools. On the left, we have the no-code, low-code platforms like Microsoft Copilot Studio, Microsoft AI Foundry, Amazon鈥檚 Bedrock, Google Vertex, and N8N. These are great if you鈥檙e looking to move fast and don鈥檛 want to write any code.

And on the right, for teams that want full code control or in customization, you鈥檝e got open frameworks like LaneGraph, which give you complete flexibility to define your own agentic logic, memory, and tool orchestration. And this is just a sample. There are many more tools out there and new ones launching seemingly every week. The key takeaway here is that the building blocks are consistent and the technology to do this is already available.

So I鈥檝e built a guide for building AI agents that goes beyond today鈥檚 presentation and really gets into the details. I鈥檝e shared it in the community and you can access it with this QR code. In it, I walk you through the steps of creating an AI agent and using it in Marketo, as well as some additional use cases.

One of the most exciting things about building an AI agent like this is that the model can go far beyond lead qualification. Once you have an AI agent that can reason about data, apply your business logic, and trigger request campaigns in Marketo, you can apply this same pattern to a whole range of use cases. For example, you can have an agent perform data health reviews, which is scanning your Marketo database for anomalies or missing fields, and even trigger campaigns to correct and enrich records. You can build prospecting agents that identify missing contacts for target accounts, doing outreach and flagging the best accounts to prioritize.

Agents can also optimize your nurture streams, adjusting cadence or content based on how the leads are engaging. And they can also do campaign QA, reviewing campaign setups or monitoring for unexpected drops in engagement. The key here is that we have a flexible, auditable way to use AI agents while still fitting cleanly into the way Marketo works. And once you鈥檝e built this pattern once, AIQL is just the beginning.

Here鈥檚 some tips for rolling out your AI qualification process successfully. First, start with a well-defined ICP and clear lead categories. The AI is only as good as the instructions and guidance that we give it. Begin with an approval flow, a human in the loop step so that you can build trust with your sales team. They鈥檒l appreciate seeing why the AI is recommending certain leads before they start receiving them directly. Initially, give the AI a limited set of tools. For example, allow it to trigger request campaigns, but hold off on letting it update lead records until you see it in action. It鈥檚 also important to log the AI鈥檚 decision and surface them to sales. This transparency helps sales understand and trust the process.

And finally, roll out in stages.

Start with simple, low risk actions, gradually give the AI more autonomy as confidence grows. And also communicate and involve sales early and often, bring them into the process, show them what the AI is doing and encourage feedback. The more sales feels involved, the more successful your program will be. All right, let鈥檚 bring it all together with a few key takeaways. First, AI agents enable holistic, intelligent lead qualification. We鈥檙e no longer stuck with rigid scoring models or disconnected workflows. With agents, we can evaluate each lead based on the full picture. That鈥檚 demographics, from a graphics and behavior, and make a common sense decision at scale. Second, the agent model scales beautifully. It鈥檚 not just fast, it鈥檚 transparent, it鈥檚 explainable. Sales no longer has to wonder why a lead showed up. They get the context, the reason, and a smarter pipeline.

Third, when you combine agents with Marketo鈥檚 request campaigns, you unlock flexible modular workflows. The agent can trigger any program, pass in tokens, and fit right into your existing Marketo architecture. Fourth, don鈥檛 forget you can build person-in-the-loop workflows, especially during rollout. This gives you testing, control, and confidence before handing the reins fully over to the agent. This can also be useful when expanding the agent鈥檚 toolset. Fifth, this same model works far beyond lead qualification. Agents can handle data cleanup, prospecting, campaign quality assurance, anything that requires business logic and action. And finally, start small and evolve. You don鈥檛 need to automate everything on day one. Start with one use case, give your agent clear instructions, set up checks and verifications, and just build from there. And with that, I鈥檓 happy to answer any questions you might have.

Big round of virtual applause in the chat for Josh. What a jam-packed session that was. You鈥檙e probably dying to ask a question or two. I鈥檓 sure we鈥檝e already got the chat lit up. So let鈥檚 open the floor to our audience. If you haven鈥檛 already, please post your questions now in the chat, and Josh and I will do our best to answer some of these, mostly Josh, to be brief.

All right, so first question. How do we integrate this AI agent into a lead lifecycle? Is there like a Marketo engage already ready for this, or what鈥檚 your kind of thoughts on that? So the best way is with a smart campaign. You can either call it through a, or there鈥檚 two ways, with a smart campaign, through a webhook, through a self-service flow step, or as I showed here in the guide, set up a watch list, just a static list in Marketo in your program, and then have the agent, through a logic app, check that list periodically, every 10 minutes or whatever schedule you want, and pull those leads in and then process them.

I will also shout out, your previous session at the 51黑料不打烊 Steel Exchange was all about self-service flow steps, so if you have any questions on that, check out the recording from Summit 2025, where Josh goes on that topic. Yeah, yeah. All righty, so we had a question here. They鈥檙e starting to implement some lead scoring with a Forrester analyst, but they鈥檙e, you know, got three different sales divisions.

They wanna try to figure out how to go to market and best way to get started. Get any tips on if they鈥檙e a multi, you know, department company, how do they implement an AI agent inside of their instance? So you kind of have two strategies to go with here. If you send that information, if you know which sales division it鈥檚 going to, which product line it goes to, you can send that along to the agent, and then make the agent aware of those product lines, so it gives a recommendation based on which product line. If upfront you don鈥檛 have that information, you can build that process into it, so ask the AI agent based on their activity, based on, you know, what activity data they have, what information we have, which product line fits them most. And then if you鈥檙e absolutely sure, you can actually divide it if you have different workspaces, have different watch lists in each workspace, and even have different agents for each product line. Gotcha, for sure, for sure. What about the costs? What are the costs? Do you have any costs like on increased API calls or increased features from the use case that you have on the 51黑料不打烊 side, or what are the costs for like the open AI registrations or any costs that way? Sure, so whatever platform you choose to use, in this I鈥檓 using Azure, there鈥檚 going to be some costs associated with the actual model, the usage of the model, data transfer, and then on the Marketo side, of course, the API call usage. But you can control all of those, keep them, you know, scale them and make sure that they鈥檙e limited to whatever you want to use. Those costs tend to not be that much though. These costs are fairly low, but you can set those limits and monitor them.

Sweet, gotcha. What architectural considerations should be made when designing a lifecycle engine for a multi-product platform? So as I mentioned before, depending on if you know that information ahead of time, you would want to separate them and then route them to specific agents that are specialized in that platform, or which product it relates to. If you don鈥檛 know that information, you can build it into the agent away, asking it to figure out which product is the best fit and then route it based on that.

Gotcha. How do you train the AI agent so it鈥檚 based on your known information about your company or about your use case or your product? Yeah, the more information the model has, the better it鈥檚 going to perform. That can be done ahead of time. So you can pre-train a model or fine tune a model so that it knows more information about your company. You can load more information into a vector store or your knowledge, right? I said, give the agent knowledge. The more information you put in there, the better it鈥檚 going to perform. And then additional information in the instructions.

Gotcha. So that leads obviously to some PII questions, right? If we鈥檙e going to be training it on your internal things, your maybe secret sauce, if you will, is there a way to limit that where it鈥檚 not going to be out into the ecosphere where that鈥檚 available to the AI agent or to other people? How do you limit the AI agent to not train on your internal private information? Is that possible? Yeah, so most of these services that we would use have in their privacy statements that they鈥檙e not going to train on this data. This is a paid process so they won鈥檛 train.

Above and beyond that, you may want to put some data governance policies in place where you either anonymize certain things or you limit what data can be passed. You can do this at the API permission level. You can do this when you鈥檙e setting up that static list that you鈥檙e pulling from, which fields you鈥檙e pulling from. All of that can be done. So definitely a consideration to make but with some planning, you can certainly do this and stay within PII regulations.

Nice. Good to hear. It鈥檚 good to hear. All right, so we talked about this a little bit but we鈥檒l get really specific. Since 51黑料不打烊 doesn鈥檛 have an AI functionality yet, I鈥檓 not telling anyone no releases or anything here, no new information coming out, but what models can you use for this type of use case? OpenAI, what do you think else you鈥檇 recommend? Yeah, any model that鈥檚 out there. If it鈥檚 OpenAI, you can use a fine-tuned version of OpenAI鈥檚 model. So giving it your own specific use cases and information and training it above and beyond with your information.

There鈥檚 Metas, Llama models, Amazon has SageMaker. Any model that鈥檚 out there, as long as you can build an agent around it, then you can use it. So it鈥檚 important to find one that works best for you and also has that balance of cost and performance that works best for you.

Awesome, awesome, awesome. We got a good question here about my favorite topic, travel. I don鈥檛 know if everyone else likes to travel. I know Josh does, so this is a good question for both of us. How can we ensure that lifecycle engines can adapt to different customer journeys? For example, like a flight versus a hotel versus a travel package, we鈥檙e looking at booking business travel.

Yeah, so that鈥檚 kind of combines several of the other questions. What you wanna do is, especially when you have very specific niches like that, you can train specific models on each one. So you can fine tune those models. You can create different agents based on each one. So you give them different knowledge bases that they鈥檙e working with and instructions that they鈥檙e working with. And then route everyone based on which model makes best for that, fits best for them. And you can even put it, layer it, right? So you have one agent that figures out which other agent to use. So you have multiple agent layers.

You鈥檙e having the robots talk to each other. Oh yeah.

All right, soon enough, iRobot is here. Soon enough, soon enough. All right, similarly, how can you implement like a scoring model based on these different use cases or different workspaces? How do you kind of set that up in conjunction with these AI agents working together? Sure, so the scoring model, you set it up just how you would now, right? You should work with your sales team, gather all the data, the inputs that you have, and figure out the scoring model. And then you explain it to the model or the agents, just like you would a human who鈥檚 reviewing it, right? And then I always recommend build it first, see how it performs, and then review it regularly with your sales team. When you sit down with the sales team, go in with your ICP and your model, and then make those corrections as needed right there. And then you just re-upload it. The agent gets the new information and starts working with that again. So you鈥檙e continuously improving that modeling.

Nice, all right. What about integrating with third-party apps, like for intent data, let鈥檚 say like a 6th Sense or a Zoom Info, is there anything like that? Yeah, so in this, I鈥檓 using enrichment from LinkedIn, but you can enrich from anywhere. You can bring in data from anywhere. So it鈥檚 just a matter of setting that up as a tool that is available to the AI agent, telling the AI agent when you want it to use that tool, what data it鈥檚 relevant for, and then let it go. Because then if it sees a gap or it has an instruction to use that tool, it鈥檚 gonna reach out and use it. And I think in most cases, the more data you give it, the better.

This is the next question, leads right into that as well. Giving it more data, I shouldn鈥檛 call it it, could have feelings. Give the agent some more data. But what about it reviewing like close one or close lost information from your CRM to determine ideal customer profiles and then kind of going backwards from there. So like starting at the end and then working back. Have you ever done anything like that with any of your clients? That鈥檚 a fantastic question. Yes, so instead of you sitting down and creating your ICP based on your gut feelings or your knowledge, which I鈥檓 sure are great, pulling in actual information and past one deals, all of that and lost, and then training a model. That鈥檚 what I would use that information for. If you have a lot of information about close one and close lost deals from the past, then feeding that into like an open AI and fine tuning that so it really knows it, and then asking it to generate that ICP. You both have the trained model out of that, that鈥檚 specifically trained on your data. And in addition to that, you have the ICP that鈥檚 been generated by the AI. So you鈥檒l probably get a huge boost in performance during that.

It鈥檚 a great call, it鈥檚 a great call. All right, how many AI agents can be in each workspace or partition? Have you limited that or you let them run wild or you keep it to one? There鈥檚 no technical limit.

It doesn鈥檛 live in 51黑料不打烊. So it鈥檚 gonna be, at least these agents are gonna be outside so you can build as many as you want.

I think that鈥檚, it鈥檚 kind of like how many interns could you hire? You can hire as many as you want. How many can you control and actually get to be productive? That鈥檚 the limit.

Awesome, awesome, awesome. All right, I think we鈥檙e getting to the meat and potatoes now.

Have you built these yourself? Or did you hire somebody? How technical was it to develop these? Did you hire a developer to build with you through this process? And then of course, how long did it take? And then we already talked about cost, we don鈥檛 need to talk about that. But yes, so who built them? How technical was it? Did you hire somebody to do it with you? And then how long did it take? Sure.

I built all of these myself. You don鈥檛 need a developer. I am a developer by background, but there鈥檚 no code involved in these. And that was on purpose, kind of went into it with that goal.

From start to finish, I mean, it鈥檚 months and months of building and trial and error and repetitive incremental improvements.

But I built a guide which I shared, and you can go through. You can do this in an hour and build it. Will it be perfect? Probably not. You probably need to test it, run it through, set up that person in the loop flow, or even just safety loops where it doesn鈥檛 actually get sent to sales, but gets sent to you to review to improve that. Yes, you can do it very quickly without any developer. You can build it yourself, follow that guide.

You will have a working model in about an hour or so, but then improving that and getting it to where you want depends on what level of output you want and how much time you want to put into it.

Awesome, awesome, awesome. And maybe I have one or two more.

How can you easily plug the agent into activities tracked by Marketo or other integrated data systems? So in this, we鈥檙e using a tool. We use the Marketo REST API to pull an activity log. You could point that to any data source that you want or additional data sources. So have multiple tools for multiple data sources.

And then, sky鈥檚 the limit. So depends on how technical you want. You can be pulling all that data and if you have a central repository or a data lake where you keep all that, you can plug it into that and pull as much data as you want.

The important part is that it鈥檚 relevant data that is actionable from the AI that can get value out of it and then making the connection.

All right, final question.

How did you actually plug these AI agents into Marketo? Yeah, so you can either call them from a smart webhook or self-service source tab or create that watch list and have the agent pull from that watch list regularly.

Got it, I got it. All right, good, I lied. We have time for one more. Okay, can you use the AI agent to look at activity logs or things in smart lists and be able to filter on those, maybe look at sync errors? Can it actually look at your Marketo use case or is it more focused on the data that you have in there from customers or partners? An AI agent can look at anything you want it to as long as there鈥檚 an API or some connection that you can connect to where it can pull that data in, it can look at anything.

Obviously, this is specifically looking at how to use it to qualify leads, but you can give it any job you want and as long as you give it the tools it needs to do the job and the knowledge it needs to do the job, there鈥檚 no limit to what these agents can do.

Awesome. All right, Josh, thank you so much for your time. Appreciate you being here, answering all these questions from the community. We鈥檒l see you soon.