Meet agentic AI: Your AI agent just leveled up to teammate (Brains Byte Back Podcast)

You’ve probably been coming across the term “agentic AI” a lot more recently, and in this episode of Brains Byte Back, we explain why that is and what it is. Ruben Phukan, co-founder of Goodgist, returns to share how this upgraded form of AI is changing how businesses work. Unlike chatbots or rule-based automation, agentic AI can handle tasks by making decisions, planning next steps, and adapting when needed, more like a human. It’s built on top of generative AI, but it goes further. It doesn’t just respond to a prompt. It carries out full workflows. Mind-blowing, right?

Ruben explains the difference through a practical use case: imagine receiving an invoice by email. Instead of needing someone to manually extract details or program rules for each format, an AI agent can classify the document, extract vendor and billing details, cross-check it with purchase orders, and forward it for approval if everything matches. All of this happens without a human having to define every rule.

This becomes especially useful when you realize most business data—over 80%—is unstructured. Emails, PDFs, reports, and slide decks are hard for traditional automation tools to handle. But agentic AI is evolving to assist companies in those overlooked areas. The double-edged sword is that instead of hiring more staff, companies can now use AI agents to handle the manual data entry and processing, saving time and money.

But Ruben stresses that enterprise adoption isn’t just about plugging in an AI tool. It has to be trustworthy. That means clear audit trails, predictable behavior, and proper access controls. Especially when AI is connected to internal systems. His team at Goodgist built their platform to ensure that agentic systems are not only smart but also reliable and safe to use.

Toward the end of the episode, the serial entrepreneur offers advice for builders entering the space. He warns that many startups build quick prototypes but hit a wall when trying to scale. The key is to solve real problems and build for production from the start. Don’t lead with the technology, but start with a real business need and go from there.

If your company handles repetitive tasks tied to unstructured data, this conversation might give you a glimpse of what’s possible and what’s next in the future of work.

You can listen to the full episode below, or on Spotify, Anchor, Apple Podcasts, Breaker,, Google Podcasts, Stitcher, Overcast, Listen Notes, PodBean, and Radio Public.

Find out more about Oded Hahum here.

Reach out to today’s host, Erick Espinosa – erick@sociable.co

Transcript:

Erick Espinosa:
Ruben, thank you so much again for joining me on Brains Byte Back. I think it’s been about a year since the last time you were on the show, which is kind of nostalgic because that was around the time I was just joining The Sociable. A lot has happened since that year. I noticed that there’s been a bit of a shift in the direction that Goodgist has taken.

The last time you were here, we started with your entrepreneurial journey. From what I recall, your journey began while working alongside one of the co-founders of Yahoo, which set you on a path that eventually led you to become something of an expert in AI and machine learning—before it became trendy, arguably. Would you agree?

Ruben Phukan:
Yeah. Yeah, absolutely.

Erick Espinosa:
Awesome. So it’s been a bit of a makeover. Today, you’re joining us to talk about something a little bit different. We’re exploring how agentic AI—hopefully I’m pronouncing that correctly—is reshaping the future of work, which ties into your latest venture. Can you start by explaining what agentic AI is, and how it differs from other forms of automation and AI?

Ruben Phukan:
Absolutely. Agentic AI is finally getting closer to the AI that Hollywood has always sold us on—but in a practical, real-world way. In simple terms, agentic AI is about goal-directed autonomy.

Let me give you an example. Imagine you have a conversation with a customer discussing implementation schedules and milestones. Now you want to take that information and get it into a system like your CRM or support ticketing software. With agentic AI, you can forward the conversation to an AI agent with a simple instruction—something like, “Extract relevant information and update the correct systems.” The agent, given access to your tools and knowledge base, figures out what to do next: it reads the conversation, plans its actions, invokes the necessary tools (CRM, email, ticketing system), and makes updates accordingly. It observes the results of each action and decides what to do next—all without you having to spell out each step.

This is very different from traditional automation approaches, like robotic process automation (RPA).

Erick Espinosa:
Which is what Gen AI uses, right?

Ruben Phukan:
So Gen AI is the foundation that agentic AI is built on top of. RPA, on the other hand, is like manual automation. You have to program every rule—read the email, find the email address, log in to CRM, update the record, etc. Or maybe you record your screen and have the bot mimic that behavior. But if anything changes, the automation breaks. Then you need new rules, which quickly becomes expensive and difficult to maintain.

Classical AI, or machine learning, takes a different approach. It needs a lot of labeled training data—like thousands of emails and their correct mappings to CRM records—so that it can learn patterns. But it’s still constrained. You’d need a different model for every type of task—one for CRM, another for support tickets, another for ERP.

Agentic AI built on top of Gen AI goes beyond that. These new language models understand human language, images, video, and context. When you add the agentic layer, you enable the system to interact more like a human: planning, reasoning, taking actions, and adjusting based on outcomes. It brings us closer to truly intelligent systems—with lower cost, less infrastructure, and much higher ROI.

Erick Espinosa:
Right, because from what I’ve seen, people building Gen AI solutions still work closely with clients to develop very specific prompts. But I read recently that chatbots using Gen AI usually just respond based on single interactions—like when I use ChatGPT, I ask a question, it answers.

What you’re talking about sounds more advanced—like the AI can hold a more sophisticated, ongoing conversation. It can respond in a way that feels more human and handle more complexity, even behind the scenes.

Ruben Phukan:
Exactly. Chatbots are the first generation of Gen AI—they can handle somewhat complex tasks, but only one step at a time. You ask a question, it gives you an answer.

With agentic AI, we’re automating workflows, entire processes with multiple steps. You don’t tell it each action. You give it access to tools and a knowledge base, and it decides the best way to complete the task.

Erick Espinosa:
I was just thinking—how do you see this impacting customer service? Would you say it’s a positive thing? Because right now, if a customer issue is complex, the chatbot usually escalates it to a human agent.

Do you think agentic AI could eventually handle those more complex issues too—making suggestions and troubleshooting more like a human would?

Ruben Phukan:
Yes, absolutely. And I’d go even further. With agentic AI, the system can be connected to your back-end services—payment processors, CRMs, usage analytics—and can proactively detect issues before the customer even notices.

In that case, the human support team gets a heads-up and can reach out with a solution before the customer complains. That alone dramatically improves the customer experience—and things like Net Promoter Score.

For example, say there’s a billing failure. Normally, you’d wait for the customer to reach out. But with agentic AI monitoring the system, it can identify the issue, perform a root cause analysis, and suggest a fix. Then it can either inform a human agent or—if the risk is low—even contact the customer directly and resolve the issue.

That changes the game for customer retention and trust.

Erick Espinosa:
Yeah, arguably, that’s the most important touchpoint in a customer relationship—payment. If there’s a hiccup and you’re proactive about solving it, the customer is far more likely to stick around.

Let’s talk about unstructured data—emails, PDFs, documents. You mentioned before we started recording that most companies overlook this. Why do you think that is? And how are you trying to fill that gap?

Ruben Phukan:
Good question. Traditional AI performs best when the data is structured—tables, rows, columns—because that’s what those models were designed to handle.

So, for example, if I feed structured data about a customer’s behavior into a model, it can predict whether the customer is likely to renew their contract. Similarly, if I feed in financial spreadsheets, it’s easy to write formulas or rules—multiply column A by column B, etc.

But the reality is, more than 80% of business data today is unstructured—emails, reports, PDFs, slide decks, images. That’s a blind spot for classical AI, and even for many companies.

That’s the gap Gen AI fills—and now agentic AI takes it further. At Goodgist, we’re building tools that can understand, reason over, and act on that unstructured data. We think that’s the next frontier—and the biggest opportunity right now.

So, these are all very unstructured. Now, if you want to define automation for that workflow—say, every time I receive an invoice—I want to extract key pieces of information: who is the vendor, what are the line items they’re billing for, what’s the total amount. Then I want to match that against a purchase order in my system and ensure everything checks out—the item, the billing details, the payment date and schedule. If all of that matches, then send it for approval.

Now, trying to do that through classical models or traditional automation is extremely difficult. Visually, humans can look at invoices and understand them intuitively—we know where to find the relevant pieces. But doing that programmatically, by writing rules or training models with hundreds of invoice formats, is incredibly hard. There are just so many variations—no two businesses send invoices in exactly the same way.

That creates a challenge. But with Gen AI and agentic technologies, you can now set a high-level instruction. For instance: “Whenever I receive an email with an attachment, extract the document. First classify whether it’s an invoice or something else. If it’s an invoice, pull out the line items, pricing, and vendor information. Then look up the relevant purchase order in a directory structure, match it, and if it all checks out, send it for approval.”

Generative models today are capable of understanding these instructions, identifying the right information, and performing those tasks. And because they have access to tools—like a file reader or OCR—they can extract text from images, too. The AI agent can then invoke the right tools for the job, almost like a human would. That’s where the real shift happens.

This is why automating unstructured data wasn’t possible before—but now it is. And it unlocks massive efficiency gains. If 80% of your organization’s data is unstructured and you suddenly gain the ability to automate tasks using that data, the potential for improving operations is enormous.

Erick Espinosa:
Is it still arguably affordable? I noticed on the Goodgist website you emphasize a “no-code” platform and accessibility. Why is democratizing automation so important right now?

Ruben Phukan:
That’s a great question. Businesses today are facing significant challenges—resource constraints, geopolitical tensions, inflationary pressures. So operational efficiency has never been more crucial than it is right now.

Companies are trying to squeeze out every bit of margin they can. And being able to automate tasks that aren’t worth spending human resources on is a huge advantage. Take a legal professional, for example. Think about their hourly rate. If they’re spending hours reviewing routine documents—just because someone has to—it’s not a productive use of their time.

Now, if AI can handle the heavy lifting—analyzing a legal contract, pulling out the key points—then all the lawyer has to do is review and make a decision. It saves time, reduces costs, and improves outcomes.

We also work with supply chain and logistics companies. Freight forwarders, for instance, deal with a mountain of documents for every shipment—bills of lading, invoices, customs documents. Right now, these are processed manually because no reliable automation existed before. But is it worth using human labor for that? Absolutely not.

With agentic AI, we can now automate those processes. So instead of hiring massive teams just to process paperwork, businesses can redirect those resources to more meaningful work. It not only boosts efficiency—it also directly impacts the bottom line.

Erick Espinosa:
It sounds like it’s not just about saving time, but also about scaling operations.

Ruben Phukan:
Exactly. One of our supply chain customers used to process about 500 shipments a day—because that’s all their team could handle. They couldn’t scale without hiring more people. But after deploying our AI agents to handle document processing, they’re now processing three times that volume—1,500 shipments per day.

So that’s not just efficiency. That’s increased revenue.

In today’s business environment, where things are moving so quickly and margins are tighter than ever, this kind of efficiency gain is critical. And the good news is that the technology is finally ready to meet those challenges.

Erick Espinosa:
Efficiency really is at the core of AI—and that’s why it’s being adopted across so many industries. It’s eliminating repetitive tasks and freeing up employees to focus on high-value work.

On the tech side of things, I noticed Goodgist’s platform uses patent-pending technology. Can you talk about the innovations behind that and what you see for the future?

Ruben Phukan:
Sure. So when you talk about workflow automation—especially at the enterprise level—workflows have to be deterministic. That means the system must behave in a predictable, consistent way every single time. You can’t have an AI that does something one way today and then a completely different way tomorrow.

But here’s the challenge: underlying agentic AI models are still probabilistic systems. They operate with a degree of randomness or variability. So how do you make a probabilistic system act deterministically for enterprise-grade tasks?

That’s essentially where our innovation lies. We’re focused on making AI reliable—ensuring it performs workflow tasks deterministically without falling back to hard-coded rules like RPA. We want the intelligence and flexibility of Gen AI, but with the consistency and reliability that enterprises need.

Another big issue in enterprise deployments is trust. How do you know the AI is doing the right thing? That’s why auditability is crucial. You need a complete audit trail—every decision, every step, fully logged and transparent.

Ruben Phukan:
These are all very unstructured workflows. Now, to define automation for those workflows, you’d say something like: whenever I receive an invoice, I want to extract key information — who is the vendor sending the invoice? What are the line items they’re billing for? What’s the amount? Then, I want to match it against a purchase order in my system and ensure that the billing details and payment schedule are correct. If everything checks out, then send it for approval.

Doing this through traditional automation models is extremely hard. As humans, when we look at invoices, we intuitively understand where everything is. But programmatically, writing rules or training models to handle the wide variety of invoice formats is very difficult — no two businesses send invoices the same way.

This is where GenAI and agentic technologies change the game. Now, we can say: whenever I receive an email, extract the document, classify it as an invoice or something else. If it’s an invoice, get the line items, the price, and match that against purchase orders stored in a given directory. If there’s a match, send it for approval. That’s how we can now express the task.

Generative models can understand the document, extract the right information, and perform matching — because they have access to tools. You can equip the agent with a file reader or OCR tool to process images and extract text. The agent can invoke those tools as needed, almost like a human would. That’s the big shift — unstructured data, which was previously inaccessible to automation, is now addressable. Since 80% of enterprise data is unstructured, this unlocks huge efficiency gains.

Erick Espinosa:
Is it still arguably affordable? Because I noticed on your website, you emphasize no-code and accessibility. Why is democratizing automation so critical right now?

Rube Phukan:
It’s more critical than ever. Businesses today are facing massive challenges — resource shortages, geopolitical instability, economic pressures. Becoming more operationally efficient has never been more important.

Think of a legal professional, for example. Their hourly rate is high, but if they’re spending time just reading routine documents, that’s a waste. With AI, you can automate that first layer — extracting key parameters from a legal contract — and then let the lawyer focus only on reviewing and making decisions. Same goes for logistics and supply chain companies. Freight forwarders, for example, deal with a huge number of documents per shipment: bills of lading, invoices, customs papers. These have traditionally been processed manually, which is time-consuming and inefficient.

Now, AI can take over that tedious work. Companies no longer need to hire large teams just to input data into transport management systems. They can reassign those resources to more strategic tasks.

One of our supply chain customers used to process 500 shipments per day — the limit of their manual team. With AI agents, they’re now processing 3x that volume, which directly increases their revenue.

So yes, it’s about efficiency, but it’s also about growth. In a business environment where margins are shrinking and speed is critical, the ability to scale with technology is a game-changer.

Erick Espinosa:
I noticed Google’s platform is built with patent-pending technology. Can you share more about the innovation at its core and the future of this tech?

Rube Phukan:
Sure. Workflow automation, especially in enterprise environments, has to be deterministic. You can’t afford for the AI to behave differently each time. But the underlying foundation models in AI are probabilistic. So the challenge — and the core of our innovation — is how to make a probabilistic system behave in a deterministic way.

That’s what we’ve focused on: making AI reliable and predictable in performing workflow tasks, without hardcoding rules, and without reverting to traditional RPA methods. That balance between flexibility and control is essential.

Another critical area is trust. Enterprises need to trust the system. That means full audit trails — every decision the AI makes must be reviewable. It can’t be a black box. You also need role-based access control. In agentic AI, tools can be powerful — so whoever has access to the AI has access to those tools. That’s not acceptable in an enterprise context, so we’ve built strong access control systems around it.

Erick Espinosa:
Is that the kind of advice you’d give to others entering this space? A lot of startups are jumping into AI now — what pitfalls have you seen?

Ruben Phukan:
Yes, absolutely. Over the last year, there’s been a lot of hype around GenAI and agentic AI. APIs became accessible, people could experiment easily, and a lot of quick proof-of-concepts were built. But the real challenge is going from proof-of-concept to production.

That’s where most efforts fail — because real enterprise use requires things like cost control, access management, guardrails, audit trails, and robust reliability. That’s when people realize the true complexity and cost.

We’re seeing a shift now. A year ago, many companies thought, “This is easy, we’ll build it in-house.” They’d get a developer to build something over a weekend. But productionizing it? That’s where they hit a wall.

Now, CIOs and CTOs are more strategic. They realize that deploying enterprise-grade AI requires partnering with companies who do this full time — who understand the nuances and infrastructure needed. We’ve been working on this since early 2023, and even now our systems are evolving daily. There’s just so much depth to building an enterprise-grade system.

Erick Espinosa:
So what’s your advice to startups and builders?

Ruben Phukan:
There are no shortcuts. Whether you’re a startup or a large company building in-house, that’s fine — but you need to understand the full picture. It’s not just about prototyping. You need to plan for production, understand the technology deeply, and treat AI like any complex software system.

Also, don’t start with the technology and look for a problem. Start with a real business problem, then work backward to the technology that can solve it. Too many try to create a general-purpose system that can do everything. That’s not where the value is. Focus on solving real, targeted problems — and build out from there.

Erick Espinosa:
That’s the response of a true engineer — solving problems from the ground up. And it’s true: if your business can’t keep up, while others move forward with AI, you risk falling behind. Great insights, Rube. I really appreciate you joining us again. I learned a lot about agentic AI — this was very new to me, and I’m sure we’ll be hearing much more about it soon.

If people want to reach out and pick your brain, what’s the best way?

Ruben Phukan:
They can reach me on LinkedIn — just search my name. Or on Twitter. Or send me an email at ruben.fukunatzogis.com.

Erick Espinosa:
Perfect. I’ll include all that info at the bottom. Thanks, Rube. Hope to chat again soon!

Ruben Phukan:
Thank you so much, Erick. Thanks for having me.