Agentic AI could fundamentally change how you interact with ad platforms and perform your job. Whether you work in ad ops, account management, sales, or strategy, if your job requires you to log in to an ad platform to perform a task, then you should understand what agentic AI means for you.

We adopt technology when it enables something new or helps us accomplish something more efficiently. This has been the case from the printing press to the iPhone. After speaking with product leaders and builders across the industry, I can tell you that agentic AI will enable new use cases and increase productivity in ad tech.

Agentic AI, coupled with generative AI, could boost human productivity to levels previously unimaginable. This technology can multiply the output of a single worker many times, which is a daunting prospect for the future job market but also a wondrous opportunity for those who learn to harness its power.

I wrote this article to learn more about agentic AI, understand what's behind the buzzword, and share with you what I learned. I interviewed founders and product leaders who are riding the agentic AI wave. These are the people who must figure out how agentic workflows enhance their existing platforms or how to build entirely new companies based on agentic products.

But before we go any further, we have to learn what agentic AI means.

What is agentic AI?

Agentic AI is the ability of a system to take independent actions toward goals, operating proactively rather than just reactively and adapting its behavior as needed to achieve them. An AI agent is a tool or system that interacts with users, software, or data to accomplish specific goals autonomously.

Agentic AI is like a person's ability to plan and act independently. An AI agent is the person applying those skills to get work done and achieve specific results.

It also helps to compare this to generative AI. Generative AI is a type of AI that creates new content—like text, images, code, or audio—based on patterns it learned from training data. Generative AI is a capability that agentic AI may use to produce outputs as part of achieving goals. An agent might use generative AI to write a personalized email or summarize a meeting.

The real magic is when all these AI concepts come together. Generative AI text output ushered in the ability to interface with an application via chat, which allows a user to set up agentic use cases via the same natural language input mechanism. It also allows an AI to produce results from those interactions in the form of text, images, or video. This combination of agentic AI plus generative AI is what will usher in a profound shift in the way we interact with technology. 

AI agents can reason and figure out the best way to accomplish tasks for a user, but to do so, they need access to ad platforms and systems and an understanding of their available capabilities. That’s why we now have to touch on MCP. This term came up frequently in all my conversations. 

What is MCP?

Anthropic, the company behind the AI assistant Claude, introduced MCP, or model context protocol, last year in an attempt to standardize how AI assistants connect to other systems. The idea is that a system could host an MCP server that allows AI assistants to understand what tools are available and how they work, along with providing a natural language description of each tool. 

Think of it like API documentation for humans today. APIs allow external systems to push or pull information to or from a platform. But to integrate your system into a platform, you have to understand what endpoints to call, what information to provide, and how to authenticate. This information would all be outlined in human-readable API documentation and implemented by engineers.

A chatbot or AI assistant can query an MCP server and ask for everything it can do, alongside natural language descriptions of what the tool does and the required information to accomplish that task.

Without an MCP or a similar standard, any agentic tool would have to build out a bespoke integration for every app it interacts with using legacy APIs. With MCP, this process could be mostly automated and enable any platform to support agentic workflows, allowing agentic solutions to scale rapidly.

Integrating agentic AI

When I spoke to David Dworin, FreeWheel's Chief Product Officer, he told me he is thinking about agentic AI from two frames of reference:

1. How can FreeWheel better accommodate third-party agents that need to access FreeWheel?

He is thinking about how to enhance FreeWheel APIs or data products to make it easy for agents to interact with the platform. He mentioned that integration with third-party agents could even leverage a FreeWheel MCP server, which his team is considering.

2. How can FreeWheel enhance its system with agentic workflows?

David touted agentic as an evolution of typical workflow automation (if this then that). So he and his team are looking at potential use cases where he could supercharge automated workflows with agentic AI:

He mentioned a potential troubleshooting use case where you can run a diagnostic, and then FreeWheel tools could recommend the action to take or even take that action on your behalf.

I loved this two-part framing, as it establishes a baseline framework to explore agentic solutions in ad tech — not an easy task with novel technology. It also sets us up to explore the different approaches we may see play out regarding model and agent ownership. Let's use David's framing to explore deeper. 

Evolving platforms 

Humans interacting with a user interface has served as the lone method of making an ad platform do something. Agentic AI could shake up this paradigm for the first time in the history of ad tech. Joe Hirsch, CEO of Swivel, who was previously CEO of video ad server SpringServe (acquired by Magnite), put it more directly when we spoke:

Joe is especially invested in this future as his company is looking to power the agents that could serve as members of your ad ops team. Swivel "is on a mission using AI and machine learning to automate media's most complex ad delivery strategies, freeing humans to achieve their loftiest business goals."

This is a lofty goal in and of itself, but Joe thinks Swivel can empower users with the help of agentic AI and unlock possibilities that humans aren't capable of:

AI agents don't sleep; they could have instantaneous access to historical and real-time data, and you can train them on your entire history of changes to ad serving, deal setup, operational policy, and goals. 

An ad ops agent either baked into an ad platform or through a third party like Swivel could make thousands of changes throughout the day based on all these inputs — a truly impossible task for a human. 

But in the meantime, man and machine still must work together — and other platform operators like PubMatic are already launching and testing out new ideas of what this could look like. 

Ankur Srivastava, Vice President of Product Management at PubMatic, told me that the company has released PubMatic Assistant, an LLM-powered tool that is currently in closed beta.

PubMatic has integrated its assistant into reporting tools to extract complex insights quickly, and is now focusing on integrating agentic AI into the deal management lifecycle (demand insights, deal creation, troubleshooting, and optimization). PubMatic Assistant, for example, can now help with programmatic deal troubleshooting.

Ankur observed that troubleshooting skills among ad operations teams can vary. PubMatic Assistant sets a consistent baseline of competency for all customers out of the box.

As someone who has spent 14 years troubleshooting programmatic issues, I've found that there is truly an art to the practice. Whenever I attack a new troubleshooting issue, I have 14 years of experience and pattern recognition to fall back on. But I also don't have millions of programmatic transactions plugged into my brain. When will these tools outperform me in troubleshooting? I fear very soon. 

One advantage I have over AI currently is that humans are still the primary users trafficking campaigns. An AI may be able to tell you that a DSP has a 0% response rate, but if it's not responding, it has no data to work with to troubleshoot the issue. This situation is where human-to-human communication comes into play, which may determine that a human error (like setting incorrect flight dates) is to blame. 

As long as humans configure the operational setup of supply and demand, human workers will remain a vital resource. If agents start taking over both sides, we may see a gradual decline in human necessity and a rapid increase in the need for agent-to-agent communication.

Agentic communication

The topic of agentic communication and interoperability came up in every conversation I had — and for good reasons. 

First, this topic is bound to be top of mind given interoperability among platforms is the heart and soul of programmatic advertising, dating back to the first real-time transaction between platforms — which, curiously enough, I recently learned may have originated from PubMatic. (Thank you for the history lesson, Ari. “Yield“ should be required reading for anybody working in ad tech.) 

Second, agentic communication drastically increases the capabilities of any agentic tool. If agents cannot break free from their systems, then they can only perform actions within a closed ecosystem. 

Outside ad tech, I often hear a recurring example of how I could one day ask my personal AI agent to book me a flight for my upcoming trip. This agent will have access to my calendar to know when my trip is and access to airline booking systems to make the reservation. 

In ad tech, users typically interact with multiple systems, such as DSPs, brand safety vendors, creative ad servers, order management systems, and more. These systems typically already have integrations together, but the usefulness of those integrations could increase significantly through agentic AI.

Let's return to our programmatic troubleshooting example, where a particular programmatic deal booked in a DSP is not responding. An AI agent confined to an SSP may not be able to identify or fix the issue since the problem lies within the DSP. What if our SSP troubleshooting agent could talk to a DSP troubleshooting agent?

Ankur from PubMatic presented me with a potential use case for agentic communication between an SSP agent and a DSP agent.

The SSP agent could inform the DSP agent that the publisher is asking why this deal is not bidding, and the DSP agent could inform both the SSP agent and the advertiser why. Then the humans could take the necessary steps to solve the issue. Further in the future, the agents could have more autonomy to resolve the issue themselves (with guardrails in place).

Integrations among platforms today leverage APIs that allow one system to perform an action on another system. There could be an API action to read or set a campaign budget, deactivate a creative, or update a deal name. Platform operators can define all these actions within an MCP that an external agent can read and understand. It can then interpret a user's input and reason which abilities it needs to accomplish the intent.

Integrating external agents into platforms aligns with another common subject that came up in my interviews: model and agent ownership. While platforms will likely have their own internal AI systems, another school of thought is that publishers, advertisers, or agencies could own or license an internal AI model that will interact with all of their ad systems. 

This model could be pre-trained on all internal brand safety standards, operational processes, and preferences, along with seeding it with baseline ad tech knowledge, including jargon, reporting terms, and understanding of how ad platforms work. The company-level agent would then plug into all the platforms and tools to run an advertising business. It would know exactly what it could do by interacting with the MCP servers of those platforms and tools. 

Joe from Swivel envisions this concept as a possible future component of his company:

He went on to explain how a customer can use natural language to perform ad ops, gather reporting, and do anything from a single entry point.

As an advertiser, imagine you just convinced your client to increase the budget. You could ask your agent:

Find all deals for my advertiser, Pizza Shack, and pause all creatives with a conversion rate by zip code lower than 5%, and shift the budgets allocated for the paused creatives evenly amongst the remaining creatives, and append '— zip code optimized' to the deal name of all updated deals. Then analyze the deactivated creatives and tell me if they offered a promotion and produce a report broken down by state to show me conversions by promotion vs no promotion.

This prompt is an extremely loaded request, but bear with me for the thought experiment. Your agent will have access to the DSP to make creative and deal changes, and internal reporting aggregation tools to pull reporting data. The agent can also access your creative library and use multi-modal visual and text caption scanning to understand if the creative included a coupon code.

Your agent understands exactly which system to interact with to accomplish its goal. But given the complexity of the prompt, it might be a good time to review the concept of "human in the loop."

Humans and guardrails

I don't know about you, but I would not trust an AI at this point to go out and adjust live deals immediately after a prompt. But this is where the concept of "human in the loop" comes into play. After my prompt above, the model could generate a response that outlines what it thinks you want it to do before it performs that action. For example:

Here is what I think you are asking me to do:

1. Find all deals assigned to the key name "Advertiser" = "Pizza_Shack" 

2. Analyze reporting data for "Advertiser" = "Pizza_Shack" broken down by all United States zip codes with "Metric" = "Conversion_Rate"

3. When "Conversion_Rate" < 0.05 for a specific zip code, do not allow that creative to serve that zip code.

4. Create a .csv file with columns = "State", "Promotion or No Promotion" (as a boolean 1 = promotion, 0= no promotion). 

Questions:

1. What date range should I look back to when analyzing the conversion rate for Pizza Shack creatives?

2. You don't want me to pause certain creatives across all zip codes, but you currently don't have any zip code targeting mechanism applied. Would you like me to create a custom rule to exclude the low-performing zip codes from specific zip codes?

3. When analyzing creative video, can you define "promotion"? Does this mean a coupon code is present on screen for any amount of time?

At this point, you would provide further clarifications to your in-house LLM and confirm its approach, with the entire interaction training the model to better understand in the future. After a final confirmation, it would automatically accomplish your intended deal updates, creative scanning, and produce a report for you. Welcome to agentic ad tech.

What if you wanted to take this a step further and enrich your campaigns with first-party or licensed audience data? 

Dstillery, an AI ad targeting company, has already developed for this use case. Dstillery uses artificial intelligence and machine learning to analyze large datasets and create custom audience segments for digital advertising campaigns.

I spoke with their Chief Product Officer, TJ In, and VP of AI Products, Mark Jung, to learn more about how they are incorporating agentic workflows into their products. Dstillery has already stood up an MCP server that defines all the tools available through their platform.

We walked through a demo leveraging these tools where their AI agent (which they referred to as 'DS-1') can query Dstillery's system to build and activate audiences that agencies and advertisers could later use for targeting ad campaigns. In our example, we looked at two separate approaches to identify relevant audiences. First, we looked for people who want to buy bobbleheads, and the tool returned segments of action figure enthusiasts, sports fans, and other groupings of people who the AI thought would buy bobbleheads.

Mark told me this is one of the benefits of AI, where legacy practices would be a human searching a DMP for "bobbleheads" and that DMP only returning segments with bobbleheads in the name or description. But with AI, it can reason the type of people who collect bobbleheads and what kinds of audiences relate to these people. So rather than relying on humans to make these connections and searching for keywords of audiences that could loosely fit your targeting criteria, these agentic workflows can find relevant audiences better than any human could.

In the second example, we tasked our AI agent to look at the first-party data of a large Asian clothing retailer and to return lookalike first-party segments along with relevant third-party segments, and to predict the potential performance lift that these segments would yield if applied to a campaign. 

The tool returned lookalike segments and third-party data segments, including women's fashion brand shoppers, UK boutique women's fashion, luggage shoppers, and non-endemic segments like drama show watchers. The tool then provided the predicted performance lift that each segment would yield for a campaign before an advertiser served a single impression.

TJ pointed out that tasks like identifying the best audience segments, creating custom audiences, and activation could take anywhere from days to weeks, depending on the number of back-and-forth emails and processes involved. When AI agents like DS-1 are available directly in the hands of agencies and media buyers, they can complete these same tasks in minutes with more accuracy.

Job killer or supercharger?

Many discussions about AI and its impact on the economy ultimately drift toward the awkward problem of white-collar job cannibalization — and it's a valid concern in ad tech. If AI agents can traffic campaigns, troubleshoot programmatic deals, produce data analysis, and everything in between, where does that leave us? Will companies need large ad ops teams at publishers, media managers at agencies, or data scientists anywhere?

At this beginning stage, while the AI models are still learning, entry-level jobs are most at risk. Couple that with an uncertain economy due to tariffs, and Gen Z grads are already facing an extremely bleak job market. Employees aged 20-24 occupied 6.5% of all jobs in advertising, public relations, and other related services last year, the lowest since 2020 (down from 10.5% in 2019).

Against the backdrop of this daunting job market, the wealthy person most responsible for the gloomy outlook, Sam Altman, is telling them, "This is probably the most exciting time to be starting one's career, maybe ever."

Despite the tone deaf response, maybe we should listen to the OpenAI chief. Agentic AI opens up some truly limitless possibilities, enabling individuals to accomplish goals that were previously impossible. 

Take the area of data science, for example. Often, data scientists are facing endless queues of requests from different departments. Sometimes companies cannot hire fast enough or keep up with the work required to fulfill the requests.

At various points in my career, I've waited months for a data science request or analysis. But what if anybody had a team of 10 data scientists at their disposal at any point? That's the pitch I heard during my conversation with John Hoctor, CEO and founder of Newton Research.

He told me that he realized that many publishers, agencies, or brands do not have enough data scientists or analysts to keep up with the number of requests they receive every day. This realization led to the genesis of Newton Research, which arms customers with AI agents that supercharge anybody's ability to pull meaning out of media and advertising data. 

His agents come pre-trained on a wide range of analytics, including complex and time-consuming analytics tasks like MMM, incrementality testing, attribution, predictive modeling, campaign insights, and measurement reports. Newton Research enables its customers to "talk to their data" and essentially gives them access to an unlimited number of junior data scientists. So what does this mean in practice? 

Rather than aggregating data, Newton works with the data wherever a customer has it stored, so it doesn't have to leave their preferred system. It allows customers to interact with their data through a large language model and perform agentic actions using that data. 

For example, as an advertiser or agency, you could interface with Newton via chat and ask it to analyze the ROAS of a particular CTV campaign and compare it to past similar campaigns.

To make sure all their customers don't start from scratch, John and the team train Newton on media analytics, advertising metrics, and industry jargon, along with an understanding of integrations — something that could take a junior-level data scientist years to grasp fully. So the model immediately pulls meaning from any request filled with jargon and referencing historical context.

The agentic workflow kicks off once the user provides their goal. The model can reason where it should pull the data from, calculate the ROAS, compare it to past campaigns, and break it down by the highest-performing audience or creative because it knows that's what you like to optimize against. 

From there, the user could choose to leverage those insights into activation. You could ask Newton to create an audience and reallocate budget distribution across creative variations based on performance. The tool could then build a list of user IDs similar to those performing well, redistribute the budget skewed toward the highest-performing creatives, and push all these changes directly to your DSP. 

This example shows how Newton can save media buyers countless hours of work or accomplish tasks they would never be able to achieve alone by leveraging AI agents. Further, it demonstrates how integrating with other ad systems supercharges the capabilities of agentic tools.

There is even development underway between Newton and Dstillery that does precisely this. Mutual customers can use agentic AI to connect Newton's measurement capabilities to Dstillery's audience modeling and activation capabilities to create a closed conversion to activation feedback loop. The idea is that users can find conversion data through Newton, which Distillery could use to generate audience targeting models for activation — a tedious task or outright impossibility for humans on an ongoing basis.

Identifying gaps in the workforce or any area where a human is performing repetitive tasks to free up their time to spend it better elsewhere is something I heard from another founder. 

Adam Epstein is building Gigi, a platform that hooks into the Amazon DSP and serves as an agency's always-on media manager. Users can interact with Gigi via natural language to set up campaigns, optimize performance, or run custom queries for insights. 

Adam gave me a demo of his slick product, and I later heard him on Eric Seufert's Mobile Dev Memo Podcast describe eloquently how he figured out how to build an impactful agentic AI company and product:

Adam identified enterprise media managers as a great place to start, and he pivoted his company to build an agentic AI media manager. Adam went on to explain, "You don't buy software, you hire software."

Companies may soon view AI agents as a supplement to a team rather than a tool that a team uses. Software has the added benefit of never quitting your company and is theoretically infinitely scalable.

Our agentic future

I had always held great envy for those at working age during the dawn of the Internet, a time when you had a vast blue ocean of opportunity literally at your fingertips. At this time, you could leverage an outrageously radical concept, a worldwide interconnected computer network, to dream up infinite possibilities.

I no longer have a valid reason to hold onto that jealousy because we are living in a similar moment. Agentic AI opens the door for founders, product builders, or anybody to imagine use cases that were previously impossible. 

While the threat of human job attrition looms over this groundbreaking set of technologies, those who embrace the coming wave of agentic automation and learn to harness these tools may be living in a time of boundless opportunity that we may never again see in our lifetime.

When I began writing this piece, I had a weak grasp on the concept of agentic AI, and after completing it, I still feel like I have only scratched the surface. But I hope that you, like me, now have a framework to think about agentic AI and how it impacts advertising technology, your company, and yourself.

Now comes the time to figure out collectively how we will integrate agentic AI features and services into our daily workflows. Ad platforms must do the dirty work of figuring out how to properly interoperate with each other — a goal of the Agentic Advertising Collective, organized by none other than Joe Hirsch of Swivel. 

Artificial intelligence is poised to take advertising technology on a ride into a new world of possibilities — but it could also fundamentally change how you succeed in your role.

For every seasoned ad tech professional employing legacy practices, there will be a hungry young worker who will leverage AI to crack into a worrying job market. I recommend you learn how to surf this new wave of AI technology rather than letting the swell sweep you out into a sea of irrelevance.