Eric Picard, Senior Vice President of Product at Fluency, is an experienced product and technology executive who has spent decades developing digital advertising platforms, automated marketplaces, data systems, and consumer-facing products. Before joining Fluency, he held senior leadership roles at Microsoft, Pandora, MediaMath, Yieldmo, and BARK, where his responsibilities ranged from advertising technology strategy and programmatic media to product management, data science, engineering, and large-scale platform modernization. Picard is also a serial entrepreneur who founded Bluestreak, Rare Crowds, Questori, and other technology ventures, with Bluestreak and Rare Crowds ultimately being acquired. His career has included helping architect an early real-time bidding ad exchange, launching programmatic audio advertising products, developing automated inventory curation technology, and advising companies on artificial intelligence, product strategy, corporate transformation, and mergers and acquisitions.
Fluency develops a Digital Advertising Operating System designed to help agencies, brands, and multi-location businesses manage paid media campaigns across search, social, programmatic platforms, walled gardens, and the open web. The platform centralizes data, creative assets, inventory, compliance requirements, audiences, campaign strategy, reporting, and execution workflows, using automation and artificial intelligence to reduce repetitive operational work and coordinate campaigns at scale. Fluency is increasingly focused on agentic advertising, including governed AI-driven workflows that can execute and optimize campaigns while maintaining defined strategic and brand controls. The company says its technology supports more than 250,000 campaigns each month and manages approximately $3 billion in annual advertising spend.
Your career has spanned several foundational periods in digital advertising, from founding Bluestreak and helping establish early digital advertising standards to contributing to the architecture of one of the first real-time bidding exchanges. How have those experiences shaped your approach to building the next generation of agentic advertising systems at Fluency?
I’ve been very fortunate in my career to have found a space that was interesting in a multifaceted way with digital advertising. I got to be part of the mechanism that funded the growth of the Internet. I have been able to invent or be part of the invention of some of the fundamental mechanics of how digital advertising functions. And the technical challenges in this space rival any in software. So it’s always interesting.
With Bluestreak we got to build some of the first of almost every ad format including some of the first video and audio ads. We were the first to do multi-touch attribution, the first to offer a tag management system, and the first to offer multi-touch across all digital media types. This led to being invited to co-author the first impression standard working with the Interactive Advertising Bureau (IAB) and Media Rating Council (MRC).
When I went to Microsoft, the company had decided to invest in advertising technology because they saw that Google was making more money per copy of Windows than Microsoft was, all because of digital advertising. They offered me a role that was incredibly juicy, the opportunity to write the global strategy for how Microsoft should go into the advertising technology space.
We decided to go big in ad-tech, and I was on point to figure out the build/buy/partner strategy. I got to perform due diligence on over 150 companies, we bought seven companies and we invested heavily in growing internal teams. When I arrived, we had about 200 engineers working on ad-tech, and we had about $750MM in media revenue. By the time I left, we had 1,500 engineers working on ad-tech, and over $5B in revenue. Today, search advertising revenue accounts for around $15B, but at the time it was an incredible number.
One of the companies I looked at, and we ended up acquiring, was AdECN. This was Jeff Green’s first company, before he founded The Trade Desk. We had looked at buying Right Media, which is how I got to know Brian O’Kelley pretty well, but we passed on them. We felt like there was an opportunity for an asynchronous global auction that allowed for multiple external bidders to connect to the exchange. So we acquired them, and helped them build a robust engineering team to execute the vision.
When AdECN was ready to launch the first real-time bidding (RTB) exchange, there was a battle over approaches, and it basically never exited Beta testing.
There were two outcomes of this that really changed the entire ad-tech ecosystem. First was that Brian O’Kelley started AppNexus after he left Right Media, and when it was clear that AdECN was never going to get off the ground, I championed the opportunity to partner with AppNexus. This ultimately led to a huge investment in the company, and a strong partnership that finally ended with them being acquired by Microsoft. The second major outcome was that Jeff Green left Microsoft, took a few of the engineers who’d built the RTB exchange at AdECN, and started The Trade Desk.
When you look at that first arc of my career, which took us from 1997 to 2010, it planted some long term seeds that have now sprouted some interesting fruit. I went off to various roles, including my startup Rare Crowds, where we invented Programmatic Curation. I spent time at companies like Pandora, where we had a great audio advertising business, including the biggest audio ad exchange, and Yieldmo, where we built the first privacy-centric ad exchange. During that time I had the opportunity to be part of the board of directors of the IAB Techlab, and was invited to be on the executive committee of the board. And I’ve continued to work on industry standards all these years.
Last year, Brian O’Kelley saw the opportunity to build an industry standard for Agentic Advertising, which is called AdCP, managed by AgenticAdvertising.org. Brian and I talked about where there were holes in the work being done, and I offered to write the first specification for Governance and Compliance for AdCP.
A couple of months after I wrote it, I was contacted by Fluency. When I began to learn what they’d built, I was blown away. Effectively they’d already solved everything I had called out in my specification. They’d been doing agentic advertising for eight years, well before LLMs were ready to go “agentic”.
I saw that the deterministic approach they’d taken with their agents was exactly what was needed to fulfill the needs of LLM-powered agents, which are inherently probabilistic. All that really means is that while LLMs are great at “thinking” and being creative, they’re terrible at doing the same thing, the same way every time. Each approach alone has weaknesses, but together, they solve the problem of agentic advertising.
My vision was compatible with the Fluency leadership team’s and I joined to lead Product here. Now we’re running full speed toward a world where agents can safely transact on advertising with humans in-the-loop where needed, and LLM-powered agents are given deterministic guardrails to manage their behavior.
I feel incredibly fortunate to have ridden so many of the waves of major change in digital advertising, all the way from the first impressions standard, to the invention of RTB, to the invention of Curation, to the advent of Agentic Advertising. It’s been a wild ride, and I’m still having fun and excited about the opportunities in this space.
Real-time bidding gave software the ability to make financial decisions in milliseconds. What lessons from the development of programmatic advertising should today’s developers consider when giving AI agents greater operational autonomy?
In programmatic advertising we used Machine Learning, which is a form of AI, to build decision engines that could make tens of millions of decisions per second. This requires an incredible amount of computing horsepower, data processing, vast amounts of infrastructure, and before the advent of generative AI, I would argue it was one of the most profound developments in all of computing.
For all its power, there are also massive inefficiencies in the programmatic model, and a huge, complex ecosystem that burns far too much electricity (and therefore creates far too much carbon) per ad impression. Don’t get me wrong, I love the programmatic space, and I’m in awe of what we created. But it feels like we’ve built a machine far more complex than the payoff we get for all that complexity.
While the same exact arguments can be made for Generative AI, the places LLM-powered agents need to live in the ecosystem are at the interfaces, not at the heart of every single transaction. And by judiciously mixing LLMs with deterministic agents, we can much more efficiently manage transactions while vastly reducing complexity.
There are two major initiatives in the industry, aiming to inject LLM-powered agents into the workflow of how transactions take place. The IAB TechLab has spun up the Agentic Advertising Management Protocols (AAMP) initiative, which enables mechanisms to proxy LLMs into the RTB infrastructure by making decisions in advance and caching them in the real-time systems of the ecosystem. The AdCP project is focused on what I think of as “account level” integrations. Meaning that all the work done by humans in the campaign management and creative workflows for advertising can be automated. While AAMP is super valuable and interesting, it’s improving a complex modern system. AdCP is going after a problem that has existed since the beginning of the industry, and one that Fluency has been solving for eight years. So we’re digging in deep on AdCP, but we’ll eventually dive into AAMP as well.
Fluency has effectively solved the last mile problem that was never addressed in the ad-tech space: The massive inefficiencies in creating and managing ad campaigns at the account, campaign and creative level. And I truly believe that the work we’re doing is as profound an opportunity for advertising as what we invented with RTB. As we connect LLMs into the workflows that Fluency has already automated, we are increasing the value of the whole system.
Fluency separates probabilistic AI reasoning from the deterministic systems that execute campaign changes. Why is it currently too risky to allow a large language model to directly control advertising budgets?
I think this is a fundamental kind of problem that LLMs will likely always have. Probabilistic engines are designed to give different answers to the same question every time. Even when they’re incredibly good at thinking, they’re terrible at deterministic things. Let’s use arithmetic as an example:
LLMs literally cannot add, subtract, multiply or divide. Now, you might not realize this, because when you ask Claude or ChatGPT a question that requires arithmetic, they can give you the right answer. But what most people don’t know is that they do this by writing scripts and executing that code to handle calculations. That’s what we’re talking about here: The intersection of a probabilistic LLM that understands it cannot do arithmetic, building its own deterministic piece of software that does calculations in order to solve the task requested.
Even though the LLM can now “magically” do arithmetic, I would never let an LLM manage my finances and make stock trading decisions for me. I would never let an LLM do my taxes and file them without any kind of review. Hell, I wouldn’t even let an LLM write an email or blog post without carefully editing it to ensure everything is true.
As they get smarter and better at executing, they may well be able to do these things at some point in the distant future. But my guess is that they’ll do that by writing their own deterministic software to handle the parts of the problem that require it.
Fluency is there today – we have the deterministic pipes to manage advertising. We are integrating LLMs into that platform, using them for the things they’re great at. And we’re exposing our platform to external agents so that they can execute over our pipes. Ultimately I see this as building the gateway into the ad ecosystem, where an LLM-powered agent can transact on any media platform by integrating in one place. But we’re all a bit far from that long term outcome!
How does Fluency’s deterministic execution layer evaluate an AI agent’s recommendations before allowing them to affect a live campaign, and what kinds of rules or constraints can advertisers establish?
We provide mechanisms for humans (or LLMs) to create deterministic instructions that will execute the same way every time on ad platforms. This is a big problem that needs to be solved in every industry. For advertising the company solving this is Fluency. We enable the creation of business rules and heuristics for any action that one of our customers might take in an ad platform. If it’s possible to create IF/THEN logic for a problem, our platform enables that to be encoded into business rules that are followed the same way every time.
This is critically important to understand. There are lots of decisions humans make the same way every time they see a specific situation. This allows us to do things like encode budget management rules, or generate campaigns or creatives while enforcing naming taxonomies, or automatically spinning up and down ad campaigns and creatives based on triggers in external datasets.
For example, a car dealership might get a new shipment of cars. Fluency can read the inventory management system the dealership uses, and automatically generate campaigns and creatives for the exact make, model, even the color and specific attributes of each specific vehicle, within minutes of the cars being on the lot. And when the last car is sold, automatically shut those campaigns off.
All this is pretty profound, because the business decisions to do all these things are very simple when taken one at a time. It’s the “death by a thousand cuts” problem in managing advertising today that goes away. Our customers find that this frees their teams up from putting out fires all day to doing the work that matters to move the needle for their clients.
Advertising strategy and creative development often require experimentation and subjective judgment. How can organizations preserve the flexibility of generative AI while ensuring that campaign execution remains predictable and compliant?
Everything we’ve talked about already has built the framework to answer this in a simple way. The things that can be encoded into business rules are set up properly to execute those easy decisions that are the same every time. LLMs can help analyze and pattern match, as can all other forms of AI like Machine Learning, which has a long history in ad-tech. But the real trick is to have the mechanisms in place to reach out and ask a human being to intervene if the problem is not so clearly answered. Our customers can tune their risk tolerance on this, so if the issue is low risk, they can decide to let the machines make the decision. But for things that truly need human decisions to be made, we alert them and enable them to take the steering wheel.
Fluency’s platform operates across search, social and programmatic advertising channels. What technical challenges arise when an AI agent must reason across platforms that have different data structures, optimization systems and policy requirements?
Every ad platform has a unique set of best practices and capabilities. APIs are the fixed pipes we connect to in order to drive those platforms externally. APIs are very well documented, but that documentation is really the set of rules for what CAN be done on those platforms. They aren’t the embodiment of best practices for that specific platform.
LLMs can ingest lots of information into context so that the best practices for how to use each platform can be made available through them. Since Fluency is operating many platforms for many agencies and advertisers, we can encode each of our customers’ best practices (on their behalf) into the instructions that an LLM can follow to drive differentiated best practices for each customer. Our customers all have their own “playbooks” for how they operate their businesses, and Fluency lets them encode these in deterministic rules where possible, and over time we’ll enable them as instruction sets for LLM-powered agents where that makes sense.
We’re starting to work with our customers now on this next-generation of development to encode these more complex scenarios into their accounts on our platform. It’s very early days. Anyone who tells you they’ve solved any part of this using LLMs is stretching the truth, because things are so early, and the models are evolving so quickly that it hasn’t been fully solved by anyone. But we’re very well positioned to work with our customers to solve this.
What should effective human oversight look like in an agentic advertising environment? Which decisions can safely become autonomous, and which should continue to require human review or approval?
Every customer has a different risk tolerance, and most of them have never had a system that could express it. In Fluency, the judgment a good campaign manager applies in their head gets encoded once and then applied identically across every account, every platform, every time. The situations that don’t have a clean answer route to a person automatically, with the context they need to make the call.
What that looks like in practice varies a lot. We’re running a vast number of different rules for our customers today, helping them encode human decisions that don’t vary, and pushing to a human when the answer isn’t clear.
Which decisions can safely be autonomous depends on the customer, and I don’t think a universal answer is worth much. The platform has to enforce the line wherever a customer draws it, and let them move it as their confidence grows.
As agents make more campaign decisions, how should platforms document the reasoning, data and rules behind each action so that advertisers can audit performance, investigate mistakes and demonstrate compliance?
One of the core value propositions that Fluency offers is that we are SOC2 Compliant, and fully auditable. Changes are logged across the system. I’ve been a CPTO of a publicly traded company, and have been on the hook for compliance before, so I know how complex this can be. It’s another reason I was excited to join Fluency, because every aspect of what we do is fully auditable. We also maintain change logs for every change made in the system, logged against which user made it.
There’s a second layer that matters. Because the business rules are encoded in software, the rule is its own specification. There’s no drift between the documentation and the behavior, which is where audits usually fall apart. And in an LLM-enabled world, that encoded logic can be rendered back into plain business language on demand, so the person doing the auditing doesn’t have to read code to understand what the system was instructed to do.
For an agency or brand beginning to adopt agentic advertising, which workflows provide the best starting point, and what evidence should the organization require before expanding an agent’s authority?
From Fluency’s point of view, agentic advertising is well travelled ground. We’ve been doing it for eight years. We have hundreds of agencies and brands using our platform to do agentic advertising today, and in the multi-local space, that’s across many tens of thousands of locations.
We automate the majority of what humans do repetitively in setting up and managing campaigns, across search, social and programmatic, so there’s no workflow that makes a better or worse starting point. The best starting point is whichever channel our customer finds the most labor intensive.
LLMs are helping us speed up onboarding our customers as well, because as you can imagine, encoding a customer’s business rules is akin to performing a CRM implementation, measured in months. We’re building tools that use LLMs to help build the rulesets and write the automation scripts. This will make the onboarding process much quicker, and will reduce the complexity for our customers.
Advertising is an early testing ground for autonomous systems managing real money at scale. Which lessons from this industry could apply to financial services, healthcare, enterprise procurement or other regulated sectors adopting AI agents?
Advertising has been the proving ground for autonomous systems spending real money for 20 years. Programmatic taught the industry how to let software commit billions of dollars a day with no human approving individual transactions. Most of the current agentic AI conversation is happening at a scale advertising left behind a long time ago.
Fluency has over $3B of annual spend running through our platform, across hundreds of agencies and brands, and in the multi-local space that reaches tens of thousands of locations. Most agentic advertising conversations right now are about pilot budgets and a handful of accounts.
What transfers to other sectors is the architecture. You separate the systems that decide from the systems that execute. The deciding layer can be probabilistic, because judgment benefits from flexibility and creativity. The executing layer has to be deterministic, because money moving is not a place you want variance. Every action gets logged against the user or the agent that initiated it, which gives you an audit trail that holds up in front of a regulator.
That requirement is identical in financial services, in healthcare claims, in enterprise procurement. The stakes are higher and the tooling is generally worse. The design we built for advertising is the right one: Full auditability with change logs on every modification is table stakes. I’m particularly concerned about non-human judgement in things like approving medical procedures and approving insurance claims. I don’t think LLMs are there yet, and I have personal examples in my own life of automated systems making very bad decisions about approving or rejecting insurance claims.
My expectation is that the pattern advertising establishes becomes the pattern those sectors adopt. We’ve had a twenty-year head start on the problem of machines spending money hundreds of millions of times per second. Fluency’s approach extends from the transaction to the workflow, and I’m confident that our approach will similarly be the one that wins.
Thank you for the great interview, readers who wish to learn more should visit Fluency.