Earley AI Podcast - Episode 90: Federated AI, Decision Intelligence, and the Data Architecture Reset with Todd Barr

In this episode, Seth Earley speaks with Todd Barr, CEO of Axonis.ai, a company spun out of a government defense integrator that is bringing federated AI and decision intelligence to high-consequence enterprise workflows. They explore why the demo-to-production gap is one of the most costly misconceptions in enterprise AI today, why centralization was built for business intelligence and not for AI, and what it really means to send your AI to your data rather than the other way around. Todd shares a candid and direct perspective on decision artifacts, AI cost exposure, the risks of vendor lock-in, and why enterprises that give away how they make decisions may be giving away the most valuable thing they own.

Key Takeaways:

• The demo-to-production gap is a form of malpractice - polished AI demos built on curated data create executive expectations that production reality cannot meet.
• Centralized data infrastructure was built for business intelligence, not AI - it is optimized for reporting, not reasoning or prediction.
• The premise of agentic AI is decentralization - if agents have to wait for data to be synced and centralized before acting, the architecture is working against itself.
• Data resists centralization for three distinct reasons: technical constraints, regulatory and compliance requirements, and organizational politics.
• Decision artifacts - cryptographically sealed records of data used, model applied, and reasoning followed - turn AI-assisted decisions into auditable, improvable corporate assets.
• Enterprises now face a clear choice: pay in tokens, pay in vendor lock-in, or invest in owning their own AI infrastructure through open source models.
• How an organization makes decisions is its most proprietary asset - giving that context to a third-party AI platform may be the most consequential thing enterprises are doing right now without fully understanding it.

Insightful Quotes:

"The misconception is really the gap between prototype and reality, and that's where a lot of these things are falling down right now. Getting people excited about something they can't have is almost malpractice." - Todd Barr

"The whole premise of agents is that they go out and do work for you autonomously. If you have to centralize and sync data before the agent can act on it, you're missing the point." - Todd Barr

"Centralization is almost a fallacy in itself. Whenever you are using data you are changing it, enriching it, doing something with it. It is a fractal nature of data that defies the whole concept of centralization." - Seth Earley

Tune in to discover why the most important AI infrastructure decision an enterprise can make right now is not which model to use - but whether they are building a foundation they actually own.

Links

LinkedIn: https://www.linkedin.com/in/tbarr/

Website: https://axonis.ai

Ways to Tune In:

Earley AI Podcast: https://www.earley.com/earley-ai-podcast-home Apple Podcast: https://podcasts.apple.com/podcast/id1586654770 Spotify: https://open.spotify.com/show/5nkcZvVYjHHj6wtBABqLbEiHeart Radio: https://www.iheart.com/podcast/269-earley-ai-podcast-87108370/ Stitcher: https://www.stitcher.com/show/earley-ai-podcast Amazon Music: https://music.amazon.com/podcasts/18524b67-09cf-433f-82db-07b6213ad3ba/earley-ai-podcast Buzzsprout: https://earleyai.buzzsprout.com/

Podcast Transcript: Federated AI, Decision Intelligence, and the Data Architecture Reset

Transcript introduction

This transcript captures a conversation between Seth Earley and Todd Barr about why the AI infrastructure assumptions most enterprises are working from are fundamentally misaligned with where AI is going. They cover the gap between demos and production, why centralized data lakes were built for the wrong era, how federated AI addresses the technical, regulatory, and political reasons data resists movement, what decision artifacts are and why they matter, and the hard choices enterprises face right now around AI costs, lock-in, and ownership.

Transcript

Seth Earley:
Welcome to today's Earley AI Podcast. I'm your host, Seth Earley, and in each episode we explore how artificial intelligence and data are reshaping business strategy and operations. Today, we're talking about one of the most persistent obstacles in enterprise AI - the gap between where data lives and where AI needs it to be.

Joining me today is Todd Barr, CEO of Axonis.ai, a company that was spun out of a government defense integrator and is now bringing federated AI and decision intelligence to high-consequence enterprise workflows. Todd has been part of major infrastructure shifts at Red Hat, Ansible, GitLab, and Chainlink Labs, and brings a pattern recognition perspective to what may be the defining infrastructure shift of our era. Todd, welcome to the show.

Todd Barr:
Thanks, Seth. Thanks for having me.

Seth Earley:
We like to start with common misconceptions. What is the biggest misconception executives have about AI readiness and the role that data architecture plays in it?

Todd Barr:
The misconception starts with how easy it is to make a really great demo today. Claude makes it super easy to show an executive something that looks real, feels real - you can click on it, it looks like their data. And in fact it is real, it just is not actually connected to anything. Getting people excited about something they cannot necessarily have is almost malpractice, because all of these things rely on access to their data, and in an enterprise that is not a trivial thing. The misconception is really the gap between prototype and reality, and that is where a lot of these things are falling down right now.

Seth Earley:
I call it the pilot paradox. The pilot looks great, people say give it to me, but it was built with lovingly curated, artisan-treated data - a limited, pristine dataset. Then you try to go to production and it is just not there.

Todd Barr:
Right. And the next misconception builds on that - the consultant says, well, it is great, you just have to connect it to your data lake, your Snowflake. And that is not exactly true either. Some of those tools were made for an era of business intelligence, not artificial intelligence. When you put your data in Snowflake, you are sanitizing it, normalizing it, cleaning it, syncing it - it is optimized for quick reporting, not for reasoning or making predictions from your data.

The other misconception is that everybody has centralized their data. The truth is, nobody has centralized all their data. You may have some centralized data. You may be in the middle of a multi-year centralization project that never ends. But some of your most valuable production data has not been centralized, and that is by design. The data you have centralized has a different purpose than the data out at the edge of your organization. Centralization is not the answer to AI readiness - that is really at the core of where the problem is today.

Seth Earley:
And centralization is almost a fallacy in itself. Whenever you are using data you are changing it, enriching it, doing something with it. If two different parts of the organization operate on the same data, where is the source of truth now? Different business units, different technology initiatives, campaigns, experiments - all of it is touching the data and changing it. It is a fractal nature of data that defies the whole concept of centralization.

Todd Barr:
And the whole premise of agents is that agents go out and do work for you - they can be autonomous. If you have to centralize and sync data before the agent can act on it, you are missing the point. The whole idea of an agentic approach is decentralized. You are sending agents out to do work for you, and so you really need to start thinking about how the agent is going to interact with your decentralized data.

Seth Earley:
So the centralization fallacy - organizations spent years and significant resources building data warehouses and data lakes, thinking that was the foundation for AI. What actually needs to change?

Todd Barr:
We do not want to throw away that work. There are great use cases for using AI to rapidly create better BI-type workflows, and centralized clean data will be used for that. But the mind shift is: I am going to send my AI to my data instead of bringing data to my AI. AI execution needs to happen where the data is. That is the thing that needs to change. Most people are conditioned to centralization first, but when you start asking what does AI execution look like when it is happening at the edge - who owns the data, what is the AI allowed to do, where is the security - that is when the architecture question really opens up.

Seth Earley:
You talk about data that resists centralization for three distinct reasons - technical, regulatory, and political. Walk us through those.

Todd Barr:
The first is a use case reason - you have transactions or real-time decisions where you cannot wait for data to be synced, transformed, and centralized before you act. The second is sensitive data - PII, HIPAA, GDDR - where there is a compliance cost and a risk cost to moving and having multiple copies of the data. The third is data that is expensive to move because it is large - video data, sensor data from a factory floor - where the bandwidth and cost of moving it simply does not make sense.

And then the last one is data that is organizationally resistant to movement. That is a social problem inside a company - application owners who say this is my data, or situations where you want to share intelligence across organizations but you cannot share the raw data. Hospitals are not going to share their patient data with another hospital system, but if you could share intelligence and train models without sharing the underlying data, that is a real unlock.

We actually came to this from our parent company operating in secure government and intelligence environments, where all three of those problems exist simultaneously. You have to be able to go offline and come back online. Security has to be applied at the decentralized node level - down to the cell level with attribute-based access control - and that security has to be obeyed throughout the entire federation.

In a government example: a person with top-secret clearance can run a model training job on top-secret data. A person who only has confidential clearance will never see the top-secret models or the top-secret data. They will not even know it exists in our system. Security at the source, not at the perimeter.

Seth Earley:
Tell me about decision artifacts. What are they, and why does capturing AI-assisted decisions as structured, auditable data matter for high-consequence workflows?

Todd Barr:
Once we had the federated AI layer, we started thinking about what people are actually doing with it - they are making decisions with the help of AI, and they want to do more of that because it speeds things up. We realized we sit in the perfect place to capture everything happening across their data and the decisions being made using AI.

The EU AI Act and similar regulations are creating requirements for human oversight, record keeping, traceability, auditability, and monitoring after the fact. People tend to think about traceability as being about the model itself - model drift, model bias, compliance. But the other side is what are humans and AI doing together, or what are agents doing by themselves? We call that the decision plane for compliance.

So we created a standard called the decision evidence specification, and implemented it. It captures: what data was used, what snapshot of that data was in effect at the time, what AI model or LLM was used, what the chat or reasoning trace looked like, and whether the model stayed within the approved data space or hallucinated outside it. All of that gets packaged as a cryptographically sealed artifact.

Two years from now, customers will have built up a decisions artifact database that is a genuine corporate asset - the record of how their company makes decisions. You can then run retrospective analysis: how accurate were these decisions, which model performed better on which type of judgment, should we fine-tune a smaller model for this specific use case? In every company I have worked for, decisions were only recorded in board meeting notes or OKRs. Now you can turn decision-making into a flywheel that improves over time.

Seth Earley:
How do you communicate the realistic path to production without scaring executives off, given how much they have already been burned?

Todd Barr:
There is a new wrench in the problem - the industry is moving from subscriptions to consumption pricing. Microsoft just announced Copilot is moving to a full consumption model. Nobody knows how many tokens their use cases are actually going to require at scale, and the costs can spiral quickly when agents start regenerating work. So executives now face a choice they did not have two years ago.

My three recommendations for C-level leaders: first, start incremental. Solve one real problem, prove it, then add another data source and another use case. Not wholesale architecture shifts. Second, build optionality in. Do not commit your entire AI strategy to one vendor - use gateways, build the ability to swap models and vendors into your architecture from the start. Third, start thinking now about owning your own AI. Models are improving fast, open source is advancing quickly, and there will be a point - probably within one to four years - where you will want your own models running to control costs on a capital basis rather than as unbounded operating expense.

The core decision is: are you going to pay in tokens, pay in lock-in, or build toward ownership? Because right now, April 2026, the era of being subsidized by big tech to explore AI is over.

Seth Earley:
You made the point that how an organization makes decisions is its most proprietary asset. Say more about that.

Todd Barr:
If I am an enterprise, what do I own in this day and age? I own how I make decisions. Which data I use to make decisions. If I commit all of that context to a third-party AI platform - and they have every assurance in the world that they are not training on it - do I really know that? I am not saying big tech is bad. I use Google all the time, and Google knows more about me than my wife does. But if I am an enterprise trying to maintain and grow shareholder value, giving away how I make decisions is like giving away my brain. That is the wisdom generated from being in your industry. That is what is proprietary. And right now, we are on a path to give it all to trillion-dollar companies who will then have the capability to essentially replicate what every industry does. Enterprises need to understand what they are doing here.

Seth Earley:
And on the Decision Evidence Specification - give us a quick thumbnail of why the industry needs that shared standard.

Todd Barr:
There is a whole set of standards needed around AI, and open standards tend to give both regulators and the companies being regulated some certainty about what they can do next. There is a lot of uncertainty in AI right now, and open standards can provide a foundation to move forward. The decision evidence specification is our contribution from our work in government and intelligence, where this kind of traceability has been required for a long time. We are currently working with healthcare and finance organizations to get it out into the world. There is a lot more that needs to happen around agents and delegating authority and the legal frameworks around agentic systems, but this is a starting point.

Seth Earley:
Todd, thank you so much for joining us. It has been a great conversation with a lot of important and candid insights. And a final thank you to our sponsor, VKTR. To our listeners, thank you for tuning in to the Earley AI Podcast. You can also find us on dLogos at dlogos.xyz/podcasts/earley-ai-podcast-271271ce, where you can vote on upcoming guests and submit questions and topics for future episodes. Be sure to subscribe for more conversations about how AI is shaping the future of business.

Todd Barr:
Thanks, Seth.