ABOUT THIS EPISODE
Traey Hatch of New Math Data on agentic AI: why every AI project is really a data project, and what agents need before they can run in production. The conversation covers the terminology blur between generative, physical, and agentic AI, the context work that decides whether a model is useful inside a specific enterprise, the cost logic behind frontier versus specialist models, and the identity, permissions, and audit gaps still keeping agents-supervising-agents out of production.
🎧 Episode Highlights
[2:22] Traey's path from data warehousing and machine learning into a full AI practice, and why nearly every client conversation has shifted.
[4:14] Reframing the McKinsey and MIT numbers on stalled AI projects. This is an experimental phase, not a failure phase.
[6:45] The word AI now absorbs machine learning, workflow automation, and analytics. What that blur costs organizations trying to scope work.
[10:51] Provenance as the adoption gate. People abandon a tool the first time it embarrasses them in front of someone else.
[15:43] What a harness actually is, and why the hard part is structuring context rather than picking a model.
[19:19] SAP means something different in every enterprise. Why tribal context has to be handed to a model explicitly.
[23:26] Onboarding an agent the way you onboard an employee: credentials, permissions, roles, and where the data actually lives.
[24:44] Choosing between frontier, specialist, and open source models as a cost decision tied to what the process is worth.
[28:55] The shoebox story. Fifteen years of uncashed checks in a permit office, and what it says about process nobody designed.
[33:34] The principal consultant approach. A negative proof of concept is a valid result, and off-ramps belong in the contract.
[37:37] The age of the operations person, and why process fluency is the durable advantage.
[39:37] Agents managing agents. Identity, permissions, and audit trails are the gaps, not model capability.
[43:15] The agent that cancelled a leadership team meeting, and where the human has to sit in the loop.
🔑 Key Takeaways
· Every AI project is a data project with a thin model layer on top. The visible work is the model. The actual work is information retrieval, data modeling, and structuring context so a language model can reach the answer you wanted. Organizations that scope these as model projects underestimate them by an order of magnitude.
· Process discipline predicts success better than model selection. Companies that already design, document, and run their business processes rigorously do well, because the technology fills gaps in work they can describe. Companies that cannot describe their own process end up with overlapping agents, messy transitions, and handoffs that never happen.
· Multi-agent systems are blocked on governance, not capability. Parallelizing work across sub-agents is well-trodden ground. Agents supervising other agents on real business decisions is not, because agent identity, permission propagation, spend thresholds, and audit trails are still being invented.
🧭 Frameworks Worth Saving
Onboard the model like a new employee. The pattern Traey returns to throughout the conversation:
· Terminology. Define your organization's specific language. A lane on a highway is not a lane in a transportation contract.
· Location. Tell the model where the data actually lives: which systems, which APIs, which warehouse.
· Permissions. Give it credentials and role-based access, the same way you would a new hire.
· Provenance. Surface how it reached an answer, so the people using it can judge the result.
Model choice is a cost decision, not only a capability decision. Route high-value, high-complexity work to frontier models. Route repetitive, well-defined work like document sorting or data entry to smaller, cheaper ones. The deciding question is what optimizing this process actually saves, and whether that justifies the engineering spend.
Crawl, walk, run, and most organizations are in the middle. Single-agent workflows and parallelized sub-agents are established. Agents supervising other agents on real business decisions is early, held back by unresolved questions on identity, permissions, and audit trails.
💬 Notable Quotes
It's a gigantic data project with a very thin layer of AI model use over the top of it.
Traey Hatch
A valid result of a proof of concept is negative.
Traey Hatch
I think that this is the age of the operations person.
Traey Hatch
👤 About The Guest
Traey Hatch, CEO, New Math Data
Traey Hatch is a co-founder of New Math Data and a cloud and data engineering leader focused on AWS, AI, and scalable analytics platforms. He works with organizations to modernize infrastructure, unlock data value, and deploy production-grade AI systems with an emphasis on security, governance, and long-term operability.
🎙️ About The Host
Derek Aranda
Derek Aranda spent over two decades as a global executive operating across commercial, supply chain, and digital transformation roles at scale. That span across the full value chain shapes his lens: the decisions, incentives, trust, economics, and governance that determine whether technology actually works inside real supply chains. On Supply ChAInge, he pressure-tests the autonomous future and helps leaders shape the framework to build it on their own terms.
Stay Connected
April 12 Advisors newmathdata.com
Produced by Speakerbox Media.
