Most enterprises now have the compute for AI. What they don't have is a clean way to feed it. One line comes up in almost every customer meeting Michael Hay sits in: "I've got my compute, and it's doing a really good job at making heat." The GPUs are paid for and the models are capable, yet the AI agents still can't answer a business question, because the data they need sits in Salesforce, an on-prem Oracle database and a Databricks workspace, with no clean path between them.
That gap is what an AI data harness is built to close. It connects AI agents to data where it already lives, applies business rules and quality checks in one place, and gives the agent a single, governed view of live data. On Mike's own side project, that meant going from about 800 lines of networking code per source to zero.
In the latest episode of Talk Data To Me, Zetaris CEO Vinay Samuel asks Mike, now CTO of Zetaris after years at Teradata and Hitachi Vantara, why data has become the brake on AI, what a data harness changes for engineers, and which skills matter as AI writes more of the code. Watch the full conversation below, or read on for the key ideas.
[VIDEO EMBED: Why AI Agents Are Only as Good as Their Data | Talk Data To Me, https://youtu.be/LiloNn1Uhrk]
Jump to a chapter: 01:18 Mike's background · 06:08 The Open Agent Hackathon · 09:02 Compute is heat, data is fuel · 09:41 How the data harness works · 11:00 800 lines versus zero · 12:42 Getting started · 15:14 Advice for engineers
Every agentic AI ambition runs on the same fuel. Mike's framing is blunt: without data, compute does nothing more useful than spin the fans and pump the water.
The issue is rarely a shortage of data. It's that the data an agent needs is spread across systems that were never designed to talk to each other, each with its own access method, schema and quality problems. An agent asked to explain a drop in renewals might need CRM records, support tickets and product usage before it can reason about anything.
Vinay named the irony in the episode. Data people have become the stars of the AI era, and data is also the thing slowing AI down.
The default fix is custom pipelines. Every new source brings networking code, authentication, schema mapping and quality checks, written by hand and maintained for as long as the source exists.
Mike has measured this on his own time. As a personal project, he maintains a Python pipeline that pulls industry classification data, the coding systems used to group businesses by sector, from several public sources. Per classification system, it takes:
Multiply that by every source an agent needs and it's clear where the engineering time goes. Most of it is plumbing. Very little of it is the agent.
An AI data harness starts from a different premise: send the query to the data instead of moving the data to the query. Mike describes three layers.
The query layer at the base. Zetaris connects to sources where they live and runs distributed queries across them, with caching underneath to keep repeat queries fast. Nothing has to be copied into a new store first.
A unified semantic layer in the middle. The unified semantic layer holds business logic, business rules and data quality rules in one place, so every consumer gets the same definition of "active customer" or "net revenue."
Location-blind consumers on top. An agent, an analyst or another system queries one interface. As Mike puts it, the consumer "doesn't know where the heck it came from," whether that's Salesforce, on-prem Oracle or Databricks. It just gets clean, governed access.
Mike's phrase for the result is "deceptively simple." The engineering underneath is deep, so the experience on top doesn't have to be. You can read more about the architecture in the Zetaris Knowledge Base (https://kbase.zetaris.com/knowledge).
Here is the comparison Mike shared from that hobby pipeline. Doing nearly the same job on Zetaris takes:
It's one engineer's side project, not a controlled benchmark, and Mike is careful to call it "nearly equivalent." But it shows where the effort moves. When connectivity and quality rules are platform features, the week you would have spent on plumbing goes into the agent itself.
The clearest current example is the Open Agent Hackathon 2026, an online event run by GenAI Works from 22 to 27 October, with Zetaris, NVIDIA and SpaceXAI among the sponsors. More than 5,000 builders registered to compete for up to $20,000 in prizes. Mike led the Zetaris side, working with GenAI Works and Steve Nouri, Zetaris AI Strategy Advisor.
The hackathon stack follows the same layers:
Zetaris engineers, product staff and forward deployed engineers are covering support around the clock across Europe, US Central and US West time zones. Mike is also a judge, which, as he told Vinay, puts him "on both ends of the discussion."
The same path works outside the hackathon. Anyone can follow the guide and run a first query against sample data on a free Zetaris Cloud account.
Vinay closed by asking what early and mid-career engineers should be doing as AI reshapes the job. Mike's answer reaches back decades: the return of partial waterfall development.
His reasoning is that when generative AI and agentic tools handle more of the execution, the scarce input is context. Someone has to understand what customers need, what the market wants and what competitors are doing, then write it down precisely enough for human and AI engineers to build from. Mike calls it spec-based development, and he sees the same shift happening in marketing and other functions, not just engineering.
It's a view shaped by where he started, in customer support and professional services. The skills he recommends building now:
"Context is king," Mike says. That's the part of the job the tools can't do for you.
An AI data harness is a layer between AI agents and enterprise data. It connects to sources where they already live, applies business rules and data quality checks in a unified semantic layer, and gives agents one governed interface to query. The agent gets consistent, trustworthy data without engineers building a custom pipeline for every source.
No. Zetaris queries data where it lives, across cloud, on-prem and SaaS sources, and uses caching to keep repeat queries fast. Your existing systems stay in place, and the agent or analyst sees one consistent view across them.
Spec-based development means writing detailed specifications, grounded in customer, market and competitor context, before execution starts. As AI tools write more of the code, Mike Hay argues that the quality of the spec increasingly decides the quality of what gets built.
Sign up for Zetaris Cloud, which has a free tier, then follow the getting started guide on GitHub to run your first query against sample data. Questions along the way can go to the Zetaris community on Discord.
Sign up for Zetaris Cloud for free, follow the getting started guide, and see what your agents can do when the plumbing is already handled.