The ODI Story

Data infrastructure for agents you trust.

Open Data Infrastructure is the architecture for a world where agents query data constantly, autonomously, across systems. Every demo on this site is an instance of that architecture. This page is the narrative behind them.

Browse the demos Foundational blog post →

Why now

The Modern Data Stack was optimized for humans.
ODI is designed for agents.

Data infrastructure has always evolved to meet the moment. The Modern Data Stack solved the problems legacy systems weren’t built for: siloed data, scaled analytics, low barrier to entry. It worked beautifully — for humans.

The moment changed. AI is woven into every tech stack and is getting smarter, more capable, more deeply embedded. Human analysts query periodically. AI agents query constantly, around the clock, across systems, making autonomous decisions with every interaction.

The shift from hundreds of queries a day to thousands isn’t just volume. It’s architectural. Infrastructure built for human consumption can’t support agent-scale workloads reliably or economically.

ODI is the answer to this new moment — the same way MDS was the answer to the last one. Sellers who can speak to this shift earn seats in bigger, more strategic conversations.

GTM North Star

The goal isn’t to sell “ODI.” ODI is the architectural vision you use to open doors. What we sell is “Data infrastructure for agents you trust” — open storage on managed data lakes, data integration with connectors, and transformations that feed context and semantics for agents.

Agents & agentic AI

Agents are the new primary data consumer.

AI agents are autonomous software systems that operate continuously, take action on data, and make decisions without human intervention for each step. Unlike humans who use dashboards between 9 and 5, agents query at scale around the clock. That shift in consumer behavior amplifies both the value and the fragility of existing data infrastructure.

Modern Data Stack layer	Why it worked for humans	Why it breaks for agents
BI as the interface	Humans consume insights periodically via dashboards and reports.	→ Agents need data delivered directly into models, workflows, and applications — continuously.
Warehouse-centric architecture	Centralization provided unified data for reporting SLAs.	→ Agents need access to the freshest data across applications, operational systems, and models.
Semantics & business context lived in people's heads	Analysts and operators supplied business context and judgment.	→ Agents need context to be explicit, structured, and machine-ready.

What happens when agents act on bad data

Two scenes that explain why the foundation matters.

The Supply Chain Manager vs. The Supply Chain Agent

Human: A human supply-chain manager reviews inventory each morning and notices a recent shipment hasn't landed in the system. She doesn't panic — she cross-references the store portal, calls the warehouse, asks the customer. Judgment fills the gap.

Agent: An AI agent doesn't do that. It sees low inventory, follows its instructions, and places a large replenishment order for product that's already sitting in the back room. Overstocked. Capital tied up. Someone has to unwind the mistake.

The Uber Driver vs. Waymo

Human: Surprise parade downtown, not in any navigation app yet. Your Uber driver hits the closure, rolls down the window, asks a cop, finds an alternate route. No problem.

Agent: The Waymo knows what the map tells it. The map doesn't know about the parade. The Waymo sits there confused, takes a gridlocked route, or in a worst case makes a decision that puts it in front of the marching band.

This is why the foundation matters

Access to fresh data, reliable pipelines, reusable semantics, and centralized governance aren’t technical nice-to-haves — they’re must-haves. They provide context so you can trust the agents. They’re what separates an agent that helps the business from one that quietly causes damage.

Are we talking about Fivetran AI?

No. That’s a common confusion. Fivetran AI (the product) is still in development. The ODI conversation is about supporting the buyer’s AI initiatives — autonomous sales agents, AI-powered forecasting, operational agents, customer intelligence — by providing the data foundation those agents need to act reliably and economically.

ODI 101

Three pillars: Open, Data, Infrastructure

OPEN

Open standards — Iceberg, Delta, Polaris, SQL, dbt, MCP. Anyone can read it, anyone can build on it. No secret sauce that traps your data inside one vendor.

DATA

AI-ready data: normalized, deduplicated, schema-migrated, idempotently replicated. Trustworthy enough to feed an agent without a human in the loop.

INFRASTRUCTURE

The operating system for AI agents. Storage separated from compute, with the freedom to attach the right engine to each job — Snowflake, Databricks, Athena, Trino, Spark, Cortex, Claude.

The vanilla analogy

Open standards are like vanilla ice cream.

Anyone can make it. Anyone can serve it. Any topping works with it. No one owns vanilla.

Vendors often take that vanilla base and add their own secret sauce. Maybe the secret sauce tastes better. Maybe it works perfectly with their proprietary toppings. But the moment you want to take your ice cream somewhere else — a different shop, a different topping — the secret sauce becomes a problem.

Fivetran uses the vanilla. No secret sauce. Data is stored and moved using open, publicly-documented specifications any system can read, process, and build on — not formats that only work inside one vendor’s ecosystem.

The market shift

Legacy → MDS → ODI

How to walk a buyer or customer through the evolution chronologically.

1st

The legacy era

Siloed data, brittle DIY pipelines, slow analytics, high barrier to entry for small teams. It worked at the time. Scale was the ceiling.

2nd

The Modern Data Stack

Best-of-breed tools + cloud warehouses + self-service BI. Optimized for human analysts. It worked beautifully — for a while.

3rd

AI changes everything (again)

Data now serves AI agents alongside humans. Agents scale analytics work 1000×. More queries, more compute, more spend. A new scalability problem MDS was never designed for.

Now

ODI solves agent-scale

Open Data Infrastructure further separates storage from compute. Cheap, scalable storage. Interoperability to choose the best compute per job. Cost savings and data portability now; scalability without lock-in later.

Talk track

“MDS was optimized for humans. ODI is designed for a future with humans and production agents at scale.”

The five structural shifts

MDS → ODI, without the jargon

From — Modern Data Stack	To — Open Data Infrastructure
Warehouse-centric architecture	→ Decoupled storage (data lake) and compute (multi-engine)
Closed, proprietary systems	→ Standards-based infrastructure — interoperability, flexibility, portability
Analytics-driven decisions	→ AI-driven operations and intelligence analytics
Implicit vendor lock-in	→ Explicit freedom of choice
Fixed stacks; periodic rearchitecting	→ Continuously evolving architecture on open standards

What you sell today

Fivetran + dbt Labs make ODI possible

Sell the vision of ODI. Sell the features that bring it to reality. Plant the seed even when a full ODI conversation isn’t ready.

Managed Data Lake Service (MDLS)

The foundational product that operationalizes ODI. Open universal storage layer — separates storage from compute and automates data movement into Iceberg / Delta.

Connectors

750+ pre-built connectors deliver AI-ready data through normalization, deduplication, schema migration, and idempotent replication.

dbt Transformations

Standards-based transformations that don't lock logic to one compute engine. Centralized logic, lineage, orchestration, CI/CD, and observability — agents reason from trusted data.

The economic advantage

How ODI lets buyers reallocate spend

One of the most powerful hooks for CFOs and architecture owners. The chance to build trust by keeping their best interests at the forefront of the pitch.

The problem: compute-coupled architecture

When data is stored inside a warehouse, every query — including AI agent queries — runs on that warehouse's compute and pricing model. As agents scale from hundreds to thousands of interactions, costs multiply. If you 1000× your queries, can you afford 1000× the cost?

The compounding cost of lock-in

Vendor-bundled pricing means storage, compute, transformation, and governance all come at a premium. There's no way to route workloads to cheaper engines or avoid paying the vendor's margins on every layer.

ODI approach: storage first, then compute

With Managed Data Lake Service, data is ingested once into your storage budget. Multiple compute engines — a warehouse for analytics, an ML runtime for training, a vector DB for retrieval — access the same data without re-ingestion.

The result: dramatically lower TCO

Eliminate duplicate data storage and ingestion costs. Route workloads to the most cost-effective engine. Avoid vendor-driven pricing lock-in. As AI agent traffic scales, you don't pay exponentially more.

Buyer analogy

Like the move from proprietary chargers to USB-C.

Before USB-C, devices frequently had their own cables, ports, and adapters. Adding a new device often meant buying more accessories or staying tied to one manufacturer’s ecosystem.

USB-C changed that. One universal standard lets many devices connect to many power sources, screens, and accessories — without a different setup every time. ODI brings that same idea to data infrastructure.

Instead of locking data inside one platform’s compute and pricing model, ODI stores data once in an open foundation so different engines can plug in for the right job — analytics, ML, retrieval, AI agents. The result: more flexibility, less duplication, lower TCO as AI workloads scale.

ODI & partners

Vendor-neutral by design.

No single company — including Fivetran and dbt Labs — delivers all of ODI. Partners are essential because ODI is best-of-breed infrastructure built on open standards, not a single-vendor stack.

Fivetran + dbt Labs

→Move — data movement from source. ODI's journey starts at the source and ends at the source with activations.
→Manage — Fivetran's Managed Data Lake Service as the open, universal storage foundation. Data lands in customer-owned storage in open formats.
→Transform — standards-based, dbt-powered transformation logic. Governed semantic models, shared metric definitions.

Destination Partners

→Compute engines that read open table formats — Snowflake, Databricks, Athena, Trino, Spark, ClickHouse, DuckDB.
→Specialized runtimes for ML, vector retrieval, and downstream applications.
→Native catalog and governance integrations against the same lake.

Hyperscalers

→Cloud infrastructure: object storage (S3, ADLS, GCS), networking, identity, and security.
→Private networking endpoints that satisfy enterprise security requirements.
→The substrate that makes open storage cheap, durable, and accessible from anywhere.

The warehouse and lakehouse vendors are already here.

Leaders from existing warehouse and lakehouse vendors openly admit that decoupling storage and compute is the way of the AI future. They built the foundation we’re selling on top of.

Databricks

Donated Delta Lake. Acquired Tabular (the company founded by the original Iceberg author). Coined the term 'lakehouse.'

Snowflake

Built Polaris — an open-source catalog implementation for Iceberg.

What’s your ODI framing for partners?

When Snowflake or Databricks is already on the deal, this is the line. Stay above the data layer — they handle ingestion natively, but every agent on top of them needs governed, fresh data at scale.

When Databricks is on the deal

“Every AI agent running in Databricks needs governed, fresh data. We're the infrastructure layer that gets it there reliably, at scale, across every source.”

When Snowflake is on the deal

“Every AI agent running in Snowflake needs governed, fresh data. We're the infrastructure layer that gets it there reliably, at scale, across every source.”

Partner framing

ODI is a partner ecosystem, not competitive. We’re enabling freedom of choice across compute engines, clouds, LLMs, and tools — not replacing them. Fivetran sits at the foundation (data movement + open storage) alongside dbt Labs as the transformations and semantic layer.

GTM pro tip — discovery questions

Two questions that surface ODI opportunity:

1
“Are you finding that as you scale AI workloads, your compute costs are growing faster than expected?”
2
“If you needed to switch your primary compute engine tomorrow, or add a second one, how painful would that be with your current architecture?”

If the answer to either is “yes, it’s a problem,” you have an ODI conversation. If they’re Snowflake or Databricks heavy, the follow-up is the MDLS line: “What we do is put a neutral open storage layer underneath your compute so you have the flexibility to evolve without re-platforming.”

Objection handling

They see the value but still hesitate.

Show them you’re listening. Handle objections head-on and honestly — that’s how you become the trusted advisor.

Objection 1

“Why centralize at all? Can't we just federate?”

Federation tools are real — Snowflake data sharing, Unity Catalog federation, OneLake shortcuts. They work when a human runs an occasional cross-system query. They break when agents run continuously at scale against production source systems — your ERP, CRM, payments platform. Source load, unpredictable costs, governance gaps that compound invisibly. More importantly: you've already done the hard work of centralization. That's your competitive advantage. ODI doesn't ask you to tear it down. It puts an open lake underneath it so you're not locked into any single vendor's compute pricing as agents scale.

Keeping it real

MDLS isn’t available on hybrid. Here’s how to navigate.

Non-sensitive sources move to MDLS; mission-critical stay on hybrid.

MDLS covers ingestion cost. Lakes offer cheap storage at scale. Open table formats unlock interoperability with downstream compute, letting teams choose the most cost-effective engine per job. Even partial migration delivers significant savings.

MDLS does offer private networking.

Traffic routes over the cloud provider's internal network using AWS Gateway Endpoints, Azure VNet Service Endpoints, or Google Private Access. Object-storage endpoints are public, but bucket policies lock access to the right processing region. Satisfies most security requirements.

Plant the seed now; agent pressure is coming.

As access scales from hundreds of human queries to thousands of agent actions, breakages and compute spend spiral fast. The architectural decisions made now shape how much flexibility exists later. Even when MDLS isn't the right fit today, keep the ODI conversation alive.

Every demo on this site is ODI in working form.

Pick one. Walk through it. Use it to open the conversation.

Browse the demos