Table of Contents
TL;DR
Most AI agent projects don’t fail because the model is weak, they fail because the data underneath is fragmented, stale, or ungoverned. Mosaic AI and Agent Bricks assume trusted, accessible data; when that assumption is wrong, teams end up building one-off pipelines and shadow copies just to get a demo working, and the project stalls before production. The fix is inverting the usual sequence: pick one domain, build governed Unity Catalog tables with clear lineage and access policies first, then validate with a single narrowly-scoped agent before expanding. Unity Catalog governs what you register, it doesn’t clean your data for you, that engineering work is the actual prerequisite most teams skip.
Executive summary
Mosaic AI and Databricks’ new Agent Bricks make it easier than ever to build AI agents on proprietary data. But they assume that data is already trusted, accessible, and governed. In practice, most failures trace back to messy, ungoverned data, not models. The fastest route to production‑grade agents is to build Unity Catalog–backed governance and clean domain data first, then layer agents on top.
Introduction
Every board deck now has an “AI agents” slide. Leaders want autonomous systems that resolve customer tickets, optimize supply chains, detect fraud in real time, and personalize experiences at scale. Databricks Mosaic AI and Agent Bricks promise to connect powerful models to your enterprise data and orchestrate complex workflows in the lakehouse.
Yet the reality in 2026 is sobering. Most AI agent initiatives stall or never leave pilot. Postmortems show a common root cause.
Agents are being asked to reason on top of fragmented, stale, or ungoverned data. A governed data foundation with consistent schemas, clear lineage, and enforceable policies – is the prerequisite that separates pilot demos from production‑grade agents.
Our guide today explains why Mosaic AI needs clean, governed data to succeed, how Unity Catalog underpins trusted agents, and a practical “data first, agents second” rollout pattern that actually gets deployed. It closes with how Wishtree helps engineering teams build the bridge between messy operational systems and production‑ready AI agents on Databricks.
The pattern of failure
A familiar sequence plays out inside many enterprises.
- The company adopts Databricks and greenlights Mosaic AI or Agent Bricks pilots.
- A data science team starts building agents that must access CRM, ERP, warehouse tables, and multiple SaaS platforms.
- They discover that:
- – CRM customer IDs do not align with ERP account IDs.
- – The warehouse holds ten years of history with inconsistent schemas and undocumented ETL.
- – SaaS systems expose partial APIs or have incompatible auth and rate limits.
- To move forward, the team builds one‑off pipelines, lands shadow copies of data, and manually reconciles keys and semantics.
- After months of wrangling, nobody is fully confident which datasets the agent can trust. The project stalls or is restricted to a narrow demo.
Recent analyses of agentic AI rollouts underline the same issues – weak lineage, ad‑hoc access, and lack of centralized governance make it impossible to certify that agents are safe and reliable, particularly in regulated environments.
Leadership concludes that agents are not ready, when in reality the data platform and governance model were not ready.
This pattern underscores why data readiness for AI must come before agent development – cleaning, harmonizing, and governing operational data so Mosaic AI can build on trusted, consistent foundations rather than fragile shadow copies.
Why Mosaic AI needs governed data
Mosaic AI and the broader Databricks agent stack (including Agent Bricks, AI Gateway, and MLflow 3.0) are designed for enterprise‑grade generative and agentic applications.
But governed AI agents need trustworthy data, accessible through a shared catalog, with controls that ensure only approved datasets and tools are used.
If those properties are missing, Mosaic AI and Agent Bricks are forced to work on inconsistent, unvetted inputs. Databricks’ own materials emphasize that the Lakehouse plus Unity Catalog are the architectural answer that makes agents controllable by standardizing governance and discovery across federated systems.
Without that foundation, you get impressive demos but not production‑grade agents.
The missing layer: Unity Catalog as foundation
Unity Catalog is the unified governance layer for Databricks’ Data Intelligence Platform. For AI agents, it provides the substrate on which Mosaic AI can safely operate:
- – Lineage for data and AI assets
Unity Catalog captures lineage for tables, views, and pipelines, and recent guidance extends this to AI assets so you can trace which datasets and features feed a given agent or model.
This lineage capability is central to lakehouse governance, enabling organizations to answer regulators’ questions about which data influenced which decisions. Again, this is essential for fintech and healthtech deployments. - – Access control and policy enforcement
Fine‑grained privileges, row filters, and column masks let you limit what an agent can see, not just what a human user can. Databricks positions Unity Catalog and AI Gateway together as the way to constrain models, tools, and data access for each agent. - – Discovery of data, models, and tools
With a catalog that spans data, features, and AI endpoints, Agent Bricks and custom Mosaic AI agents can be configured to use approved sources and tools only. This reduces shadow data usage and accidental access to unvetted tables.
Unity Catalog, however, does not create clean data. It governs what you register. The hard work is taking messy operational systems and turning them into curated, well‑documented, governed assets in the catalog. That is the missing layer in many agent projects.
A better approach: data first, agents second
Instead of starting with the agent and hoping the data will sort itself out, invert the sequence.
1. Choose one domain
Select a domain where:
- – Data quality matters.
- – Business value is visible.
- – Scope is manageable.
Examples: Customer 360, transaction monitoring, product catalog, claims history, or support tickets. Industry playbooks on agentic AI recommend domain‑focused agents rather than giant, cross‑everything agents for precisely this reason.
2. Build your foundation in Unity Catalog
For that domain:
- – Identify all contributing systems (CRM, ERP, core banking, claims, marketing tools, etc.).
- – Engineer pipelines that bring the relevant data into Delta Lake with consistent schemas and keys.
For use cases requiring low‑latency decisions – fraud detection, ad bidding, inventory optimization, these pipelines must be built as real-time data foundations. This way, you can ensure that agents act on fresh information.
- – Register curated bronze/silver/gold tables in Unity Catalog with clear ownership, tags, and lineage.
- – Define and enforce access policies and, where needed, row/column‑level controls for sensitive attributes.
Data‑readiness guidance makes the same recommendation. Treat domain data products with clear contracts as prerequisites for AI workloads.
3. Validate with one AI agent
Once the domain foundation is in place:
- – Design a single Mosaic AI–backed agent that answers a tightly scoped set of questions or actions on top of those governed tables.
- – Use evaluation tooling (for example, MLflow 3.0–based evals mentioned alongside Agent Bricks) to measure accuracy, safety, and robustness before exposing it broadly.
- – Confirm that governance works – the agent only sees allowed data, and lineage and logs capture its activity.
This turns the agent from a speculative experiment into a concrete test of your data and governance maturity.
4. Expand to the next domain
Once the pattern works for one domain:
- – Reapply the same data modeling, governance, and agent design disciplines to adjacent domains.
- – Gradually connect agents or enable multi‑agent workflows where appropriate, still anchored in Unity Catalog policies and lineage.
This “data first, agents second” approach is less glamorous than jumping straight into LLM prompts, but it consistently produces agents that can actually be put into production.
Engineering the bridge between data and agents
For enterprises under time pressure, the gap is rarely in inspiration. It is in engineering the foundation that Mosaic AI requires.
Wishtree, a Databricks Partner, is a product engineering company focused on AI‑native systems. For Databricks customers, Wishtree helps build the bridge between messy operational realities and governed data that agents can trust.
- – Industry‑specific data models for Unity Catalog
Reusable schemas for fintech transaction processing, healthtech claims, adtech audience and campaign data, supply‑chain logistics, and HVACR operations that land directly into Delta + Unity Catalog with ownership, tags, and lineage defined from day one. - – Governance strategy and implementation
Practical Unity Catalog governance models – domains, catalogs, schemas, roles, filters, masks – that scale across business units, regions, and regulatory regimes without blocking experimentation. This includes coordinating policies with AI Gateway so Mosaic AI agents operate under the same rules. - – Production AI agents on Mosaic AI
Once the foundation is in place, Wishtree designs and implements custom agents using Mosaic AI and Agent Bricks, wiring them to governed data, tools, and workflows within the lakehouse. Agents are evaluated, monitored, and iterated with MLflow‑based metrics rather than left as fragile prototypes.
For deployments on AWS, Wishtree aligns Databricks with existing security controls – building compliant AI infrastructure where governance spans from networking and identity through Unity Catalog to the AI layer, meeting regulatory requirements for fintech and healthtech.
Contact us today to get started!
Frequently Asked Questions (FAQs)
Can AI agents be built without Unity Catalog?
You can build demos or small pilots without centralized governance, but Databricks positions Unity Catalog as the core governance layer for production‑grade AI agents, covering data, features, and AI assets. Without it, controlling and auditing what agents can see and do becomes extremely difficult at scale.
How long does it take to get one domain ready for AI agents?
Timelines depend on complexity, but many organizations can bring a single domain such as Customer 360 or transaction monitoring – from fragmented sources to curated, governed Unity Catalog tables within weeks if they stay focused. Building and validating the first agent on top then adds incremental time based on scope and evaluation depth.
What if our data spans multiple domains and systems?
That is normal. Agentic AI reports and data‑readiness frameworks recommend starting with one domain, establishing patterns, and then expanding, rather than trying to clean and govern the entire enterprise landscape at once. Multi‑domain agents can be introduced later, once the building blocks are stable.
How does Mosaic AI enforce governance when agents use external models?
Mosaic AI uses components such as AI Gateway and Agent Bricks to centralize control over models, endpoints, and tools, while Unity Catalog governs the data and AI artifacts they interact with. This combination lets you enforce policies, monitor usage, and audit agent actions across heterogeneous models.
Does Wishtree work with AWS‑hosted Databricks deployments?
Yes. Wishtree works with customers running Databricks on AWS, aligning Unity Catalog governance with AWS identity, networking, and security services.
What makes Wishtree different from traditional consulting firms?
Wishtree is an engineering‑led partner. Teams design data models, build pipelines, implement Unity Catalog governance, and ship production Mosaic AI agents. Customers retain IP and patterns, and Wishtree’s deliverables are running systems and reusable assets, not just strategy presentations.
How do we measure whether our data is “agent‑ready”?
Signals include governed domain tables in Unity Catalog with clear lineage and ownership, consistent schemas across key entities, enforceable access policies, and the ability to build at least one evaluated agent in a domain without extensive one‑off data wrangling. If an agent requires new custom pipelines for every query, the data is not ready yet.
Are small, domain‑focused agents really better than one giant enterprise agent?
Databricks and ecosystem partners increasingly advocate for smaller, domain‑specific agents with clearly defined capabilities, data access, and governance boundaries. This reduces hallucinations, improves predictability, and makes it easier to reason about and audit what each agent can do.
How does lineage help with AI agents specifically?
Lineage lets you see which tables, features, and pipelines feed an agent or model, and how outputs are used downstream. This is vital for debugging unexpected behavior, conducting impact analysis, and satisfying regulators that high‑impact decisions can be explained and traced.
What is the first concrete step if our Mosaic AI initiative is stuck?
Pick one high‑value domain where the agent is blocked by data issues, and run a focused data‑readiness and governance design exercise: model the domain in Unity Catalog, clean and integrate data into Delta, define policies, and then relaunch a narrow Mosaic AI agent on top. Proving success there creates a template for the rest of your portfolio.
