— AI on the lakehouse · Genie, agents and retrieval
Databricks AI consulting that makes AI tell the truth.
Genie, Agent Bricks and Databricks AI Search all read whatever your gold layer says. If that layer does not reconcile, the assistant answers confidently and wrongly. We build the governed lakehouse data first, then the AI that sits on top of it.
30+ years inside IBM · EY · J&J · McKesson
Lean Six Sigma Master Black Belt
AIMContext from $3,500
60-Day Ship Guarantee
— The pattern · Three failure modes
Most AI pilots fail on the data, not the model.
01
Answers nobody can defend
Natural-language questions over a table nobody conformed return a number nobody can defend. The fix is semantics in Unity Catalog and a gold layer that means one thing.
02
Agents that read too much
An agent that can read everything will eventually quote something it should not. Access, evaluation and retrieval scope have to be designed, not bolted on after the demo.
03
Retrieval that goes stale
Retrieval fails quietly when the index holds stale, duplicate or unowned content. The index is a data product. It needs an owner, a refresh schedule and a quality check.
— The work · What we build
Databricks AI consulting, built from the gold layer up.
Most engagements start with the governed data and bring in the AI once the numbers reconcile. These are the six pieces we build, in your workspace, with your team in the room.
Governed gold layer
Gold views that conform your sources into one definition per metric, so every assistant and every dashboard on the lakehouse reads the same number.
Semantics in Unity Catalog
Table and column definitions, certified metrics and ownership written into Unity Catalog, the layer Genie treats as its source of truth for data and meaning.
Curated Genie spaces
AI/BI Genie, generally available since 12 June 2025, set up on curated tables and tested against questions your finance team signs off before launch.
Agents with Agent Bricks
Agents scoped to governed data. Agent Bricks launched in Beta in June 2025, and we confirm which components are generally available before we design around them.
Retrieval with AI Search
Databricks AI Search, formerly Vector Search, over governed documents, with the index run as a data product that has an owner and a refresh schedule.
Lakebase for application state
Lakebase, the managed Postgres that reached general availability on AWS on 3 February 2026 and on Azure on 3 March 2026, where an application needs transactional state.
— Why the gold layer comes first
The model is rarely the problem. The context is.
When an agent hallucinates on company data, it is usually doing exactly what it was asked with the context it was given. Two tables disagree, a metric has three definitions, a document is five versions old. We fix that underneath, in the lakehouse, so the model has one version of the truth to read.
The mechanism is set out in why AI agents hallucinate on company data, and what retrieval costs to build and run is broken down in enterprise RAG implementation cost.
— The method · AIM-IT
Five phases. Five artifacts you can sign off.
Assess, Innovate, Model, Implement, Track. Each phase of AIM-IT ends in a named deliverable, not a slide.
01 · ASSESS
Assess
Question inventory
Which questions the business will actually ask, and which lakehouse tables can answer them today.
02 · INNOVATE
Innovate
Gold layer design
The gold layer and semantics the assistant will read, designed before any prompt is written.
03 · MODEL
Model
Tested answer set
Proof on real data: a test set of questions and answers your finance team agrees with.
04 · IMPLEMENT
Implement
Live Genie space and agents
The Genie space, agents and retrieval built in your workspace, with your team in the room.
05 · TRACK
Track
Answer quality control plan
Evaluation, owners and review dates, so answer quality is still measured in ninety days.
— The deliverable · What you get
What you own at the end.
- A gold layer and Unity Catalog semantics the AI reads from.
- A curated Genie space, tested against questions your team signs off.
- Agents and retrieval scoped to governed data, with evaluation built in.
- A written control plan for answer quality: what is measured, how often, and who owns it.
- Your team able to run it. We are not trying to become permanent.
14 days
Typical first build
5 phases
AIM-IT, end to end
$3,500+
Starting price
60-Day
Ship Guarantee
On every Sprint engagement: if we don’t deliver the agreed working artifact in 60 days, you don’t pay the final invoice.
— Who does the work
Principals on the work, not junior analysts.
Frank “Rio” Shines spent 30+ years implementing enterprise systems inside IBM, Ernst & Young, Johnson & Johnson and McKesson, and is a Lean Six Sigma Master Black Belt. That is the discipline behind the gold layer: define the question, measure the defect, then build.
The platform side, from medallion design to Unity Catalog, is covered on our Databricks consulting page, and the pipelines that feed the lakehouse are covered under data engineering.
— Related reading · From the blog
Go deeper on AI on the lakehouse.
Genie
Databricks Genie: making self-service BI answers trustworthy
Why Genie answers only as well as the gold layer and semantics it reads.
Agents
Agent Bricks: building production AI agents on the lakehouse
What moves an agent from demo to production on governed data.
Retrieval
RAG on Databricks: AI Search, formerly Vector Search
How to build retrieval that stays accurate as documents change.
Lakebase
Databricks Lakebase: when Postgres on the lakehouse makes sense
Where transactional state belongs next to your analytical data.
Cost
Enterprise RAG implementation cost: a real breakdown
What retrieval costs to build and run, line by line.
Context
Why AI agents hallucinate on company data
Why it is a context problem, not a model problem.
— Straight answers · The objections answered
Frequently asked questions (FAQs)
Are you a certified Databricks partner?
No, and we will not imply otherwise. Certification is in preparation. What we bring is thirty years of delivering the systems this data comes out of (ERP, quality and BI inside IBM, Ernst and Young, Johnson and Johnson and McKesson), plus Lean Six Sigma at Master Black Belt level. Ask us for the work, not the badge.
Is Agent Bricks generally available?
Parts of it. Agent Bricks launched in Beta in June 2025, and Databricks has since confirmed named components, including Document Intelligence and Custom Agents, as generally available. We check the status of each feature before we design around it, and tell you which parts are still in preview.
Can Genie replace our BI tool?
Usually not, and we would not sell it that way. Genie is strong for ad hoc questions over curated tables. Your governed dashboards still matter, and both should read the same gold layer so they never disagree.
Do we need a lakehouse before we do AI?
You need governed, reconciled data. On Databricks that is the lakehouse. If your data is not ready we will say so, and the first engagement becomes the data readiness work, not the agent.
What does Databricks AI consulting cost?
AIMContext starts at $3,500 for a fixed-scope governed data and retrieval foundation. Larger agent programmes are quoted after the assessment. Every engagement carries the 60-Day Ship Guarantee.
Start with the data, not the demo.
One call, your real questions on the table, and an honest read on whether your lakehouse is ready for AI.