Microsoft Fabric Microsoft Fabric AE Talk to us
Agentic data engineering

From engineering-heavy teams to AI-led delivery.

Data teams have long been the bottleneck between a question and an answer. Agents change the shape of the team, not its accountability: they generate the code, engineers set the direction and validate the result.

See how the work is changing ↓
The evolution of AI in engineering

From autocomplete to agents that do the work.

Models can now use tools, read a repository, run a test and correct themselves. That turns a suggestion into work you can delegate and review.

Completion
Line by line

The model predicts the next few tokens inside the file you have open. Useful, but it knows nothing about your platform.

Conversation
Ask and paste

Chat assistants explain, refactor and draft. The engineer still carries every fragment into the codebase by hand.

Tool use
Agents that act

With MCP connections the agent reads the repo, queries Fabric, opens a pull request and runs the tests it just wrote.

Delegation
Agents that work together

Narrow agents hand work to each other on backlog tickets, with skills and reusable blocks as their guardrails and a human at every gate.

The evolution

Four steps from typing code to directing it.

Teams move along one task at a time, and the instruction gets shorter as the guardrails get better. Most data teams we work with sit between step two and three today. Slide to see where the hours go.

Step 02
Human with assistant

Engineers use an AI assistant that helps them work better and faster. The work is still theirs, line by line.

“Write a SQL statement that…”
Human effort per data product75%
Human judgement per decision100%
Time to a first data product6 weeks

Judgement never drops: someone still decides what should exist and confirms it is correct. What falls is the typing in between.

What changes for the team

The same roles, spending their time differently.

A data product still needs planning, a definition, a build and active monitoring. What shifts is where the hours go: away from writing transformations, towards deciding what should exist and confirming it is correct.

RoleEngineering-heavy teamAI-assisted team
Data product ownerWaits for a delivery slot in the backlog.Owns vision, value and ROI, and steers the agents from definition through delivery.
Business / domain expertWrites requirements, then disappears until UAT.Brings the use case and validates meaning while the product is being generated, in days rather than months.
Data engineerIngests, transforms, models — by hand, ticket by ticket.Directs agents, reviews generated code, and owns the hard cases agents cannot reason about.
Data architectFrames the target design, then polices drift after the fact.Encodes standards as agent skills and reusable blocks, so the guardrails are executable.
Platform / DevOpsProvisions environments and automates pipelines on request.Provides the agent runtime, the MCP connections and the monitoring the agents report into.
The agents

Seven agents, one task each.

Each agent owns one task in the data product lifecycle and reports in human-readable text on backlog tickets, so an engineer can follow the process and step in at any point.

01
Analysis

Turns the data product definition into a backlog of stories.

02
Infrastructure

Environments as code, CI/CD pipelines, access groups and secret management.

03
Ingestion

Ingests data from the different source systems.

04
Transformation

Writes the ETL transformations into the layers of the medallion architecture.

05
Reporting

Creates the semantic model and the dashboards in a tool such as Power BI.

06
Profiling

Profiles the source system and suggests data quality checks.

07
Monitoring

Monitors pipelines, investigates issues and suggests fixes.

Speed up your data delivery
Trusted data products in days, not quarters.

Get the return on your data investment while the business case is still relevant.

I want to speed up my data delivery
Agentic data engineering

From engineering-heavy teams to AI-led delivery.

Data teams have long been the bottleneck between a question and an answer. Agents change the shape of the team, not its accountability: they generate the code, engineers set the direction and validate the result.

See how the work is changing ↓
The evolution of AI in engineering

From autocomplete to agents that do the work.

Models can now use tools, read a repository, run a test and correct themselves. That turns a suggestion into work you can delegate and review.

Completion
Line by line

The model predicts the next few tokens inside the file you have open. Useful, but it knows nothing about your platform.

Conversation
Ask and paste

Chat assistants explain, refactor and draft. The engineer still carries every fragment into the codebase by hand.

Tool use
Agents that act

With MCP connections the agent reads the repo, queries Fabric, opens a pull request and runs the tests it just wrote.

Delegation
Agents that work together

Narrow agents hand work to each other on backlog tickets, with skills and reusable blocks as their guardrails and a human at every gate.

The evolution

Four steps from typing code to directing it.

Teams move along one task at a time, and the instruction gets shorter as the guardrails get better. Most data teams we work with sit between step two and three today. Slide to see where the hours go.

Step 0{{ step }}
{{ stepTitle }}

{{ stepText }}

{{ stepPrompt }}
Human effort per data product{{ effort }}%
Human judgement per decision{{ judgement }}%
Time to a first data product{{ lead }}

Judgement never drops: someone still decides what should exist and confirms it is correct. What falls is the typing in between.

What changes for the team

The same roles, spending their time differently.

A data product still needs planning, a definition, a build and active monitoring. What shifts is where the hours go: away from writing transformations, towards deciding what should exist and confirming it is correct.

RoleEngineering-heavy teamAI-assisted team
{{ r.role }}{{ r.before }}{{ r.after }}
The agents

Seven agents, one task each.

Each agent owns one task in the data product lifecycle and reports in human-readable text on backlog tickets, so an engineer can follow the process and step in at any point.

{{ a.n }}
{{ a.name }}

{{ a.d }}
Speed up your data delivery
Trusted data products in days, not quarters.

Get the return on your data investment while the business case is still relevant.

I want to speed up my data delivery