
Why Every Data Team Needs a AI Context Engineer
An AI context engineer designs and maintains the enterprise context AI needs to answer questions accurately, including metric definitions, business rules, verified queries, and institutional knowledge.
Demand for the role is growing fast. 73% of companies already have someone dedicated to owning AI context, and most of the rest plan to hire within 2 years.
The role combines SQL, semantic modeling, and context delivery, with a growing toolset that includes MCP, dbt, and metadata catalogs.
PUBLISHED:
UPDATED:
Picture this: your data team just rolled out an AI analyst, and everyone’s stoked. Instead of filing a ticket and waiting days for an answer, users can now ask the AI directly and get a response in minutes. For the first week, it feels like the ticketing nightmare is finally over.
Then the Slack messages start. A sales director wants to know why AI says revenue is up 12% when the dashboard says 8%. Minutes later, someone asks why the customer count looks off. By the end of the second week, the thread is growing faster than you can answer it.
Before long, you’re spending more time verifying the AI’s answers than you ever spent building dashboards. And you're not alone. In our survey, only 1 in 5 data leaders said they’re confident in AI-generated answers, with most still relying on an expert to review insights before making a decision.
Simply put, AI doesn’t understand your business. And closing the gap falls to the AI context engineer, a role companies are racing to hire.
What does an AI context engineer do?
An AI Context Engineer designs, manages, and optimizes the information entering an LLM’s context window. Their job is to determine what AI context an LLM needs to complete a task, where this information should come from, and how to organize and deliver it so the model can use it effectively.
The role is easy to confuse with prompt engineering and analytics engineering, but the two solve different problems.
Prompt engineering vs. context engineering
A prompt engineer focuses on perfecting a single interaction, often by refining how a question is phrased or adding instructions until the model gives a better answer. This approach works well for stable, single-turn tasks like drafting an email or summarizing a document, where the model already has enough context.
Complex analysis is a different story. You can’t expect hundreds of business users to write the perfect prompt every time or to know every exception their analysts apply. And when a business rule changes, hundreds of saved prompts can quietly go stale.
An AI context engineer takes the opposite approach, treating context as a production system. It’s versioned, governed, and built into a shared context layer for every AI agent, dashboard, and user to draw from. Most practitioners start by encoding business rules and metric definitions, then documenting the institutional knowledge living in people’s heads.
With this foundation in place, the focus shifts to building context, running evals, investigating failures, and keeping answers accurate even as the business changes.
Why demand for AI context engineers is rising
AI has reshaped nearly every role in the enterprise, and the data team is no exception. In our survey, 86% of data leaders said AI has fundamentally changed their team’s core responsibilities, and half said the change has been dramatic. More importantly, most leaders are still figuring out how to make AI analytics reliable at scale.
Take Anthropic. Its data team built an internal analytics system on Claude, and the project took months of serious context engineering before the agent could work reliably in production. Yet within a month of launch, accuracy dropped from 95% to 65%, all because the context had gone stale and there was no dedicated owner to keep it in sync with the business.
Stories like these explain why demand for AI context engineers is rising. Building trust in AI takes a rare mix of skills: a deep understanding of the business, a working knowledge of how AI systems work, and the technical ability to build complex processes to retrieve and manage context. Few people on today’s data teams have all three, and without a dedicated owner, AI failures pile up.
What does the career outlook look like?
As more companies move AI into production, they need someone who can define business logic, manage context retrieval, test AI-generated answers, and catch gaps. This is exactly the job AI context engineers are hired for.
The skills behind this work are already in high demand. Our report found some interesting facts:
73% of companies already have someone dedicated to owning AI context. Of those who don’t, a third have open positions, and the rest plan to hire within the next 2 years.
74% are still building a complete, shared definition of enterprise context, and 94% plan to invest in better ways to manage it within 18 months.
57% of data leaders plan to upskill their existing teams, while about 20% plan to hire from outside.
These numbers point to a career path with real promise. If you already work with data, you have a head start. And if you’re just starting out, you’re getting in on the ground floor of a role still being defined, where a few years of hands-on work can make you an expert.

Real-world example: How do context engineers work?
Context engineering is a very different job from the ticket-taking work data teams have endured for years. Success in this role is measured by how confidently the business makes decisions, how much users trust the answers they get, and how quickly you catch and fix AI failures before they disrupt the business. Arm’s data team offers a good example of what this looks like in practice.
Arm, a nearly $5 billion supplier of chipsets and software, used to rely on manual reporting to answer even basic procurement questions. Tom Smith, Arm’s Procurement Governance, Policy, and Reporting Lead, describes it as “a heavyweight manual process.” Today, the team uses WisdomAI to let employees explore spend, supplier, and contract data on their own.
The bigger shift is in the role of the data team itself. Smith’s team used to be defined by how many reports and data requests it fulfilled. Now, with AI handling much of the ticket queue, he calls that work “the easy part.”
Now, their team is focused on helping AI understand the business. They spend their days reviewing how people use AI, examining the queries it generates, and tracing wrong answers back to where context broke down. Smith calls it “a totally different problem from just building dashboards,” and it already takes up about a quarter of the team’s time, a share he expects to grow as adoption deepens.
Roles and responsibilities of an AI context engineer
As an AI context engineer, your role sits at the crossroads of data engineering, analytics, and machine learning. You’ll lean on each of those skill sets as you take on these core responsibilities:
Write and review SQL
SQL is still the foundation of AI data analysis. When AI writes a query, you need to read it and spot problems quickly, whether it’s a join fanning out rows, an aggregation at the wrong grain, or a window function partitioned on the wrong key.
Fluency with CTEs and subqueries matters too, along with knowing how SQL dialects differ across warehouses like Snowflake, Databricks, and BigQuery. Most importantly, strong SQL skills help you tell whether a bad answer came from the model’s query logic or from missing context, so you know which layer to fix.
Map schemas
Schema mapping is the process of documenting how your tables relate to one another, so AI knows which keys connect them and how to join them correctly. By mapping table relationships and enriching schemas with metadata, descriptions, and ownership details, you give the AI agent a path to retrieve the precise information requested rather than guessing, hallucinating a join, or querying the wrong table entirely.
Model semantics
Many warehouse columns are named in ways only their creators understand. A field like flg_x2_adj might make sense to the engineer who built it last year, but to anyone else, it’s just a string of letters.
You fix this ambiguity by giving every important field a clear, documented meaning. You’ll be defining what each business metric represents, documenting how it’s calculated, and writing clear descriptions. These definitions then go into a semantic model, which sits on top of the warehouse and translates raw tables into business concepts like customers, orders, and active users. As a result, when someone asks how many active users signed up last month, AI uses the exact definition your analysts use.
Structure the context layer
Even with metrics defined in a semantic model, a bigger problem remains. Some of your business’s most important context was never documented. It lives in your analysts’ heads, or it's buried across old Slack threads, docs, and dbt comments nobody remembers writing.
When AI can’t find this context, it improvises, so two people asking the same question get completely different answers. Bringing everything into a context layer is one of the most important parts of your job. A well-built layer includes:
Metric definitions: The agreed-upon formula for each business metric, documented with its exact calculation logic
Business rules: The conditions, filters, and exceptions your analysts apply to get correct results
Institutional knowledge: Undocumented know-how your team has built up over years
Verified queries: SQL your team has reviewed and approved for AI to use as trusted examples
Version history: A record of who owns each piece of context and how it has changed, so every update is traceable
Keeping the layer accurate over time is just as essential as building it. You’ll have to decide what to add, resolve conflicts, check queries, and retire context once it goes stale.
Context delivery
Context only helps if it reaches the model at the right moment. Luckily, there are several mechanisms for context delivery:
Retrieval-augmented generation (RAG) fetches relevant documents and definitions when a question is asked
Model Context Protocol (MCP) connects AI agents to external data sources and tools
Context layers and system prompts provide the knowledge and instructions the model needs
However, these mechanisms don't always guarantee a consistent answer. Even when your context layer supplies the correct definition, the model can take a different reasoning path or call the wrong tool, delivering different answers to the same question.
This is why you also need to configure and tune an agent harness, specifically an analytics harness. It’s the system wrapped around the model, tying these mechanisms together. The harness decides what context each question receives, which sources take priority, and how updates flow through. And because everything runs through one path, a single correction applies everywhere.
Optimize context window
Every model has a limited context window, and giving it more information doesn't necessarily lead to better answers. In fact, too much context can make things worse. Chroma's context rot study found model performance often declines as input length increases, even on relatively simple tasks.
To keep context lean and efficient, you’ll need to use techniques like dynamic context compression and token truncation, which decide what gets cut and what to do when context exceeds the budget. Semantic caching also helps on the efficiency side, reusing answers for similar questions.
Although your analytics harness and context layer can automate much of this, you still need to decide how they should work. As a context engineer, you'll determine which information matters for different types of questions, how much context can safely be compressed, and when cached results need to be refreshed rather than reused.

Who can become an AI Context Engineer?
If you already work with data, you’re closer to becoming an AI context engineer than you might think. Many of the skills the role demands are already part of your day-to-day work as an analyst or engineer. Find your starting point below, along with the skills to add:
Data and BI analysts
As an analyst, you already understand how the business works. You know which metrics matter, how different teams define them, and why a number looking correct in SQL might still be wrong in a business context.
On top of that, you’re probably using AI analytics tools, whether it’s the copilot in your BI platform or a model like Claude or ChatGPT helping you write SQL. This hands-on experience gives you a useful advantage: you've seen where AI performs well and where it struggles with business nuance.
To move into context engineering, start by sharpening your SQL and data modeling skills. From there, learn how semantic layers, retrieval systems, and context management tools work.
Analytics engineers
If you're an analytics engineer, you're already doing some of the foundational work of context engineering.
The next step is thinking beyond the semantic layer. Plenty of business rules and institutional knowledge never make it into code, so start by learning how to capture them. From there, learn how to make sure all of it reaches the model and works as intended. This includes grasping the basics of harness engineering, which controls how context is delivered, and also learning how to evaluate AI-generated answers so you can spot and fix problems.
Data engineers
Data engineers already understand how data moves through a business and where things tend to break. You spend your days building pipelines, connecting source systems, and tracing data lineage, so spotting stale data, duplicate records, or broken pipelines comes naturally. You're also familiar with much of the infrastructure behind AI systems, from APIs and orchestration tools to emerging standards like MCP.
Where you’ll want to invest is learning the business side of the data. Start with RAG and keep exploring how vector databases, metadata filtering, and MCP help agents find and access the right information. Also, spend time learning how different teams define their metrics and why, so you’ll be able to turn what you know into context AI can use.
Essential tools for AI context engineers
Configuring, maintaining, and optimizing context takes more than one tool, and your stack will keep changing as models and frameworks evolve. The basics stay the same, though. Here are the essential tools you need to know:
Orchestration frameworks
Orchestration frameworks control how AI applications retrieve context, call tools, and chain steps together. LangChain, LlamaIndex, and LangGraph are common choices for building and managing these workflows, while Airflow and Dagster schedule the data pipelines. Understanding both sides helps you see when context gets loaded, and why it sometimes arrives stale or incomplete.
MCP and connectors
MCP has become a common standard for connecting AI tools to data sources, business systems, and context layers. Focus on knowing how MCP servers expose data and definitions to models, and how connectors link AI to your warehouse, data catalog, and apps like Salesforce or Slack. When AI misses context it ideally should have, the problem often lies in how the connector passed the information along. Fix it there, and every tool pulling from the same source gets the right answer.
Data modeling tools
Data modeling tools turn raw tables into business logic AI can rely on. dbt is one of the most widely used tools for transformations, while semantic layers like the dbt Semantic Layer, Cube, and LookML in Looker let you define metrics once and reuse them everywhere. Fluency in any one of these gives you a good head start.
Version control tools
Context changes constantly, so it’s best to manage it like code. Hands-on practice with tools like GitHub and GitLab allows you to track every change to definitions and rules, review updates before they go live, and roll back mistakes. So when something breaks, you can simply check the version history and see exactly what changed and who changed it.
Metadata management
Before AI can choose the right source, it needs to know what data exists and which sources can be trusted. Metadata management tools track this information, including who owns each dataset, where it came from, and whether it’s certified. Atlan, Alation, Collibra, DataHub, and Unity Catalog in Databricks are common options.
Get familiar with using them to keep the most important signals current, especially certification, ownership, lineage, and descriptions, since AI relies on these to pick trusted sources over stale ones.
Be the reason your company trusts its AI
Most AI analytics tools rely on shallow context. They use MCP to pull bits and pieces of information from across data sources, then count on frontier models to overcompensate for the gap. Either way, accuracy tanks.
Accuracy starts with better context, built from the ground up by someone who knows how your business works. That’s why we created the AI Context Engineer certification, the first course of its kind for data teams. You’ll learn what AI context is, why it decides whether people can trust your answers, and how to set it up using proven context management methods. Hands-on training runs through the whole course.
Register now to be among the first certified AI context engineers.
