This article was originally published on aicoderscope.com
72% of data teams now prioritize AI-assisted coding. That number comes from dbt Labs' 2026 State of Analytics Engineering report — and it tracks with what you see on any data engineering Slack or conference panel. But buried three questions later in the same survey: 71% of respondents fear hallucinated or incorrect data reaching stakeholders. Those two numbers aren't a paradox. They're a description of exactly where the tooling is right now — useful enough to use, unreliable enough to distrust at the output boundary.
The problem for data engineers is that the AI coding tool landscape wasn't built for them. Cursor, Copilot, Cline — excellent at TypeScript, Python, Go. SQL is an afterthought in most training pipelines. Jinja-templated SQL inside a dbt project is further out still. And Airflow DAGs, with provider-specific operator parameters and API-version-sensitive syntax, are practically a landmine for any model that trained on a mix of Airflow 1.x and 2.x documentation.
Here's which tools are actually worth using for data work, and what you need to give them to stop generating garbage.
Where AI actually earns its keep
Three tasks where AI shows consistent, measurable payoff in data work:
SQL scaffolding. Window functions, complex CTEs, multi-join aggregations — this is boilerplate with structure, and structure is what language models handle best. A model that's seen a million ROW_NUMBER() OVER (PARTITION BY ... ORDER BY ...) patterns will produce working SQL faster than a human typing it cold. The failure mode is dialect mismatch: Snowflake's QUALIFY, BigQuery's DATE_TRUNC with a different signature, DuckDB's COLUMNS(*) macro. Feed your schema DDL as context and the hit rate improves dramatically.
dbt documentation. Writing description: fields for 40 staging models is exactly the kind of work that eats an afternoon and produces diminishing returns on human attention. AI does this faster and more consistently than a human who's been at it for three hours. dbt Copilot generates column-level docs from the model's SQL and column metadata — it doesn't need to query your warehouse, it works from column names and DDL it can already see.
Test generation. not_null, unique, accepted_values, relationships — these are pattern-based, and AI gets them right reliably. The more interesting case is generating business-logic tests: was revenue for any order_id recorded more than once? do all rows where active = true have a non-null activated_at? This requires more context but produces higher-value output than the boilerplate alternative.
The Airflow problem is a different category
Only 9% of data engineers report being satisfied with DAGs generated by generic AI tools. Astronomer's 2026 State of Airflow survey found 43% cite hallucinations as the primary issue, and 42% report the AI generating outdated API syntax — operators deprecated in Airflow 2.x, or provider-specific parameters that don't exist in the version they're running.
This is a training data problem that no prompting trick fully solves. Airflow has changed significantly across versions, provider packages are independently versioned, and the documentation structure makes it hard for a language model to know which API surface you're working against. A DAG that calls S3Hook.read_key() instead of S3Hook.get_key() won't error at write time — it'll fail at runtime, possibly after writing partial output.
Astronomer's architectural response is Otto, an AI agent built into Astro with access to your instance, deployment history, and warehouse connections. It's the right answer structurally — an agent needs runtime context, not just training data — but it's wrapped inside Astronomer's enterprise pricing, which starts at $0.35/AU-hour for cloud deployments with typical production workloads running $1,500–5,000+/month.
For teams not on Astro, the practical advice: use AI to generate the DAG skeleton (imports, with DAG(...):, task definitions), then hand-verify every operator parameter against your installed version's docs. The AI saves time on structure; you catch it on specifics.
Tool-by-tool breakdown
dbt Copilot + Developer Agent
Best for: Teams already on dbt Cloud who want no-friction AI within the platform.
dbt Copilot generates models, documentation, tests, and semantic layer definitions from natural language inside the dbt Cloud Studio IDE. It works from column names, model SQL, and metadata — it doesn't touch your warehouse row data. Available on Starter ($100/seat/month, up to 5 developer seats) and Enterprise plans. The free Developer plan (single user, no job scheduling) doesn't include Copilot.
The dbt Developer Agent, now in preview for platform customers, is the more capable version. It's multi-file aware — it sees every file a change would touch, which matters in dbt projects where a refactor to a staging model ripples downstream into marts. It shows its reasoning and tool calls as it works, which is useful for calibrating when to trust the output. Two modes: ask-for-approval (default) and auto-edit-files for bulk generation.
Both tools only run inside the dbt Cloud browser IDE. If your team uses VS Code or Cursor with dbt Core locally, you need a different path.
Paradime + DinoAI
Best for: Analytics engineers who want a dbt-native IDE at a lower per-seat cost than dbt Cloud.
Paradime is an IDE built for dbt work — think VS Code with dbt context baked in from the start rather than bolted on via extension. DinoAI is their AI layer: it knows your warehouse schema, model relationships, and column metadata when generating code, not just the open file.
Pricing (verified May 18, 2026): Spark plan at $25/user/month (1M AI credits), Flow at $55/user/month (5M credits). Team-level features include .dinorules (SQL style and convention enforcement across the org) and .dinoprompts (shared prompt library for analytics teams).
The warehouse-context integration is DinoAI's biggest differentiator from general coding tools. When you ask it to write a model joining orders to customers, it knows those tables exist in your warehouse and what columns they have — without pasting the schema each session.
The tradeoff: Paradime is its own IDE. If you're invested in VS Code extensions, keybindings, and configurations, the context switch costs real time.
Altimate AI / dbt Power User
Best for: Teams that want to stay in VS Code and add dbt-specific AI without switching tools.
dbt Power User is an open-source VS Code extension with 100+ AI-backed features: model generation, auto-documentation, test generation, column-level lineage visualization, cost estimation, and project health checks. The underlying AI platform is Altimate.
Pricing (verified May 18, 2026): Community tier is free (200 one-time credits for core VS Code features). Pro is $29/month (200 credits/month, Python package access). Team is $549/month (750 credits/month, SaaS web UI for docs and lineage, "defer to prod" feature). Enterprise is custom.
The separate Altimate Code open-source harness goes further: 100+ deterministic tools for dbt, SQL, and cloud warehouses. SQL validation against your actual schema runs in 2ms. Column-level lineage traces through CTEs deterministically. It supports 10 warehouses (Snowflake, BigQuery, Databricks, PostgreSQL, Redshift, DuckDB, and more) and any LLM provider (Anthropic, OpenAI, Google, Ollama). Think of it as a set of precise, warehouse-aware tools you can load alongside a general-purpose agent like Cursor or Claude Code — it provides the schema grounding those tools lack natively.
Cursor + dbt-MCP
Best for: Data engineers who already use Cursor for application code and want to extend it into dbt work.
Cursor at $20/month (Pro) isn't purpose-buil
Top comments (0)