Every month, analysts at plenty of large companies rebuild the same report by hand. They export from three systems, join the tables in Excel or an Alteryx workflow, and fix whatever broke since last month. Ask a chatbot to write the SQL instead and it will happily produce a query that runs on the first try.
Whether that query is correct is a separate question, and it’s the one the more interesting startups in this space are trying to answer.
“Almost Right” Is the Costly Kind of Wrong
A query that throws an error gets fixed. A query that runs and returns plausible numbers gets trusted.
Picture a revenue report joining orders to customers. If one customer record appears twice because of a duplicate in the CRM, the join doubles that customer’s revenue. The total goes up a little. Nobody notices for a quarter, and by then the number has been in a board deck.
AI-generated pipelines make this kind of mistake easier to produce at scale, because the person asking for the pipeline often can’t read the code that comes back. Which is why the startups worth watching are less focused on generating pipelines than on making the result something a human can inspect.
Prophecy and the Visual Pipeline
Prophecy is a good example. Its pitch is that AI agents generate visual workflows and then deploy SQL or Spark pipelines on Databricks, Snowflake, or BigQuery. In February 2026, it launched Prophecy v4, which brought agents based on Claude Code to business users for visual data preparation and analysis.
The visual layer is the important part. The agent writes a first draft, but the analyst sees it as a series of steps, joins, filters, and aggregations, and can check each one against the data. A September 2026 release added a review flow that shows every change an AI agent makes before anyone approves it. That’s exactly where the duplicate-customer problem would get caught.
The company positions itself as a cloud alternative to desktop-based tools such as Alteryx, and it offers an import for existing Alteryx pipelines. For a mid-sized or large company with analysts who already think in Alteryx-style workflows and data sitting in a cloud warehouse, that’s a sensible migration path.
It’s a poor fit for a small startup with one Postgres database and a single data person. The value shows up when many analysts build many pipelines and someone needs governance over what reaches production.
Operations Agents Depend on the Pipes
The other wave of startups works further downstream, on AI for business operations: agents that reconcile invoices, flag expense policy violations, chase overdue approvals, or pull a Monday revenue summary into Slack.
Ramp has added AI agents that review expenses against company policy. Rillet is building an accounting system around automation rather than retrofitting it onto an old ERP. Plenty of Slack-based agents now promise to answer any question about the business by querying connected tools.
They all share a weakness. An agent that summarizes churn is only as reliable as the table it reads, and if that table came from a pipeline nobody reviewed, the agent will report the wrong number confidently and quickly. Companies get more from operations agents once their core data is clean and owned. Deploying them first, hoping they’ll paper over messy data, usually produces faster confusion.
Questions Worth Asking Before a Pilot
Who reviews the agent’s output before it’s used, and is that review logged? Can someone see exactly what changed between two versions of a pipeline or a report?
Where does the work run? Tools that execute inside the company’s existing warehouse keep data under current access controls. Tools that copy data into their own environment need a closer security review.
And what does it cost per run? Agent-heavy products increasingly bill by usage, so a pilot should track spend per workflow from day one.
Trust Is the Bottleneck Now
Generating a pipeline, a query, or a report has become cheap. What still takes time is deciding whether to believe it. The startups pulling ahead seem to understand that, and they’re building review screens and change logs with the same care others put into the prompt box. That’s a less exciting demo, but a more useful product.
