DataRobot

Score records against DataRobot MLOps deployments in real time or as batch prediction jobs. Agents also manage modeling projects, browse the Leaderboard to compare candidate models, and ingest datasets into the AI Catalog.

13 actions API key available

DataRobot website ↗ Platform Documentation ↗ Capability data verified 2026-07-27

The nightly extract of new applications lands in an S3 bucket

Get Deployment pulls the production deployment's model, prediction server details, and health

Service, data drift, and accuracy health are compared against the green baseline

When drift is flagged, the model owner decides whether tonight's file scores against the current deployment or holds for review

Start Batch Prediction Job scores the released file through HTTP intake

Get Batch Prediction Job polls until the state reads COMPLETED

The scored-row count and result download link post to the team channel

What This Integration Enables

Most companies do not have a model problem, they have a distance problem: the model lives in DataRobot and the decisions live everywhere else. This connector closes that distance. FlowRunner agents score records against MLOps deployments in real time through Make Predictions, run large workloads as asynchronous batch prediction jobs, and route the scores into the systems where people act on them. On the modeling side, agents ingest datasets into the AI Catalog from a URL, spin up projects, and read the Leaderboard to compare candidate models on their actual metric scores. The stance worth stating: predictions are operational data, not analytical output. A churn score that surfaces in the CS platform while the account is still saveable is worth more than a better score delivered next quarter. And because scored output drives real decisions, the [human-in-the-loop](/concepts/human-in-the-loop) gate belongs where this connector puts it, between deployment health and the batch job that spends compute and feeds downstream systems.

Without FlowRunner

Models behind a ticket queue The data science team runs scoring jobs by hand when someone asks nicely

Predictions that arrive stale By the time the batch is scored and emailed around, the decisions it fed are already made

Silent model decay Drift accumulates for weeks because nobody checks deployment health until numbers look wrong

With FlowRunner

Scoring as a pipeline step Records score against production deployments in real time or as batch jobs inside the flow

Predictions where decisions happen Scores land in the CRM, the sheet, or the queue while the record is still being worked

Health checked before every run Agents read deployment drift and accuracy status first, and a person rules on anything amber

Use Case Scenarios

Batch scoring the pipeline runs itself

A weekly renewal-risk job stages the account file to [S3](/integrations/s3), then calls Start Batch Prediction Job against the churn deployment with the file's URL as HTTP intake. The agent polls Get Batch Prediction Job until COMPLETED, posts the scored-row count and download link to [Slack](/integrations/slack), and routes high-risk accounts into the CS team's queue in [HubSpot](/integrations/hubspot). What used to be a data scientist's Friday afternoon is a flow that runs while the team sleeps.

Real-time scores inside operational records

An operations analyst maintains a working file in [Google Sheets](/integrations/google-sheets). The agent pulls the new rows, passes them to Make Predictions with pass-through columns so identifiers survive the round trip, and writes the predictions and explanation columns back next to each record. The analyst never exports a CSV, never waits on a ticket, and every row shows why the model scored it that way.

From raw file to Leaderboard comparison

When a refreshed training extract is published, the agent calls Upload Dataset From URL, polls Get Dataset until processing completes, and creates a new modeling project from the dataset with Create Project. Once Autopilot has run, List Models returns the Leaderboard with validation and holdout scores per candidate, and the agent assembles a side-by-side comparison for the model owner. The promotion decision stays with the person accountable for it; the agent just makes sure the comparison is waiting when they arrive.

Human-in-Loop Highlight

Start Batch Prediction Job is the spend that cannot be recalled. Once the job completes and results are written to their destination, downstream systems consume the scores: renewal offers get priced, applications get routed, accounts get flagged. If the deployment was drifting when the job ran, thousands of decisions inherit the drift at once. So the agent reads Get Deployment first, and any deployment whose data drift or accuracy health is not green stops the flow. The model owner sees the health detail and the file waiting to score, and chooses: run against the current model, hold until the deployment is retrained, or score through a different deployment. The compute is cheap compared to the decisions it feeds, and that is exactly why a person releases the job.

Agent processes routinely

Detects exception requiring judgment

Clear match Continues automatically

Ambiguous Routes to human via preferred channel

Human decides

Agent resumes with decision

Agent Capabilities

13 actions

Projects

List Projects Lists modeling projects newest first with target, project type, and stage, with paging and a server-side name filter.
Get Project Retrieves a single project's full detail: target feature and type, stage, partitioning, and metric.
Create Project Creates a modeling project from an AI Catalog dataset and starts data ingestion. Target selection and Autopilot follow inside DataRobot.

Models

List Models Lists a project's Leaderboard with each model's type, blueprint, feature list, sample size, and metric scores. The comparison step behind every promotion decision.
Get Model Retrieves one Leaderboard model's detail, including processing steps and scores, for a deeper look at a candidate.

Deployments

List Deployments Lists MLOps deployments with label, model, prediction environment, and health, searchable by label. How an agent finds the right thing to score against.
Get Deployment Retrieves a deployment's model, target, prediction server URL and DataRobot-Key, and its service, data drift, and accuracy health. The health read this page's human gate depends on.

Predictions

Make Predictions Scores a batch of records in real time and returns per-row predictions, optionally with explanations and pass-through columns. Routes through the serverless path or a dedicated prediction server.
Start Batch Prediction Job Starts an asynchronous batch job for workloads beyond the real-time payload limit, reading input from a file URL and writing results inline or to an HTTP destination.
Get Batch Prediction Job Retrieves a batch job's state, rows processed, and result download links. The polling step between submission and downstream delivery.

Datasets

List Datasets Lists AI Catalog datasets with row and column counts, size, and processing state.
Get Dataset Retrieves one dataset's detail and processing state. Poll it after an ingest before building on the dataset.
Upload Dataset From URL Ingests a file from a publicly accessible URL into the AI Catalog asynchronously. The first step of every training data refresh in this page.

Frequently Asked Questions

What can FlowRunner do with DataRobot?

FlowRunner agents can run List Projects, Get Project, and Create Project in DataRobot, plus 10 more actions.

Does connecting DataRobot to FlowRunner require OAuth?

No. DataRobot connects to FlowRunner with an API key, no OAuth flow required.

Can DataRobot trigger a FlowRunner workflow automatically?

DataRobot doesn't currently expose triggers in FlowRunner. It connects as an action step inside workflows started by another trigger.

Start building with DataRobot

$100 in credits. No card required. Connect in minutes.

Start Building Free Talk to us