AI Agent for Customer Retention — Built for Datarobot
For customer success teams whose data science team already built a churn model in DataRobot: an agent that scores every account each week and gives the owning CSM a save plan built from the model's own reasons.
412 accounts scored. 17 over the threshold. First brief:
How do you turn DataRobot churn predictions into retention actions?
Keep the churn model where it is, deployed in DataRobot, and let an agent do the part after the score.
- 01
Trigger fires
Each week the agent pulls active accounts and their usage, support and billing fields from your CRM, sends them to the deployment for scoring, and reads back the churn probability and the prediction explanations.
- 02
Send each active account's fields to the churn deployment
For accounts that cross your threshold, it writes the score to the CRM record and emails the account's CSM a brief: why the model flagged it in plain words, the save play your playbook assigns to that reason, and a drafted outreach email.
- 03
Write the score and top reasons to HubSpot
The agent doesn't train or change the model.
- 04
Match each flagged account to a playbook save play
- 05
You approve
Anything under your confidence bar waits for a human.
How you tell it what to do
Built in plain English.
You write the rule the way you'd describe it to a teammate. The agent reads the rule, breaks it into the actions it'll take, and confirms the apps it'll touch — before it does anything.
- 1Send each active account's fields to the churn deployment
- 2Write the score and top reasons to HubSpot
- 3Match each flagged account to a playbook save play
- 4Email the CSM a brief with a drafted outreach
How it connects
Connect Datarobot. The agent does the rest.
Claude and ChatGPT are already running on our side. You connect Datarobot with one click, and customer retention runs inside it.
Claude and ChatGPT run on our keys. Nothing for you to configure.
- DatarobotConnect
- HubSpotConnect
- GmailConnect
Runs on your data, in your apps.
Nothing to deploy. Nothing to maintain.
Actions
What Datarobot + DeskFerry can do
Real Datarobot actions your AI agent can perform automatically — no manual work required.
Generate text content
Send prompts to Datarobot and receive generated text for emails, summaries, product descriptions, or creative content.
Classify and categorize inputs
Pass unstructured data to Datarobot for classification into predefined categories, labels, or intent types.
Extract structured data
Use Datarobot to parse documents, emails, or images and extract structured fields like names, dates, and amounts.
Analyze sentiment and tone
Send customer messages or reviews to Datarobot for sentiment scoring and emotional tone analysis.
Generate embeddings
Convert text into vector embeddings using Datarobot for semantic search, clustering, and recommendation systems.
Summarize documents
Pass long-form content to Datarobot and receive concise summaries that capture the key points and decisions.
Moderate content
Screen user-generated text and images through Datarobot to detect policy violations, toxicity, or inappropriate material.
Build conversational agents
Connect Datarobot to messaging channels to power AI chatbots that handle FAQs, triage requests, and collect information.
DataRobot flags drift on 'tickets_open_30d' since the helpdesk field was renamed Sep 9. Briefs went out; check the feature before next Monday.
Playbook suggests a renewal discount for 'budget cut' reasons. Discounts need the CS lead's approval before the CSM offers one.
Account is 19 days old with no usage history. The model returned 0.88 on mostly empty fields; not sent to the CSM.
Human in the loop
Approve before it sends.
Every draft lands in a review queue. You approve, edit, or reject — the agent never acts on its own unless you explicitly turn that on for a workflow you trust.
Governance
Every action, with the reasoning attached.
Each step the agent takes is logged with what it did, why it did it, and which app it touched. Audit-ready, so security and compliance can sign off without backfilling.
- Agent8:22 AM
Emailed Rafael the Harbor & Pine brief and saved a draft outreach in his Gmail.
- Agent8:21 AM
Updated Harbor & Pine in HubSpot: churn_score 0.74, top reasons, playbook 'champion gone quiet'.
- Agent8:15 AM
Held the Orrin Labs score from the CSM.
Reason: The account has 19 days of history and the model's top reasons are all missing-value features, so the high score says more about the data than the customer.
- Agent8:09 AM
Scored 412 active accounts against the churn deployment; 17 above 0.6.
- Agent8:00 AM
Pulled usage, ticket and renewal fields for 412 active accounts from HubSpot.
How it works
Get started in three steps
Step 01
Connect DataRobot and your CRM
Add a DataRobot API key with access to the churn deployment, and connect HubSpot or Salesforce so the agent can read account fields and write the score back.
Step 02
Write the playbook as rules
Set the threshold, map the model's common reasons to save plays, and say which plays, like discounts or exec outreach, need a lead's sign-off.
Step 03
Review what the model can't vouch for
Briefs go straight to CSMs. Drift warnings, brand-new accounts and plays that need approval wait in the queue with the reason shown.
Start automating Customer Retention for Datarobot
7-day free trial. Works with the tools you already use.
FAQ
Frequently asked questions
Does the agent build or retrain our churn model?
No. Your data science team builds, validates and deploys the model in DataRobot. The agent only sends account data to a deployment you name and reads back the prediction and its explanations. Retraining, champion and challenger models, and feature changes all stay in DataRobot with the people who own them.
What does the CSM actually receive?
A short email per flagged account: the new score and last week's, the top reasons from the prediction explanations rewritten in plain words, the save play your playbook maps to those reasons, and a drafted outreach email waiting in their drafts. The same score and reasons are written to the CRM record.
What if the model's inputs change underneath it?
If DataRobot reports data drift on the deployment, or the fields the agent sends start arriving empty, it stops sending briefs and holds that week's run for a person. CSMs acting on scores from a broken tracking pipeline is worse than a week with no briefs.
Can it offer a renewal discount on its own?
Only if your rules allow it, and most teams don't. The usual setup is that the agent proposes the play and drafts the email, while anything with commercial terms, such as a renewal discount or contract change, goes to the CS lead for approval first.
Does customer data leave our systems?
Account fields travel from your CRM to your own DataRobot deployment for scoring, and the result comes back to the CRM. The agent works inside the systems you connect, and every read, score and email is recorded in the audit log with the account it concerned.
Explore more