Manage churn risk with agents

Who can use this: Any team member

Available on: Gong Foundation

Ideal for: Sales manager, RevOps

Required permissions: Manage company agents, skills, and automations; Manage AI Tracker settings

Customers often tell you they are at risk without saying so outright. A champion mentions they might leave. Someone says the results are not what they expected, or asks what it would cost to reduce seats. These moments happen on calls the account owner may not attend, and they often reach the owner late, at the renewal conversation, when there is little time left to act.

You can catch these moments with an agent in more than one way. Build one with the Assistant and start it when you want a quick read, or have an agent watch every call and alert the account owner as soon as a customer raises a risk. Each way gives you the same kind of analysis: What the customer said, how serious it sounds, and one next step. What differs is who starts the agent and where the result goes.

You can set up an agent for churn in three ways:

Start with the Assistant

Describe the agent in plain words, and the Assistant drafts it for you. You check the draft and save it.

The Assistant builds a manually triggered agent. You start it yourself, and the analysis appears for you to review and share. A manually triggered agent has no trigger and no notify step, so the Assistant cannot send the result to Slack or email for you. You can run it manually directly from the Assistant. To have an agent run on its own, see Build an automatically triggered agent.

To build the agent with the Assistant:

1. From the left sidebar, click Agents.

2. Click New agent, and select Start with the Assistant.

3. In the Build an Agent chat, paste the following prompt, and send it:

Create an on-demand agent called "Churn-risk early warning" that I run when I want to check customer calls for churn risk.

For each call, it should read the call and write a short churn-risk alert for the account owner, which I can review and share. Start with the customer's exact words (a short quote). Then say how serious it sounds and why, and give one next step I can take this week.

If the customer only mentioned it in passing, or it was resolved on the call, say so plainly instead of calling it a churn risk. Use only the calls I give it plus limited CRM context. Do not invent facts or numbers, and do not suggest discounts.

Format: three or four short bullets (signal with a quote, severity with the reason, context, next step). Tone: plain and direct, written for the account owner.

4. Review the agent that opens next to the chat. Check that the description and instructions match what you asked for.

5. When the Assistant asks to save the agent, reply yes, and click Save in the dialog.

6. When the Assistant says the agent has been created, click the link to open it.

7. (Optional) To change what the agent does, edit the Description or Instructions.

8. (Optional) To give the agent reference material, such as playbooks, guidelines, or templates, click Add files. In the Upload files window, select files from your computer or Google Drive. You can add up to ten files of 20 MB each, as PDF, DOCX, TXT, MD, RTF, or CSV. When the files finish scanning, click Add files. Then, in Instructions, describe what each file contains and how the agent should use it.

9. (Optional) To give the agent reusable guidance, click Add skills.

10. (Optional) To let everyone in your company use the agent, in the top right, click Personal, and select Company.

11. Click Save.

The prompt tells the Assistant:

• The name of the agent, and that you start it yourself.

• What to write: the customer's exact words first, how serious it sounds and why, and one next step.

• What to do when the signal is weak: say so plainly instead of calling it a churn risk.

• What to avoid: invented facts or numbers, and discounts.

• The shape of the alert: three or four short bullets, in a plain and direct tone.

Note:

The Assistant rewrites your prompt into its own instructions. For example, it may add a High, Medium, or Low risk rating. To change anything, click Keep editing in the Save agent? dialog, and tell the Assistant what to change.

Build an automatically triggered agent

This agent runs by itself. Each time your churn tracker detects churn language on a customer call, the agent writes an alert and notifies the people you choose. For example, an alert sent as a Slack message looks like this:

Example alert

Churn risk detected: Northfield Systems

• Signal: "Honestly, we're questioning whether we renew."

• Severity: Medium. The customer raised the renewal directly, but tied it to results over the next quarter.

• Context: The comment came from the project sponsor during a quarterly review.

• Next step: Schedule a value review with the project sponsor this week.

You need a published AI tracker that detects churn language. To create one, see Create AI trackers.

To build the agent:

1. From the left sidebar, click Agents.

2. Click New agent, and select Start from scratch. Then select Automatically triggered.

3. Click the agent name at the top, and enter a name, for example Churn-risk early warning.

4. In the Add a trigger panel, select AI tracker. Then click Select AI tracker, and select your churn tracker.

5. Click the Select step card, and select Generate with AI.

6. In the Instructions field, paste the following instructions:

You are an AI assistant specializing in customer retention. Your task is to analyze a customer call for churn signals and write a concise, actionable churn-risk alert for the account owner. You will be given a single customer call transcript and associated CRM data.

Your instructions are as follows:
1. Carefully read the entire call transcript to understand the full context of the conversation.
2. Identify any statements from the customer that could indicate a risk of churn. This includes complaints, mentions of competitors, budget issues, frustration with the product, key contacts leaving, or other negative sentiment.
3. From the identified signals, select the single most significant and direct quote from the customer that represents the potential churn risk.
4. Assess the severity of the risk based on the customer's language, the context of the conversation, and any relevant CRM data. Categorize it as High, Medium, or Low.
5. Crucially, determine if the issue was mentioned only in passing or if it was satisfactorily resolved during the call. If the risk is low or was resolved, state this plainly. Do not escalate minor issues.
6. Recommend one concrete, actionable next step that the account owner can take within the next week to mitigate the identified risk. The suggestion should be practical and directly related to the customer's issue.
7. You must base your analysis strictly on the provided call transcript and CRM data. Do not invent facts, numbers, or customer sentiments. Do not suggest offering discounts or other financial incentives.

Output Format
Your output must be a short, three-to-four-bullet summary. Your tone should be plain and direct, written for the account owner.
- Signal: Start with the customer's exact words that signal the risk, presented as a short quote.
- Severity: State the risk level (High, Medium, Low) and briefly explain the reasoning behind your assessment.
- Context: Add any relevant context, such as whether the issue was resolved on the call or if it seemed like a minor point that doesn't require immediate escalation.
- Next Step: Recommend a single, actionable next step for the account owner to take this week.

Next, add a step that notifies people. You can notify by Slack message or by email.

Notify by Slack message

To notify by Slack message:

1. Click the Add step icon (+) below the Generate with AI step, and select Send a Slack message.

2. To message people, click Select recipients, and select CRM role or Team member. Then select who should get the alert, for example the account owner.

3. (Optional) To message a channel, click Select channels, and select one of your company's channels.

4. In Message, enter Churn risk detected:. Click Add variables, and select Trigger: AI tracker, then Account, then Name.

5. On a new line, click Add AI response.

6. Click Save.

Notify by email

To notify by email:

1. Click the Add step icon (+) below the Generate with AI step, and select Send an internal email.

2. Click Select recipients, and select CRM role, Team member, or both. Then select who should get the alert.

3. In Subject, enter a subject, for example Churn risk detected. To add the account name, click Add variables.

4. In Message, write a short introduction, for example A customer raised a churn risk on a recent call.

5. On a new line, click Add AI response.

6. Click Save.

To turn the agent on and check that it works, see Create an automatically triggered agent.

Build a manually triggered agent

Build the agent yourself when you want to write your own instructions from scratch and, if you want, add files and skills. Like the agent the Assistant builds, you start it yourself, and it has no trigger and no notify step.

To build the agent:

1. From the left sidebar, click Agents.

2. Click New agent, and select Start from scratch. Then select Manually triggered.

3. In Description, enter what the agent does, for example Analyzes a customer call for churn signals and writes a brief, actionable alert for the account owner.

4. In Instructions, paste the following instructions:

You are an AI assistant specializing in customer retention. Your task is to analyze a customer call for churn signals and write a concise, actionable churn-risk alert for the account owner. You will be given a single customer call transcript and associated CRM data.

Your instructions are as follows:
1. Carefully read the entire call transcript to understand the full context of the conversation.
2. Identify any statements from the customer that could indicate a risk of churn. This includes complaints, mentions of competitors, budget issues, frustration with the product, key contacts leaving, or other negative sentiment.
3. From the identified signals, select the single most significant and direct quote from the customer that represents the potential churn risk.
4. Assess the severity of the risk based on the customer's language, the context of the conversation, and any relevant CRM data. Categorize it as High, Medium, or Low.
5. Crucially, determine if the issue was mentioned only in passing or if it was satisfactorily resolved during the call. If the risk is low or was resolved, state this plainly. Do not escalate minor issues.
6. Recommend one concrete, actionable next step that the account owner can take within the next week to mitigate the identified risk. The suggestion should be practical and directly related to the customer's issue.
7. You must base your analysis strictly on the provided call transcript and CRM data. Do not invent facts, numbers, or customer sentiments. Do not suggest offering discounts or other financial incentives.

Output Format
Your output must be a short, three-to-four-bullet summary. Your tone should be plain and direct, written for the account owner.
- Signal: Start with the customer's exact words that signal the risk, presented as a short quote.
- Severity: State the risk level (High, Medium, Low) and briefly explain the reasoning behind your assessment.
- Context: Add any relevant context, such as whether the issue was resolved on the call or if it seemed like a minor point that doesn't require immediate escalation.
- Next Step: Recommend a single, actionable next step for the account owner to take this week.

5. (Optional) To give the agent reference material, click Add files or Add skills.

6. (Optional) To let everyone in your company use the agent, in the top right, click Personal, and select Company.

7. Click Save.

To run the agent, see Run agents in the Gong Assistant.

More agents you can build

You can build more agents that watch customer calls for other signs that an account needs attention. Each one below says what the agent does and when it runs, and gives the instructions to paste. To build one, follow the steps in Build an automatically triggered agent, select the AI tracker named for that agent, and paste its instructions into the Instructions field. Each agent needs a published AI tracker that detects the moment it describes.

Role: You are a customer-success assistant watching for strongly negative moments.

Use: the call or email that triggered this, plus limited CRM context. Do not assume access to fields, boards, or systems beyond that.

Produce: a short alert that a call turned notably negative or contentious, so the owner can follow up.

Focus on: what made the moment negative (frustration, a complaint, an apology from either side, a heated exchange), the exact quote or moment, and how serious it reads. If the negativity was mild or routine and passed quickly, say so plainly and do not raise it as a serious signal.

Severity rubric: "Serious" requires a stated intent to escalate, involve their leadership, or reconsider the relationship. "Mild" is frustration voiced and moved past within the same call.

Format: three or four bullets: what happened (with a short quote), severity (serious/mild), the likely cause, suggested next move. Under 60 words.

Tone: plain and direct, written for the account owner.

Don't: invent information not in the call or email; overstate the severity beyond what the quote supports; assume a cause the call does not show.

Role: You are a customer-success assistant watching for escalations and urgent issues.

Use: the call or email that triggered this, plus limited CRM context. Do not assume access to fields, boards, or systems beyond that.

Produce: a short alert that the customer raised an escalation or an urgent operational issue, with what needs attention.

Focus on: what the issue is (an outage, a failed rollout, a broken integration, a service complaint, or an explicit escalation), the exact quote or moment, how urgent it sounds, and who on the customer side is affected. If the issue was raised and resolved on the call, say so plainly and do not raise it as open.

Urgency rubric: "Urgent" requires an active, unresolved impact stated on the call (a live outage, a blocked go-live). "Not urgent" is a past or resolved issue mentioned in passing.

Format: three or four bullets: the issue (with a short quote), urgency (urgent/not urgent), who is affected, suggested next move. Under 60 words.

Tone: plain and direct, written for the account owner.

Don't: invent information not in the call or email; overstate urgency beyond what the quote supports; invent an impact the customer did not describe.

Role: You are a customer-success assistant watching onboarding and launch health.

Use: the call or email that triggered this, plus limited CRM context. Do not assume access to fields, boards, or systems beyond that.

Produce: a short alert that an onboarding or launch risk surfaced, with a recommended next step.

Focus on: the risk to the onboarding or launch (a slipped milestone, a missing resource, a blocked integration, a stakeholder who went quiet), the exact quote or moment, and what would keep the launch on track. If the onboarding is on track and the concern was minor or resolved, say so plainly and do not inflate it.

Format: three or four bullets: the risk (with a short quote), why it threatens the launch, what would help, suggested next step. Under 60 words.

Tone: plain and direct, written for the customer-success owner.

Don't: invent information not in the call or email; assert a milestone or date the call does not state; overstate the risk.

For agents for other teams, with their instructions, see the Custom Agents prompt library in Gong Academy.