01 the business
Subscription software company
70 staff, 1,200 customers
Renewal dates live in the billing system, which sales rarely opens. Account managers learn a customer is unhappy when the cancellation arrives. Customers who have outgrown their plan are not offered an upgrade because nobody is watching.
02 before
The problem
Renewal dates live in a spreadsheet or a billing system nobody in sales checks. Account managers find out a customer is unhappy when the cancellation arrives. Customers who have outgrown their plan or reorder often are not offered more because nobody spots the pattern.
- Renewal conversations start a week before the end date, if at all.
- Support complaints never reach the account manager.
- Seat limits are hit, and customers work around them instead of upgrading.
03 the build
What the AI does
- Step 1: Reads contract end dates and billing history from Stripe, Chargebee or your accounting system.
- Step 2: Combines usage or order data with recent support tickets in Intercom or Zendesk to judge account health.
- Step 3: Spots signals worth acting on, such as falling usage, repeated complaints, seat limits reached or larger repeat orders.
- Step 4: Creates a task on the account in your CRM with the reason, the evidence and a suggested approach.
- Step 5: Drafts a check-in or offer email for the account manager to edit and send.
paused: waiting for a person to approve
How it was put together
- Joined contract dates from billing with product usage and support tickets for each account.
- Defined a small set of signals with the customer success lead, such as usage down a third in a month.
- Created CRM tasks with the reason, the evidence and a suggested approach.
- Drafted check-in and offer emails for the account manager to edit.
04 human in the loop
Where a person stays in charge
The account manager decides whether to act and writes or approves every message to the customer. Discounts and plan changes still follow your normal sign-off rules.
- The account manager decides whether to act and approves every message.
- Discounts and plan changes follow the normal sign-off rules.
- Signals and thresholds are reviewed monthly against what actually happened.
05 rollout
From kickoff to live in 5–6 weeks
- Weeks 1–2Connect billing, usage and support data, and agree the first signals.
- Weeks 3–4Back-test the signals on last year's churned and expanded accounts.
- Weeks 5–6Go live with tasks and drafts; review hit rates monthly.
06 after
What changes
Renewals get attention well before the end date, and at-risk customers hear from you before they decide to leave. Expansion conversations happen because the timing is right, not because someone happened to remember.
Assume 1,200 customers with an average annual value of $6,000 and churn of 12% a year.
1,200 × 12% = 144 lost customers, or $864,000 of annual revenue. Saving one in ten of them keeps about $86,400 a year.
What to measure
- Days of notice before renewal when an account is first flagged.
- Gross revenue retention.
- Expansion revenue from flagged accounts.
- Share of flags the account manager judged useful.
07 evidence
What the research says
About 75% of generative AI's potential value sits in four areas: customer operations, marketing and sales, software engineering, and R&D.
A survey of 4,050 sales professionals found the average seller spends about 40% of their time actually selling, and 51% of sales leaders using AI say disconnected systems are slowing their AI work.
Salesforce, State of Sales report, 2026 edition (opens in a new tab) · 2026
08 lessons
Mistakes to avoid
- Alerting on everything. Too many tasks and people ignore them all.
- Automating the customer message. Renewals are relationships; a person should write or approve every word.
- Never checking the signals. Compare flags with real outcomes every month.
09 questions
Common questions
- Does it need product usage data?
- It helps, but it is not required. Ecommerce businesses use order frequency instead, and service businesses use ticket and invoice patterns.
- Is this a machine learning model?
- It starts with clear rules people can read. A model can come later if there is enough history to train one well.