01 the business
Bookkeeping and payroll firm
65 staff, 900 business clients
Clients rarely announce they are leaving. The signs are scattered: a sharp email, a repeated complaint, a question about notice periods. The account owner usually hears about the problem when the termination letter arrives.
02 before
The problem
Customers rarely announce they are about to leave. The signs are scattered across support tickets, emails and chats: a sharp tone, a repeat complaint, a question about notice periods. By the time the cancellation arrives, the account owner is hearing about the problem for the first time.
- Risk signals sit in tickets the account owner never sees.
- Repeated issues are handled one at a time.
- Exit interviews reveal problems that were visible months earlier.
03 the build
What the AI does
- Step 1: Reads new tickets, chats and emails in Zendesk, Intercom and Gmail as they arrive.
- Step 2: Scores the tone of each conversation and picks out specific risk signals, such as mentions of competitors, cancelling, contracts or repeated unresolved issues.
- Step 3: Combines those signals with account details from HubSpot, Salesforce or Stripe, such as plan size, renewal date and recent payment problems.
- Step 4: Sends a short alert to the account owner in Slack with the reason, the quotes behind it and a suggested next step.
- Step 5: Adds a weekly roll-up of at-risk accounts and common complaint themes for the support and account leads.
paused: waiting for a person to approve
How it was put together
- Scored the tone of each new ticket, chat and email.
- Picked out specific signals such as mentions of competitors, notice periods or repeat issues.
- Combined signals with account size, renewal date and payment problems.
- Sent account owners a short alert with the quotes behind it and a suggested next step.
04 human in the loop
Where a person stays in charge
Alerts only inform. The account owner decides whether to call, offer help or leave it, and no message goes to the customer without a person writing or approving it.
- Alerts inform only; the account owner decides what to do.
- No message is sent to a client without a person writing or approving it.
- Thresholds are tuned monthly against real cancellations.
05 rollout
From kickoff to live in 4 weeks
- Weeks 1–2Connect inbox, help desk and billing; agree the first signals.
- Week 3Back-test on last year's lost clients.
- Week 4Go live with alerts and a weekly roll-up for leads.
06 after
What changes
Account owners hear about unhappy customers while the relationship can still be saved. Leaders see which problems are driving frustration instead of finding out from cancellation emails.
Assume 900 clients worth $3,000 a year each, and 8% leaving each year.
900 × 8% = 72 clients, or $216,000 a year. Keeping one in five of them protects about $43,000 a year.
What to measure
- Lost clients who were flagged beforehand.
- Days of warning before cancellation.
- Alerts the owner judged useful.
- Client retention rate.
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.
Gartner research put the average cost of poor data quality at $12.9 million a year per organisation.
Gartner, Data quality: why it matters and how to achieve it (opens in a new tab) · 2020
08 lessons
Mistakes to avoid
- Alerting on tone alone. Combine it with account facts.
- Too many alerts. Owners will mute them.
- Treating scores as truth. They are prompts to look, not verdicts.
09 questions
Common questions
- Does it read every email?
- Only the shared support and client inboxes you connect, with the same access rules as your team.
- Can it predict churn exactly?
- No. It surfaces evidence early so a person can judge. The rules improve as you compare alerts with outcomes.