Renewal and Expansion Planning When Agents Watch the Account

Renewals fail quietly. The signal was available for months, nobody was watching the right thing, and the first visible event is a procurement email sixty days out. That is a detection problem, and detection is exactly what agents are good at.

What agents are not good at is the conversation that follows. The split between those two is the whole design question.

The four signal classes worth acting on

Most churn dashboards drown teams in signals with no action attached. Four classes carry enough information to justify a play.

1. Usage decay. Active users, core-action frequency and breadth of features falling over a rolling window. The most reliable single class in most software businesses.
2. Relationship decay. Fewer contacts engaged, lower seniority engaged, longer response times. This is the class engagement platforms surface well, and the same graph that flags single-threading in new business flags it in renewals.
3. Champion change. A job change, an internal move or a reorg at the account. High-value, low-volume, and the one class where speed matters most.
4. Value evidence gap. No documented outcome in the last two quarters. Not a customer behaviour, an internal failure, and the one class you fully control.

Anything outside these four is usually noise dressed as insight. Score them, weight them by your own closed-lost history, and route only what clears a threshold.

What the agent owns

• Monitoring all four classes continuously and flagging threshold breaches with the evidence attached.
• Producing the renewal brief: usage trend, support history, open issues, contact map, previous commitments, and the value evidence file.
• Drafting the QBR narrative and the expansion hypothesis, for a human to accept, reject or rewrite.
• Chasing internal follow-through. Tasks created, owners reminded, artefacts collected.
• Preparing the pricing scenario comparison, without sending it.

That is a substantial share of the work and none of it is irreversible. Each item is a draft or a flag, checked before it reaches a customer.

What the human owns

• The renewal conversation itself, including anything about pricing, terms or commitments.
• Interpreting why usage fell, which is almost always a story about the customer's business rather than about your product.
• The expansion ask, which requires standing that a system does not have.
• Any escalation involving an unhappy executive.
• Every irreversible action: contract changes, credits, discounts, termination handling.

The rule is reversibility, and it is the same rule used for production agents generally in the rollback plan requirements. A wrong draft costs an hour. A wrong commitment costs a relationship.

The 120-day renewal plan

1. Day 120. Agent. Generate the renewal brief for every account in the cohort. Score the four signal classes. Rank by risk-weighted value.
2. Day 115. Human. Review the ranking, override where you know something the data does not, and assign owners. Overrides are logged, because the log is how the scoring model gets better.
3. Day 110. Agent. Assemble the value evidence file per account: outcomes delivered, tickets resolved, milestones hit, with dates.
4. Day 100. Human. Where the evidence file is empty, that is the renewal risk. Build a ninety-day value plan before any commercial conversation happens.
5. Day 90. Human. Executive alignment meeting on the top-risk accounts, with the agent-produced brief as the pre-read.
6. Day 75. Agent. Draft the QBR narrative and the expansion hypothesis. Flag any contact-map gaps and any champion change since day 120.
7. Day 60. Human. Run the QBR. Land the value story, test the expansion hypothesis, and confirm the decision path and timeline in writing.
8. Day 45. Human. Commercial proposal, constructed by a person, with agent-prepared scenario comparisons behind it.
9. Day 30. Agent. Track every open commitment and chase internal owners daily. This is the step humans drop and machines never do.
10. Day 15. Human. Close the renewal or escalate. No new information should appear at this stage. If it does, the detection layer failed and that is a post-mortem item.

The measures that keep this honest

Four numbers, with a baseline captured before you deploy any of it.

• Warning lead time. Days between the first flagged signal and the renewal outcome. Longer is better and it is the primary proof the detection layer works.
• Surprise rate. Share of churn events with no flag raised more than sixty days out. This should fall toward zero.
• Brief usage rate. Share of renewals where the human actually used the generated brief. Low usage means the brief is wrong, not that the team is lazy.
• Net revenue retention. The only number the board cares about, and the one that moves last.

Do not report the first three without the fourth eventually appearing. That drift is optimization theater and it will be caught.

Related: Customer success and AI renewal risk · Rollback plan requirements · Signal stack · Skill: planning renewals and expansion