An AI agent can write a perfectly reasonable renewal recommendation using the wrong champion, an outdated business goal, and last quarter’s usage trend.
The writing can be excellent. The recommendation can still be wrong.
That’s the tension as agents move deeper into customer success. Generating an answer and being authorized to act on it are different things.
My take: automate the preparation aggressively. Expand decision-making authority carefully.
That doesn’t mean a CSM should approve every task. If your agent needs permission to organize account notes, you haven’t saved much time.
It means we need a clearer boundary between preparing the work and making a commitment.
The most useful agent might be the one that hands off well
Trustpilot offers an example worth studying.
In a Salesforce-published case study, the company describes Sam, its customer support agent. Salesforce reports a 99% reduction in Trustpilot’s support backlog within six months, a 28% overall case-deflection rate, and a 45% decrease in average handling time for escalated cases.
Those are vendor-reported results, not independent proof that another team will get the same outcome.
But the workflow is more interesting to me than the headline.
Sam answers common questions. When a request needs human expertise, it creates a case, summarizes the conversation, and routes it to the appropriate team. The person taking over receives the interaction history and recommended next steps.
That’s the part CS leaders should borrow.
Not “How do we remove the person?”
“How do we make sure the person starts with the context they need?”
For a renewal workflow, that could mean assembling the contract details, relevant usage changes, unresolved issues, and stakeholder gaps before the CSM reviews the account.
The agent does the gathering. The CSM evaluates what it means.
Memory is now an account-management problem
Salesforce also published an update on Agentforce Memory. In its evaluation, an updated memory policy recalled 1.8 times as many explicitly stated facts and 2.2 times as many facts that changed over time as the previous policy.
That’s a specific evaluation result, not evidence that account context is now solved.
The underlying problem should feel familiar to anyone in CS:
A champion leaves, but remains in the account plan.
The customer changes priorities, but the QBR still uses the original goals.
A renewal date moves, but the outreach sequence doesn’t.
An old implementation issue keeps appearing as a current risk.
An agent that remembers more isn’t automatically an agent that knows what’s current.
Before asking AI to draft an account strategy, I’d want a short context check:
What do we know? Where did it come from? When was it last confirmed? What has replaced it?
If two sources disagree, the agent should surface the conflict, not quietly choose whichever fact makes the cleanest narrative.
Define what your agent is allowed to do
“Help manage renewals” is too broad.
A more useful instruction defines what the agent can do at three levels.
1. Automate within clear boundaries
Good candidates include assembling internal account briefs, organizing notes, and flagging changes for review.
These tasks still need monitoring, access controls, and quality checks. But requiring approval for every internal summary defeats much of the point.
2. Prepare a recommendation for review
Think renewal-risk hypotheses, expansion opportunities, account plans, and proposed customer communications.
Here, I want the agent to show its evidence, identify missing information, and explain what would change its recommendation.
The output should make the CSM’s judgment easier—not disguise a guess as a conclusion.
Pricing changes. Contract concessions. Roadmap promises. Commercial commitments.
These deserve an explicit approval boundary, with a clear owner.
The boundary isn’t “AI versus human.” It’s the consequence of being wrong.
Some customer-facing automation can be appropriate within narrow, tested policies. A standard documentation answer is not the same decision as offering a discount to save a renewal.
Make the handoff obvious
Maventa’s announcement of its Mauri AI support pilot offers a useful design choice.
AI replies are labeled. Customers can reply in the same thread to reach a human. The expert receives the full conversation, including the AI’s response. Maventa also states that Mauri cannot see individual company data or change settings.
That’s a defined scope and a straightforward handoff.
For CS, I’d apply the same principle internally: when an agent hits missing data, conflicting evidence, or an authority limit, it should hand over enough information for someone to act.
Not “Please review this account.”
Instead: “Here’s what changed, here’s the evidence, here’s what I couldn’t verify, and here’s the decision needed.”
Prompt of the Week: Evidence before renewal risk
Review this account using only the information provided.
Separate verified observations from interpretations. For each material observation, include its source and date. Flag stale information, conflicting facts, and missing context.
Then propose up to three renewal-risk hypotheses. For each, show the supporting evidence, counterevidence, confidence level, and one question or check that could confirm or weaken it.
Do not invent customer intent, assign an unsupported churn probability, or recommend a concession.
Finish with the next action a CSM should consider and any decision requiring approval.
Use an approved tool and only data your organization permits you to share.
Experiment of the Week: Test the reasoning, not the writing
Try that prompt on three accounts you consider at risk and one you consider healthy.
Before running it, write down your own main concern for each account.
Then evaluate:
Did it distinguish current evidence from stale context?
Did it find a useful gap you had missed?
Did it generate a plausible but unsupported explanation?
Did its suggested check change your next action?
Keep the first run internal. Don’t automatically update health scores or trigger customer outreach.
The goal isn’t to see whether AI agrees with you. It’s to see whether it improves the investigation.
The strongest CS agent won’t necessarily be the one that does the most without help.
It will be the one that knows what it can handle, brings current evidence to the decision, and makes the handoff useful when it reaches its limit.
Let the agent prepare the work.
Give it room to act where the boundaries are clear.
Keep accountability attached to the decisions that matter.
Hit reply: What’s one task you’d let an agent run on its own, and one decision you’d keep with your team?
