Every customer success platform now has an AI label on it. Most of the time the label means one of two things: a chat box that summarizes a record you were already looking at, or a "health score" that is a weighted average of three usage metrics with a new name. Neither changes what a customer success manager does between 9 and 6.
This post is about the kind that does. We build one, so read it with that in mind. But the test we propose at the end works on any vendor, including us.
The job AI is supposed to take off your plate
A CSM with 60 accounts spends most of the week on four things:
- Reading. Product usage, tickets, email threads, call notes, the CRM, LinkedIn, the news. Per account. Every week. Most of it is unchanged since last week, and you only find out by reading it.
- Deciding. Which account needs attention today, and what kind. Risk or expansion. A call or an email. This week or next quarter.
- Preparing. The agenda, the deck, the "what has happened since we last spoke" paragraph before every meeting and QBR.
- Writing. Follow-ups, check-ins, renewal reminders, the internal update to the account exec.
The honest reason churn gets missed is not that CSMs are careless. It is that the reading step does not scale, so it gets rationed to the accounts that are loud. The quiet account that stopped logging in six weeks ago is the one nobody read.
Useful customer success AI attacks the reading step first, because that is where the hours go and where the misses come from.
What good customer success AI actually does
It reads everything, every day, for every account
Not a sample. Not the accounts you flagged. The whole book: usage, support, email, calls, the CRM and the public web, re-read on a schedule, so the quiet account gets the same attention as the loud one.
The output of this step should not be a dashboard. It should be a short, defensible statement per account: what changed, since when, and why it matters. If you have to open the account to find out what the AI found, it moved the reading problem instead of solving it.
It grades the account against that customer's definition of success
A generic health score ("logins down 20%") is wrong for half your book, because half your book does not log in to get value. The customer who bought you to send a monthly report is healthy if the report went out.
Good AI asks what each customer was trying to achieve when they bought, once, and grades everything after that against it. The grade comes with its reasoning, in words, so a CSM can disagree with it. A number you cannot argue with is a number you cannot defend to your VP either.
It proposes the plan, not the alert
Alerts are cheap. "Usage dropped" is an alert. "Usage dropped in the reporting module after their analyst left in August; the renewal is in 90 days; here is the champion to call, what to say, and the three follow-up steps" is a plan.
The difference matters because an alert adds work and a plan removes it. If a tool's AI output is mostly things for you to look into, it has made your week longer.
It executes the steps a human has approved
Drafting the email. Creating the task in the CRM. Preparing the meeting brief with what changed since the last call. Logging the outcome. These are the parts of the job that are mechanical once the decision is made, and they are the parts AI can do end to end.
The word that matters here is approved. Nobody should want an AI sending unsupervised email to a customer at renewal. The useful shape is: the AI proposes, a human clicks activate, the AI carries it through and reports back.
What it still can't do
We would be lying if we said the CSM role disappears. Three things stay human for the foreseeable future.
Relationships. The AI can tell you the champion is leaving. It cannot be the person the champion trusts enough to tell you first. The hours AI gives back are best spent on exactly this.
Judgment under ambiguity. When the data says one thing and the executive sponsor says another, a person decides. The AI should make that decision faster and better informed, not pretend it away.
Accountability. Someone owns the renewal. A model can hold a plan; it cannot hold a quota.
There is also a class of things AI could do and should not: anything that writes to a customer or to the CRM without a human having seen it first. The failure mode is not that the AI is wrong 2% of the time. It is that when it is wrong, nobody knows which 2%.
The questions to ask before you buy
Run these against any vendor's demo, ours included:
- Show me an account I did not pick. If the AI only has something to say about the accounts the vendor prepared, it is reading on demand, not continuously.
- Why is this account that color? The answer should be a sentence a CSM would say, not a formula. If it is a formula, ask who tuned it for your customers.
- What happens when I click this? Follow one recommendation to the end. Does a task appear in the CRM? Is an email drafted? Or does it open another screen with more things to read?
- What did it send without asking? The right answer is nothing.
- Where does my data go, and what model reads it? You are handing it every email and call transcript your customers ever sent you. The vendor should be able to tell you, per tenant, how that is encrypted and who can see it.
- What does it do on day one with 200 accounts and no playbooks configured? If the answer starts with a six-week implementation and a workshop on your health score model, you are buying a dashboard with an AI label on it.
The one-sentence version
Customer success AI earns its name when it reads the whole book every day, grades each account against what that customer wanted, proposes a plan rather than an alert, and executes the approved steps so the CSM's week gets shorter. Everything else is a very good search box.