What to automate in customer success comes down to four tests. Automate a task when a wrong action can be undone within a day, every input sits in a system software can read, the customer wouldn't expect a person at that moment, and the task happens often enough to earn the build.
Fail any test and a person owns the task. Ask the question task by task, and once per segment. "Automate onboarding" is too big to answer.
Why "automate the boring parts" fails
The common advice is to automate boring work and keep people on interesting work. It sorts tasks by how they feel to the CSM, not by what happens when the machine gets them wrong.
Updating seat counts and pricing on a renewal quote is boring. It is also a commercial commitment, and a wrong number that reaches the customer can't be taken back.
Choosing which onboarding step an account should see next feels like judgment. But a wrong choice costs one ignored email, the inputs sit in your product data and CRM, and it happens every time an account finishes a step. Boring and safe are different things.
Forrester's Shari Srebnick goes further. In an April 2026 post on Gainsight's MCP launch, she writes that "What can we automate?" isn't the right question. One she says advanced companies ask instead: "Where does human judgment create the most value, and how do we architect everything else around those moments?" The tests below answer it with criteria.
The four tests
Ask all four, every time:
Reversible. If the action is wrong, can you undo it within a day at low cost?
Readable. Is every input already in a system, not in a CSM's head or inbox?
No person expected. Would this customer, in this segment, be fine not hearing from a person here?
Frequent. Does the task happen often enough across your book to repay the setup?
Four yes or no tests, ours, asked once per task and per segment.
The one-day line is ours. A nudge that reaches the wrong contact costs an apology and a corrected email tomorrow. A renewal price stays on the record. For the review rules behind that line, see trust operations for post-sale AI agents.
For Readable, name the system and the field behind every input. The renewal date is in the CRM. The fact that the champion is leaving next month may be in one CSM's notes, or nowhere.
For Frequent, count across your book. A quarterly business review across 300 accounts happens about 100 times a month, so it passes. An annual executive sponsor reset with your ten largest customers happens ten times a year, so it doesn't.
Do B2B customers want a person?
B2B customers want to do routine post-sale steps on their own and want a person when they make or check a decision, such as renewing, expanding or escalating. That is our inference from two adjacent Gartner surveys, neither about customer success, and it is what the third test asks about.
So ask whether the customer is doing a step or deciding something. Doing passes. Deciding, or hearing news that changes the deal, fails. Either way, an automated step should never hide the way to reach a person.
Schedule, agent or person: how do you choose?
Choosing between a schedule, an AI agent and a person takes one question per lane. Use a schedule when every account in a segment should get the same step on the same trigger. Use an AI agent when a task passes all four tests and needs a choice no rule wrote. Otherwise, keep a person.
On schedule. A welcome message at signature, or a renewal reminder at a fixed point. Our glossary covers both machine lanes under zero-touch automation.
AI agent decides. Which onboarding step comes next for an account that left the planned path. Who to contact when an account goes quiet.
Person owns it. Software prepares the context, and a person makes the call. The human-in-the-loop AI pattern describes that handoff.
A two-lane view, human or automated, hides a costly mistake: paying for an agent to make a decision a rule could make. Gartner's Anushree Verma put it plainly in a June 2025 release on agentic AI: "Many use cases positioned as agentic today don't require agentic implementations."
The Base AI Customer Growth platform page puts the split in one line: "Regular touches run on schedule. The judgment calls go to your AI teammate."
The Base AI homepage adds a dial inside that lane: "what it does on its own, what it asks first and what it only suggests." A task near the line can start in "asks first" and earn more autonomy once the numbers hold.
How ZoomInfo split one kickoff
ZoomInfo had onboarded every customer with the same 90-day framework. Its customer story says the framework "still works well for mid-market and enterprise customers." For down-market SMB customers, the same time investment didn't fit. ZoomInfo is a Base AI customer, and its portal is "powered by EverAfter by Base AI."
"These customers move fast," says Nate Alcorn, ZoomInfo's Director of Lifecycle Operations. "Their window for value is measured in hours and days, not weeks."
Run the kickoff through the third test and you get two answers. For a down-market SMB team, no person is expected. For an enterprise account with layered teams, one is. Same task, two lanes.
So ZoomInfo moved one segment to a self-service portal built on the same framework, with Salesforce-synced nudges when a customer falls behind. By our lanes, a rule fires those nudges, so ZoomInfo needed no agent. Within 40 days of launch, contract to completed kickoff went from 5-7 days to 1-3 days.
Within 40 days of launch, as ZoomInfo's customer story reports it.
ZoomInfo "kept the same engagement volume" while removing manual scheduling and inbox work, and onboarding managers "now spend that time on strategic accounts." Base AI's onboarding journeys by segment are built on the same idea: keep the framework and change the delivery.
20 CS tasks, sorted by lane
We ran 20 common tasks through the tests, for a book with a scaled segment and a handful of strategic accounts. The scoring is ours, so copy the method, not the verdicts.
Our scoring of 20 common tasks. Your answers will differ by segment.
Three examples from each lane:
On schedule: the welcome message, a renewal reminder, a satisfaction survey after a ticket closes.
AI agent decides: re-engaging a quiet account, routing a health drop to the right owner, assembling the data pack for a business review.
Person owns it: renewal negotiation on a strategic account, a response to an executive escalation, telling a customer about a price increase.
In our scoring, 12 of the 20 run without a person (one starts in "asks first"), and five of those need only a schedule. Two of the eight tasks that stay with a person sit there only because the data is missing, like a champion's departure that lives in CSM notes. Those two are your data roadmap.
Run the tests on your own list
Pull your 20 most frequent CS tasks from your team's calendars and your CS platform. Score each on the four tests, once per segment. Pick the one agent-lane task with the most hours behind it and start there, in "asks first."
Decide task by task, per segment. "Automate onboarding" is too big to answer.
One no keeps it human. A task needs four yeses before a machine takes it.
Use three lanes, not two. On schedule, AI agent decides, or person owns it.
Don't pay an agent for rules. If a rule can pick the step, schedule it.
Frequently asked questions
What should you automate in customer success?
Automate a customer success task when it passes four tests: a wrong action can be undone within a day, every input the decision needs sits in a system software can read, the customer would not expect a person at that moment, and the task happens often enough to repay the build. If the task fails any one test, a person owns it.
Which customer success tasks should stay human?
Keep a person on any task that fails one of the four tests. Typical examples are renewal negotiation on a strategic account, an expansion proposal with new pricing, a reply to an executive escalation, and telling a customer about a price increase or a serious incident. Software can prepare the context, but a person makes the call and has the conversation.
When does a customer success task need an AI agent rather than a schedule?
Use a schedule when every account in a segment should get the same step, fired by a date or a trigger written in advance, such as a renewal reminder. Use an AI agent only when the step needs a choice no rule wrote in advance, such as the next onboarding step for an account that stalled. If a rule can pick the step, an agent adds cost.
Do B2B customers want to talk to a person after the sale?
No large survey of post-sale B2B customers answers this directly. In a Gartner survey of about 645 B2B buyers about their purchase journey, fielded in 2025, 67% preferred a rep-free experience and 69% preferred to validate AI-generated insights with sales reps. Applying that to post-sale is our inference: customers want to do routine steps alone, and want a person when they decide.
Why is "automate the boring parts" a bad rule for customer success?
"Automate the boring parts" is a bad rule for customer success because it sorts tasks by how they feel to the CSM, not by what a wrong action costs. Updating seats and pricing on a renewal quote is boring, but a wrong number that reaches the customer cannot be taken back. Choosing the next onboarding step feels like judgment, yet a wrong choice costs one ignored email.
How do you decide what to automate for your own customer success team?
Pull your team's 20 most frequent CS tasks from calendars and your CS platform. Score each on the four tests (reversible, readable, no person expected, frequent) once per segment, because the same task can pass for SMB customers and fail for enterprise accounts. Start with the agent task that has the most hours behind it, and let it ask before acting until results hold.
Customer experiences that drive NRR.
Building and scaling your advocacy program is easy when you can manage campaigns, rewards, track, and report - all within one AI platform.
