In customer success, an AI agent is software that reads an account’s data from several systems, decides what needs attention, takes one defined action and leaves a record a person can review. The three examples below show that sequence on real kinds of accounts, followed by what a human checks each week and where the agent should stop and ask.
How an Agent Differs From the Automation Your Team Already Runs
Your team probably has workflow rules today. When a health score drops below a threshold, create a task. When a ticket sits for three days, ping the CSM. Those rules fire on one field and do one thing, and CSMs learn to ignore them within a month because a rule can’t tell a quiet account from a troubled one.
An agent reads product usage, support tickets, call transcripts and CRM notes together instead of one field at a time, and it decides which signal matters for the account in front of it. Then it acts within limits you set and writes down what it did and why.
The wider change in how CS teams work is covered in the shift to AI-powered customer success. The rest of this post stays on what one agent does on one account.
Account Context Decides What the Agent Can Do
Before any example works, the agent needs the account’s data in one place. An agent that only sees CRM fields can only comment on CRM fields.
Carl Carell, Co-Founder and CRO of GetAccept, described how his team set this up on the Across the Funnel Podcast. Transcripts and account data go into one system of record, and the agentic workflows run on top of it:
“We feed all the transcripts and all the data into HubSpot. Everything is in that one system of record as kind of a “data lake” of everything. We use Claude integrated to HubSpot Breeze instead of using their internal to then run things. And then we have n8n on the other end to kind of do some of the agentic workflows.”
That setup matters more than the choice of model. If usage lives in one tool, conversation history in another and the renewal date in a spreadsheet, the agent will write confident summaries that miss half the account. Stop treating customer context like a search problem explains why a CSM hunting through tabs is the same failure. The examples below assume the account data is already joined up.
Example 1: A Usage Drop 70 Days Before Renewal
Picture a mid-market account renewing in 70 days. Weekly active users on its core workflow have fallen for three weeks, and two of the five admins who set the account up haven’t logged in for six weeks.
Here is what the agent does, in order:
- Notices. It sees the decline against the account’s own 90-day baseline and checks that no holiday or product outage explains it.
- Gathers context. It pulls open support tickets (none), the last call transcript (the sponsor mentioned a reorganization) and the contract terms (annual, auto-renew).
- Drafts. It writes a short brief for the CSM with the usage chart, the reorganization note and a suggested question for the sponsor.
- Waits. It does not email the customer.
The CSM reads the brief in two minutes, edits the question and sends it. Without the agent, the same decline usually surfaces at the renewal call, which is the pattern described in the six weeks before your NDR starts to decline. The agent doesn’t need a perfect health score to do this, because it compares the account with its own history.
Example 2: A Champion Who Stops Replying
The second account is in month five of an annual contract. The champion who bought the product answered every email until six weeks ago. The last three messages from the CSM got no reply, and a new VP appears on a calendar invite for an unrelated meeting.
The agent can only connect those facts because email, calendar and product data sit together, which is the problem breaking down data silos for a 360-degree customer view addresses. It flags the champion as disengaging, finds two other contacts who are active in the product and drafts an introduction to one of them for the CSM to approve. It also updates the stakeholder map, so the next person who picks up the account sees the change without reading six months of email.
The decision it leaves to a person is how to approach the new VP, because that depends on internal politics the agent can’t see.
Example 3: An Integration That Stalls in Week Two
The third account is in onboarding. Kickoff was 14 days ago, and the integration with the customer’s billing system is half configured. No ticket is open because the admin is waiting on their own IT team.
The agent sees that the setup step hasn’t moved in nine days and compares it with how long the same step took on similar accounts. It sends the admin a message with the exact permission their IT team needs to grant. If the step is still stuck a week later, it flags the onboarding lead and the CSM with a note on the delay.
This is the gap between a finished kickoff and real use, the one described in from first login to first value and the gap between customer onboarding and customer activation. A customer waiting on their own IT team seldom opens a ticket, so the account looks fine in the CRM while value hasn’t started.
What a Human Reviews Every Week
Each example ends with a record, and someone has to read it. Lindsay Liu, Co-Founder and CEO of Super, runs a voice AI platform for property management, so her agents handle live phone calls instead of renewals. The review problem is the same, and she explained how her team structures it on the Across the Funnel Podcast:
“We are structuring a summary of each conversation, so you get a kind of up to three sentence summary of what it was about, and then there’s bullet points for was it resolved? Yes or no, what was the next step, right? So did we need to bring a human in the loop? Was there a next step involved? Yes or no, and then what was the sentiment of that conversation? Was it positive, negative or neutral?”
Apply that to an account agent. Every action gets a short summary, a yes or no on whether it was resolved, the next step and whether a person was needed. A weekly review then asks:
- Which actions did a CSM change before sending, and why?
- Which flags led nowhere, and what made them look serious?
- Which accounts had no agent activity at all, and should they have?
- How many actions were handled without a person, and is that number rising for good reasons?
The last question is the one to watch over time. A rising share of unreviewed actions is only good news if the changed-before-sending count stays low. Flags that keep leading nowhere usually point to a score nobody trusts, and how to build a customer health score that your team actually uses covers that fix.
Where the Agent Should Stop and Ask
Set the stopping points before the first week. Anything involving price, discounts or contract terms goes to a person. So does any message to an executive sponsor, any account marked as escalated and any case where two sources disagree, such as strong usage and a cancelled QBR.
A short list of hard stops is easier to audit than a long list of permissions. It also tells the CSM exactly which decisions are still theirs, and who benefits more from AI, CS or account management shows how that line differs between the two roles.
The record the agent writes and the account data it reads from are the same memory. Platforms such as Hyperengage build that memory by joining CRM, product and conversation data, so the weekly review and the agent’s next decision start from one source.
Conclusion
An agent earns trust the way a new CSM does, one account at a time with someone reading its notes. Start with the usage-drop example, because the data already exists and the cost of a wrong draft is a rewritten email. The first agent worth running is the one whose work a CSM can read in two minutes and correct in one.
