AI can take over the rote, high-volume parts of customer success: renewal reminders, check-in cadences, usage reporting, and bringing a new account owner up to speed on three years of client history. It cannot build the trust that stops a client from cancelling. Telling clients which parts of your work were AI-assisted and which were written by a human is what keeps that trust intact.
I've spent five years running a company of generalist operators who sit inside other people's teams and own this function, and the sorting rule below is the one I keep coming back to. It's simple, and most teams skip it because automating feels like progress and sorting feels like homework.
how We Do: split rote work from judgment work before you automate anything
Before you buy anything, spend one week logging every client-facing action your team takes. Then run it through five steps.
- Write down each recurring task, including the ones nobody has named, like the Slack message you send when a client goes quiet for eleven days.
- Ask what a wrong answer costs. If the cost is rework, it's rote. If the cost is the relationship, it's judgment.
- Automate the rote column first, and only that column, with a person reviewing the output for the first month.
- Assign every judgment item to a named human the client can recognize by voice.
- Re-sort quarterly. Tasks migrate. A conversation that was judgment in month one can become rote by month nine, and the reverse happens after a bad quarter.
| Rote work | Judgment work | |
|---|---|---|
| Typical examples | Renewal date reminders, QBR scheduling, usage report assembly, ticket routing, meeting recaps | Pricing changes, escalations, scope disputes, bad news, expansion conversations |
| Cost of an error | Minor rework | The account |
| Right owner | A tool, checked by a person | A named human the client already knows |
| What AI contributes | Speed and consistency | Preparation, never delivery |
What AI is actually good at in client work
The legitimate wins here are unglamorous, which is why they get overlooked in favour of a chatbot on the pricing page.
- Cadence tracking. Knowing who hasn't been contacted in six weeks, across a book of forty accounts, is a memory problem, and machines are better at memory problems than you are.
- Meeting capture. Notetakers genuinely help a generalist who carries context across several very different clients in a single day.
- Draft assembly. Pulling the usage numbers, the open tickets, and the last three call summaries into one document before a review call saves an hour of prep.
- Transition support. When an account owner leaves, a well-maintained system can hand the next person the full history of what was promised, what broke, and what got fixed.
- Pattern spotting. Flagging that four clients raised the same complaint in a fortnight is something a tool will catch faster than a team comparing notes.
Every one of those sits in the rote column. Each one buys back hours you can spend on the other column.
What it can't do: the pepperoni-stick problem
My colleague Sarah Matthew put the limit better than I've managed to. Describing what a system will hand a new account owner, she said: "it's also not going to remember the story your client told you over drinks that one time about their dog's favorite treat being like pepperoni sticks and then you bringing it up six months later."
That detail is the whole job. A transcript records that a dog was mentioned. It cannot tell you that mentioning the dog again, six months later, unprompted, is the thing that makes a client feel known. The tool captured the data and missed the point of it.
Things that don't survive the transfer to a system:
- Why a client went quiet after a particular call
- Which stakeholder actually decides, regardless of the org chart
- What a client is embarrassed about
- The joke you can make and the joke you can't
- Whether "fine" meant fine
The notetaker tradeoff
The tool is in every meeting now, and it changed the room. When everyone knows the recording is running, people stop listening actively, because the listening has apparently been outsourced. Research on what distinguishes good listeners points the other way: the value is in the two-way exchange happening live, not in the record of it [1].
What we've seen go wrong
Teams start treating the recap as the meeting. Decisions get made from summaries by people who weren't paying attention while the conversation happened, and the nuance that would have prompted a follow-up question is gone by the time anyone reads it.
What to do instead
Keep the notetaker and add a rule: the person who owns the relationship does not read the transcript before writing their own three-line summary from memory. If those two documents disagree, you've learned something about how well you were listening.
Why disclosing AI use builds trust
The common worry is that admitting AI involvement makes the work look cheap, so teams quietly launder machine drafts into client emails and hope nobody notices. What I do on proposals is the opposite. I label the section I wrote myself, in plain language, so there's no ambiguity about which words are mine.
"This part is written by Britt. No editing. This is just human to human saying, 'Hey, I hear you. This is how I think we can help you.'" Britt Gage
The label does two things. It tells a client that the analysis and the assembly were fast, which is a selling point. And it tells them that the part where someone decided they understood the problem was a person's judgment. Regulators have made clear that claims about AI in your product need to match reality [2], and the same standard is worth applying to your own client communications well before anyone requires it of you.
Support bots done badly and done well
The failure mode is always the same: a bot placed where the stakes are highest, because that's where the ticket volume hurts most.
| Situation | What a bot should do | How it fails |
|---|---|---|
| Password reset, how-to question, doc lookup | Resolve it instantly, at 2am, without a queue | Rarely fails. This is the right job for it |
| Billing discrepancy | Gather the invoice details, then route to a human with the context attached | Tries to adjudicate, and gets the number wrong |
| Outage during the client's own launch week | Acknowledge, timestamp, escalate inside a minute | Offers help-centre articles to someone watching their launch break |
| Frustrated client hinting at cancelling | Nothing. Route to the account owner immediately | Asks the client to rate the interaction one to five |
The rule underneath the table: a bot should be allowed to resolve, or allowed to escalate, and the design decision is which of those two you've given it for each scenario.
Handing off a client without losing the relationship
Turnover is the moment where the split between the two kinds of work becomes expensive. A good system makes the handoff survivable. It doesn't make it invisible.
| Transfers cleanly through a system | Only transfers human to human |
|---|---|
| Contract terms, dates, renewal history | Which promise the client is still quietly annoyed about |
| Ticket history and resolution times | How much detail this particular buyer wants |
| Recorded commitments and deadlines | Who to call before a decision goes to the board |
| Product usage patterns | Whether they prefer a call or hate them |
The practical version: give the incoming person the full record a week early, then put the outgoing person on one shared call where they say out loud the things they never wrote down.
Renewals are decided long before the renewal conversation
Retention economics are well documented, and keeping the right customers is where the margin lives [3]. The part that gets modelled badly is when the decision happens: across eleven months of small interactions, most of which were too minor to log, long before anyone sits down for the renewal meeting.
Sarah Matthew's line on this has stuck with me all year: "Nobody cancels on someone who knows their kid's name."
You can automate the renewal reminder. You cannot automate the eleven months. What automation buys you is the time to do them properly, which is the entire argument for putting a tool in front of the rote column in the first place.
Where the human hours should go
Our operators at of All Trades embed inside client teams and work in the client's own tools rather than advising from outside [4], and that structure exists because the judgment column can't be delivered at arm's length. If you're a founder still carrying customer success personally, the sorting exercise is worth an afternoon on its own. Automate the reminders. Keep the calls.
- Week one: log the work and sort it
- Week two: automate one rote task and review every output
- Week three: name a human owner for each judgment item
- Week four: call the three clients you've had the least real conversation with
FAQ
Is customer success the same thing as customer support?
Support resolves problems a customer brings to you, usually through tickets, and is measured on resolution. Customer success is proactive ownership of whether a client is getting value from what they bought, measured on retention and expansion. Automation suits support workflows far more readily than it suits success workflows, because support handles bounded questions and success handles open-ended relationships.
Should we tell clients when a proposal or email was AI-assisted?
Yes, and labelling it explicitly works better than mentioning it vaguely. Marking which section a person wrote without editing tells the client that the human judgment in the document is real. Silence on the question costs more than disclosure does, because a client who discovers AI involvement on their own will wonder what else went unmentioned.
What customer success tasks are safe to automate?
Tasks where a wrong output costs rework rather than a relationship: scheduling, renewal date tracking, usage report assembly, meeting capture, ticket routing, and cadence reminders. Tasks involving pricing changes, escalations, scope disputes, bad news, or any conversation where someone is already frustrated belong with a named human. The test to apply is what a wrong answer costs.
Why do clients cancel even when the product works fine?
Cancellation decisions accumulate across months of small interactions rather than forming in the renewal meeting. A client who feels anonymous to their vendor will leave for a marginally cheaper option, while a client who feels known will absorb a fair amount of friction first. Working product with no relationship is a weak position at renewal time.
How does AI help when a CSM leaves and a new one takes over?
A maintained system can hand the incoming person the full factual record: contract terms, commitments made, ticket history, usage patterns, and recorded decisions. What it cannot hand over is unrecorded context about how the client prefers to be handled and which past promises still sting. Pairing a complete written record with one live conversation between the outgoing and incoming owner covers both halves.
References
- Zenger, J. and Folkman, J. "What Great Listeners Actually Do." Harvard Business Review, 14 July 2016. https://hbr.org/2016/07/what-great-listeners-actually-do
- Federal Trade Commission. "Keep your AI claims in check." FTC Business Blog, 27 February 2023. https://www.ftc.gov/business-guidance/blog/2023/02/keep-your-ai-claims-check
- Gallo, A. "The Value of Keeping the Right Customers." Harvard Business Review, 29 October 2014. https://hbr.org/2014/10/the-value-of-keeping-the-right-customers
- of All Trades. Company site. https://weofalltrades.com