← All insights
James Hayward-Zhu4 min read

What AI in CS Actually Changes (And What It Doesn't)

What AI in CS actually changes vs what teams think it changes
A paper collage on green. A halftone figure holds up a large hand-drawn tick at the centre, surrounded by a watering can, growing plants and sheets of data. Vintage halftone men, modern photographs of smiling people, a cartoon robot and engraved hands are arranged around them.

Every CS platform has AI features now. Call transcription, automated email drafting, churn risk scoring, sentiment analysis on support tickets, next-best-action recommendations. The demos are impressive. The implementation stories are considerably more mixed.

The reason, I think, is that most CS teams are using AI to do more of the same (faster, with less effort) rather than using it to interrogate whether what they were doing in the first place was the right thing.


The automation trap

The most common AI deployment in CS right now is some variation of "automate the administrative work so CSMs have more time for high-value activity." QBR summaries drafted from CRM notes. Follow-up emails generated from call transcripts. Health score flags surfaced automatically rather than requiring manual review.

These are real productivity gains. I'm not dismissing them. But there's an assumption embedded in "more time for high-value activity" that's worth examining: namely, that CSMs know what the high-value activity is, and that they were constrained only by time.

In most CS teams, neither of those things is reliably true.

CSMs often don't have a precise model of which activities actually drive retention versus which ones feel like they do. The team that does the most QBRs is not necessarily the team with the best retention. The CSM who makes the most check-in calls isn't necessarily the one who catches the most at-risk accounts before they churn. The correlation between input activity and output outcomes is weaker than most CS leaders want to believe, and it's almost never measured directly.

So when you free up CSM time by automating administrative tasks, you get more of whatever those CSMs were already prioritising, which may or may not be the right things.


What AI actually makes possible

The more interesting AI capability in CS isn't automation. It's pattern recognition at a scale that wasn't possible before.

Historically, a senior CSM carries around a model of "what does a healthy account look like at 6 months, in this industry segment, at this contract size?" That model is built from experience: from hundreds of accounts, from seeing what led to renewals and what led to churns. It's genuinely valuable. And it's locked inside one person's head, which means it doesn't scale, doesn't survive attrition, and isn't consistently applied across the team.

AI can externalise that model. It can look across thousands of accounts and identify the signals (product usage patterns, engagement cadences, support ticket themes, executive sponsor stability) that reliably precede churn or expansion. Not as a single health score number, but as a set of specific, actionable patterns: "accounts with this feature adoption gap at 90 days have a 40% higher churn rate in months 10 to 12, unless X intervention is applied."

That's a different kind of capability. Not "help CSMs do their existing job faster" but "give CSMs access to pattern recognition they couldn't develop on their own, based on data they couldn't hold in their heads."


The CS Ops prerequisite

This is where AI in CS runs into its real barrier, and why so many deployments underperform: the quality of the output is completely determined by the quality of the data going in.

Churn prediction that's trained on unreliable CRM data produces unreliable predictions. Sentiment analysis that surfaces support ticket themes is only useful if someone has already done the work to map ticket categories to CS-relevant signals. Next-best-action recommendations are only as good as the action library they're drawing from, which means someone has to have built, documented, and validated that library first.

All of that is CS Ops work. It predates the AI layer by years in most organisations that do it well. And it's the reason CS teams at mature SaaS companies tend to get significantly more value from AI tools than CS teams at earlier-stage companies, not because the tools are different, but because the underlying data infrastructure is better.

If you're evaluating AI tooling for your CS function and the vendor's primary use case is "summarise your calls," that's a signal that the tool is designed for teams that don't have the data infrastructure to do anything more ambitious. Which may be fine for where you are. But it's worth being honest about the ceiling.


What AI doesn't change

The things AI doesn't change: the need for human judgment in conversations where trust is being built or broken. The need for clear-eyed diagnosis of whether a customer is genuinely successful. The need for someone to tell a customer something they don't want to hear.

AI can surface the evidence that makes those conversations possible. It can prompt a CSM to have the conversation earlier. It can draft the email that starts it. But the judgment call (do we have this conversation now, at what risk, with what framing) remains human. And the relationship that determines whether the customer receives it constructively is something no AI is going to build for you.

The CS teams I expect to do best with AI over the next few years are the ones that treat it as infrastructure for better judgment, not as a replacement for it. The ones that ask "what does our data tell us that we couldn't see before?" rather than "how do we use AI to do what we're already doing with less effort?"

The first question leads somewhere different. And usually somewhere better.


Questions worth sitting with: If your CS team's AI tools were taken away tomorrow, which of your current processes would genuinely break? Which ones were already broken and the AI was just masking it? And what's the one thing you could learn from your account data that you currently can't, because the data isn't structured well enough to ask the question?

Share this

Email

Working on this in your business?

We help UK SMEs and scale-ups turn this kind of thinking into action.

Bring us the problem. Leave with a path.

Request a free 30-minute strategy call