The temptation with AI in Customer Success is easy to understand. Once a team can automate summaries, alerts, follow-ups, account prioritisation and playbook execution, it becomes possible to remove a lot of repetitive work.
I am interested in that. I have used automation in post-sale operating models and I am building with AI now.
What I am more cautious about is treating automation as the starting point.
The question I would ask first is simpler:
Is the operating logic underneath the automation good enough to scale?
Automation inherits the model beneath it
If the customer journey is clear, the important states are well defined and the signals are trustworthy, automation can be extremely useful.
If those things are unclear, the system can create activity without improving the decision. A weak playbook does not become strong because software executes it. An ambiguous health state does not become more meaningful because an AI model explains it. A missing customer outcome does not appear because the system has access to more events.
This is why I think AI strategy for Customer Success should begin one layer below the technology.
What are we trying to understand? What decision are we trying to improve? What state should trigger action? Which evidence supports that state? What can the system do safely when it sees it, and when does a human need to take over?
Those are operating-model questions.
Common language becomes operating infrastructure
A shared language across sales, implementation, Customer Success and product has always been useful. AI makes the consequences of inconsistent definitions more obvious.
Humans can often compensate when the same word is used in different ways. We hear "adoption" used three times and infer the intended meaning from context.
Automation has a harder job. If one system treats adoption as login frequency, another as completion of a use case and another as licensed-seat penetration, the AI layer can combine them into a fluent answer without resolving the underlying ambiguity.
Before automating a state, define it. Before asking AI to recommend an action, define what evidence makes the action appropriate. Before asking it to identify risk, decide what kind of risk you actually mean.
The taxonomy is part of the operating model the machine inherits.
Start narrower than the demo suggests
AI demonstrations naturally show breadth: read everything, summarise everything and act across everything.
I would start narrower.
Take one customer state that matters. Define the evidence. Define the permissible actions. Define what the system must not infer. Define when the state is too uncertain for automation. Then test whether the intervention changes something meaningful.
If it does, expand it. If it produces engagement without moving the customer towards the desired state, change it or remove it.
That is the same discipline I used in Digital Customer Success work before today's generation of AI tools existed. The technology has changed; the need to test whether an intervention improves a meaningful customer state has not.
AI should reduce reconstruction
One place I think AI can be particularly valuable is context reconstruction.
Customer-facing teams often have to assemble the same picture from several systems: what sales sold, what changed in implementation, which capabilities the customer adopted, what happened in the last conversation, what remains unresolved, who owns the next step and what evidence of value exists.
If the information already exists, the CSM should not have to become the integration layer every time they need to make a decision.
AI can help assemble the relevant context and present it at the point of use. I see that as different from asking the model to declare that the customer is healthy, at risk or ready to expand without exposing the evidence behind the conclusion.
The first reduces cognitive overhead. The second creates a decision dependency that deserves stronger evidence and controls.
Automate after the diagnostic
This is why automation and AI sit as one of the areas I examine through Routeability rather than the headline solution.
If the problem is a broken sales handoff, fix the handoff before automating it. If definitions are inconsistent, stabilise the taxonomy. If adoption is weak, establish which behaviours matter. If the operating logic is already clear and people are doing repetitive work manually, automation becomes much more attractive.
The easiest part of AI is making something happen.
The more important question is whether it was the right thing to happen.


