The Customer Success Agent Problem Nobody Is Solving Yet
The Customer Success Agent Problem Nobody Is Solving Yet
Every CCO has a Claude project open in another tab. Most of them are stuck at stage two of the AI maturity curve, confusing demos for systems. Here's what stage three actually looks like, and why building it is a stack decision.
Every CCO I've talked to in the last two months has the same Claude project open in another tab. It summarizes QBRs. It drafts renewal emails. It pulls account context before a call. Useful work, but the human is still the one taking action after Claude finishes thinking.
The next step is obvious once you see it. Claude stops being a research assistant and starts being an operator. The deal closes and onboarding is already moving. The customer goes quiet and the CSM has a save brief before the next standup. The P1 ticket hits and the manager walks into the room already briefed.
This is the shift that Claude Routines made real. A routine is an event-triggered AI worker that reads, decides, and acts. It is not a workflow builder like n8n or Zapier where you draw a flowchart. It is an outcome, described in plain English, that Claude figures out how to execute.
CSAT goes up because the first twenty-four hours after contract signed stop being dead air. Onboarding time comes down because the clock starts the hour the deal closes, not the Monday after the CSM clears their backlog. Wasted CSM cycles come down because the reading, drafting, and chasing gets done by the routine instead of the person.
Most CS leaders will try writing their first routine this weekend. Most of them will hit the same wall. This post walks through why that wall exists, what the production-ready version looks like, and why building it correctly is a stack decision, not a vendor decision.
The capability maturity model for AI in customer success
After sitting in on four or five CCO conversations where the person on the other side of the table was lost on where to start, I built a framework. I call it the Capability Maturity Model for AI in Customer Success. It has three stages, and almost every CS organization I have talked to this year is stuck at stage two.
Where your CS operation actually sits
Salesforce Agentforce
Zendesk AI Assist
Intercom Fin
QBR drafters
Renewal email writers
Lives on a laptop
Multi-system workflows
Service-account auth
Production logging
The jump from stage two to stage three is where most CS teams get stuck. The agents are the easy part. The orchestration layer underneath is where the work is.
Stage one: what ships with your stack
These are the agents that come bundled with whatever vendor is already in your stack. Every CS team should be using the agents that come with the tools they already pay for. If you are not, start there before you spend a dollar on anything custom.
Stage two: the demos that feel like progress
This is where most teams are stuck. Someone on the team took a Claude class. They built a little dashboard that pulls from Salesforce and shows account health. Maybe they built a QBR drafter. They are proud of what they shipped, and they should be, but nothing is operationalized. The agent runs when a human runs it. It writes when a human asks it to write. It lives on a laptop. It is a demo.
Stage three: where the business value lives
Orchestration means an agent that listens for signals across your stack and triggers the right action without a human pressing a button. A contract downgrade in HubSpot fires a save brief to the CSM in Slack. A product usage drop in Pendo fires a health score update and a task to the account owner. A P1 ticket in Zendesk fires a context brief to the on-call manager. The routine watches, thinks, and acts while the team is in meetings or asleep.
Most teams cannot get from stage two to stage three by themselves. They have the agents. They do not have the orchestration layer underneath.
Why the orchestration layer is the hard part
I was at a 250-person event in Boston a few weeks ago called GTM with AI, hosted at Lovable's offices. People were showing what they had built. Most of what got applause was stage two work. Dashboards that pulled from disparate data sources and showed insights. No triggers. No operational actions. No routines running in the background. Just queries, formatted as cards.
That is the state of the market right now. Companies are proud of their stage two work because it feels like progress. It is progress, but it is not what produces business outcomes.
Business outcomes come from stage three. And stage three requires an orchestration layer that most companies do not have and cannot build themselves. Here is what it actually has to do:
None of this is hard to understand. All of it is hard to build. And most importantly, none of it is the interesting part of AI in customer success. The interesting part is the agent that writes the brief or drafts the email. The orchestration layer is the plumbing.
The stack is not one product. It never has been.
The temptation for every vendor in this space right now is to claim they are the platform. HubSpot wants to be the platform. Salesforce wants to be the platform. Gainsight wants to be the platform. Every new AI startup wants to be the platform.
None of them are the platform, because the real platform is the combination of tools a customer already uses. The work falls down outside of any single product.
They don't care about the software anymore. They care about the job being done. They want to hire an agent to do work for them.
I had a long conversation this week with Jay Nathan, CEO of Balboa Solutions, and Carl Yost, Chairman of All & Sundry. Balboa is the Pendo implementation specialist. All & Sundry is an Airtable consultancy for the enterprise. Between their work and mine, we cover most of what a PE-backed SaaS company actually runs.
The conversation kept landing on the same point. A customer success routine that matters cannot live inside one product. It has to move across products. Here is what a production-ready churn risk response routine actually looks like:
One routine. Five layers. One orchestration brain on top.
The routine is not a feature of any single tool. It sits on top of all of them. This is why the platform pitch does not hold up. No platform has all the data. No platform has all the workflows.
Why Pendo matters here
Most health scores are built on data a human remembered to configure six months ago. They are static. Pendo Predict applies AI models to in-app behavioral data to forecast churn risk and upsell potential. It surfaces signals before they would ever make it into a manually-configured health score. This is the input that tells the routine something is wrong, often weeks before a renewal meeting would have caught it. Pendo's newer Agent Analytics product does the mirror-image job: it measures how users interact with the agents you deploy, so you can prove the agent is producing outcomes and not just running.
Why Airtable matters here
This is the part most teams underbuild. The customer journey has steps that do not live cleanly in a CRM. Onboarding milestones, implementation checkpoints, training completion, go-live dates, executive sponsor meetings. Airtable has become the de facto place where this stuff gets tracked, because the CRM was built for sales and the ticketing system was built for support. The workflow in the middle is where Airtable lives. A routine that ignores the workflow layer is a routine that misses the actual state of the customer relationship.
What this means for CCOs
If you are a Chief Customer Officer at a 500 to 5,000-person SaaS company, here is what is on your desk right now. You know you need to be doing something with AI. Your board has asked about it. Your CEO has asked about it. Maybe you have a few people on your team experimenting.
You also know that the experiments are not producing business outcomes. They are producing demos. And you know, at some level, that the gap between a demo and a system is wider than it looks. You have two options.
The reason option two wins is simple. The domain expertise matters more than the engineering work. Someone who has run customer success for twenty-five years knows what a churn risk response should look like. Someone who has only written code does not.
What a desktop routine is missing, and what production adds
- Runs only when the laptop is open
- Authenticates with your user token (breaks on password change)
- No log of what it did last week
- Fails silently when something goes wrong
- No permission boundaries on customer records
- Cannot trigger downstream agents safely
- Runs 24/7 on infrastructure you do not maintain
- Service-account authentication per system
- Every action logged and reviewable
- Alerts on failure, on-call rotation
- Scoped write permissions, audited quarterly
- Multi-agent orchestration with approved templates
Desktop is where you learn the pattern. Production is the system. The same routine looks nothing alike in the two environments.
How Infinite Renewals approaches this
Three things, in order:
We price this as a managed service because that is how you should buy it. Not a one-time build. Not a piece of software. An agent fleet that we build, maintain, tune, and scale as your customer base and stack evolve.
Get a read on where your CS operation actually sits
The CS AI Diagnostic is a $5,000 engagement. Two weeks, on-site or remote, with a prioritized agent roadmap at the end. The deliverable tells you which three agents would produce the most business value in your specific stack, what the build looks like, and what it would cost to operate.
Start the diagnosticThe question to ask yourself
Most CS organizations will end 2026 with a pile of stage two experiments and no stage three system. They will have spent six to eighteen months convincing themselves that the demos are progress.
The question worth asking: Is your team going to spend next year building the orchestration layer themselves, or are you going to buy the agents that actually produce outcomes?
If the answer is the second one, we should talk.
Questions CCOs are asking about customer success AI
What is a customer success AI agent?
A customer success AI agent is an event-triggered AI worker that reads data across your stack, makes a decision, and takes an action without human intervention. Unlike a dashboard or a chatbot, an agent responds to signals like contract downgrades, usage drops, or P1 tickets by writing briefs, scheduling meetings, updating records, or triggering downstream workflows. A production-grade CS agent runs on server infrastructure, authenticates with service accounts, and logs every action for compliance review.
What is the difference between a Claude Routine and a Zapier or n8n workflow?
A Zapier or n8n workflow is a flowchart. You draw every step in advance, and the tool executes each step in order. A Claude Routine is an outcome described in plain English. You tell Claude what finished looks like, and Claude figures out the execution dynamically. Routines handle ambiguity and context that workflow builders cannot, like reading unstructured notes, summarizing a customer relationship, or deciding which of three possible actions best fits the situation.
Why can't I just run Claude agents on my laptop?
Desktop agents fail in five ways that make them unfit for production. First, they only run when the laptop is open. Second, they authenticate with your personal user token, which breaks when you change your password or leave the company. Third, there is no log of what the agent did, so compliance and audit review is impossible. Fourth, when the agent fails, no one is alerted. Fifth, there are no permission boundaries, so an agent that can write to one customer record can accidentally write to the wrong customer record. Production agents run on server infrastructure with service-account authentication, full logging, alerting, and scoped permissions.
What is the capability maturity model for AI in customer success?
The Capability Maturity Model for AI in Customer Success has three stages. Stage one is out-of-the-box agents that ship with your existing vendors, such as HubSpot Breeze, Salesforce Agentforce, Zendesk AI Assist, or Intercom Fin. These are table stakes. Stage two is custom-built agents, typically dashboards or drafters built in Claude or Lovable that run when a human runs them. Most CS organizations are stuck at stage two. Stage three is orchestration, where agents watch for signals across the stack and take action automatically. Business value lives at stage three.
What tools do I need to run AI agents for customer success in production?
A production-ready CS agent stack has five data layers and one orchestration layer. The data layers are product usage analytics (Pendo Predict for behavioral churn signals), a CRM (HubSpot or Salesforce for contract and commercial data), a workflow state system (Airtable for onboarding milestones and the messy middle), a ticketing system (Zendesk, Intercom, or HubSpot Service Hub for support history), and communications (Slack, Gmail, and call recording). The orchestration layer sits on top, running Claude agents connected to these systems via MCP servers with service-account authentication, complete action logging, and scoped permissions.
How much does a customer success AI agent cost to build?
A custom CS AI agent built and operated as a managed service typically ranges from $15,000 to $30,000 for a single premium agent including orchestration setup. Infinite Renewals starts most engagements with a $5,000 CS AI Diagnostic that maps the customer journey, identifies choke points, and produces a prioritized agent roadmap. A full agent fleet covering post-sales handoff, churn risk response, escalation, renewal prep, and upsell detection is priced as an ongoing managed service rather than a one-time build.
Should I build AI agents in-house or buy them as a managed service?
Most CS organizations should buy AI agents as a managed service rather than build in-house. Building in-house requires hiring AI engineers for at least six months, another six months for integration work with platforms like HubSpot and Pendo, and domain expertise that engineering teams typically lack. A managed service provider with CS domain expertise can build, deploy, and operate the agent layer in weeks instead of months, using the client's existing stack without rip-and-replace. The domain expertise of knowing which agents to build first and what a QBR brief should actually say matters more than the engineering work.
What is an orchestration layer in AI agent architecture?
An orchestration layer is the infrastructure that sits on top of a company's existing tools and coordinates AI agents across them. It handles five things that individual tools cannot: service-account authentication to the CRM and ticketing system, complete logging of every agent action, permission boundaries that prevent agents from writing to the wrong records, observability and alerting when agents fail, and the event fabric that watches the stack for signals worth acting on. Without an orchestration layer, AI agents can only run in demo mode on a developer's desktop.