How We Built an AI Agent That Runs Our Client Success Operation

We built an internal AI agent system to run the operational layer of client success across the tools we manage for clients. It didn't replace our team. It removed the repetitive triggers so our people could spend more time on optimization, positioning, and the client work that actually moves results.

This is the architecture, the failure modes, and what we would tell any SaaS growth leader thinking about building the same thing.

The problem we were actually solving

Understory runs paid media, outbound, and creative for B2B SaaS clients. By early 2025, we had a familiar agency problem: more clients, more campaigns, more data to monitor, and no interest in doubling headcount.

Client success at a company like ours means something specific. It means knowing, at any given moment, how a client's outbound sequences are performing. It means catching a deliverability dip before the client notices. It means following up after a sales call with context from previous conversations and similar client builds.

The work is real. Pulling a cross-platform performance snapshot for one client used to mean logging into multiple dashboards, exporting CSVs, and assembling a narrative manually. Writing a meaningful post-call follow-up consumed 30 minutes of senior time per call. Monitoring reply quality across every active outbound sequence required someone watching inboxes that never closed.

When you run allbound coordination across paid media, outbound, and creative for a growing client roster, that operational layer consumes time your team should spend on optimization and positioning. Our design principle was simple: the agent moment is less about intelligence and more about removing the initiation step.

What we actually built

There is no single "AI agent" running our client success operation. We built a connected system of purpose-built agents, each tied to a specific workflow through MCP connections with Claude.

The stack breaks down into five core systems:

• Campaign intelligence via MCP: Ali built a custom LinkedIn Ads MCP internally, and we connected Claude to Instantly through MCP. Any team member can query campaign performance, spot trends, and generate reports through natural language. Wispr Flow lets us dictate questions quickly when typing slows things down.
• AI reply agent: Every positive outbound reply receives a response within minutes, generated by an agent loaded with the client's offer, value props, and answers to common questions. The knowledge base auto-updates when the offer or messaging changes.
• Context-aware follow-up: Fireflies captures the call transcript, and an MCP-connected agent produces a structured follow-up that references both the current conversation and relevant builds from other client work, without exposing client-specific details.
• Autonomous outbound pipeline: Clay handles audience building and enrichment. Claude Code drives workflow and optimization tasks. When a buying signal fires, like a recent CRO hire, a funding round, or a tech-stack change, the sequence launches automatically through Instantly for email and HeyReach for LinkedIn.
• Client reporting infrastructure: Live analytics dashboards, Slack and email webhooks for responses, HubSpot and Salesforce CRM integration, and bi-weekly meetings with quarterly business reviews.

A cross-platform snapshot for one client is now a conversation, not a CSV export.

The follow-up agent, a closer look

The MCP-connected Fireflies agent produces a structured follow-up that references both the current conversation and relevant builds from past client work, without surfacing client-specific details.

"The follow-up references not just what we discussed today, but things we've built for similar clients in the past. It's like having a perfect memory of every client case study at your fingertips," Alex explains. The reported outcome: follow-up time reduced by roughly 60x.

The autonomous outbound pipeline in practice

For one cloud cost allocation client, we connected Octave to Clay via HTTP API to analyze every contact in the target audience. The system tailored outreach messaging to relevant technical and financial stakeholder personas, and the campaign produced measurable engagement across both buyer profiles.

The client-facing reporting layer

The operational layer our clients actually see is straightforward:

• Live analytics dashboards for paid and outbound performance
• Automated Slack and email webhooks for outbound responses
• HubSpot and Salesforce CRM integration for contact records and correspondence tracking
• Bi-weekly meetings with quarterly business reviews
• Documented internal playbooks for HubSpot and OutboundSync dashboards, deliverability management protocols, and ClickUp task workflows

The internal reporting playbooks powering all of this were built and documented by our team.

Clients started asking us to offer this system as a standalone service. We are not doing that yet. As Alex has put it publicly: "Internal tools are how you build asymmetric advantages, and how we do our best client work."

The hard parts and what broke

This did not all work on day one. The failure modes we hit mirror what other teams run into when they deploy AI agents in production.

Agents need daily management

Agents drift. That is the most predictable failure mode. We designed around it. The time you used to spend doing the work manually, you now spend managing the agents that do the work. That is a good trade. It is still a trade. Going into this expecting set-and-forget automation will produce disappointment.

Deploying too fast kills quality

Rushing multiple agent deployments at once means none of them get enough attention to tune properly. Sequential rollouts with internal piloting before any client-facing exposure is the only approach that holds up under load.

The handoff is where trust breaks

The automation itself usually works fine. The break happens when the agent needs to hand off to a human. If the transition loses context, or if the agent doesn't recognize that a situation has exceeded its scope, the client experience craters at the worst possible moment.

We built escalation logic into every agent workflow. The agent has to know what it doesn't know, and every workflow needs explicit conditions for when a human takes over. Without that, the quality of the transition determines whether the overall interaction succeeds or fails.

AI should create capacity for human work, not replace it

This is the part most SaaS growth leaders miss. The instinct is to use AI to cover more accounts with fewer people. We took the opposite path. Our agents handle the operational load so our team spends more time on strategic client work: positioning refinement, creative direction, and campaign architecture.

The human layer got deeper, not thinner. The clients who would have received standardized service from a leaner team instead get senior attention on the work that compounds.

What this means for SaaS growth leaders

Operational gains from AI agents are real, but they are deployment-specific. There is no universal "deploy AI, get X% improvement" guarantee, and most of the published numbers come from teams who got the boring parts right before chasing the headline result.

The teams getting genuine value from AI agents tend to share three traits:

• They start with constrained, high-volume workflows instead of general-purpose assistants.
• They budget for ongoing agent management as a permanent operational cost, not a one-time deployment.
• They use the freed capacity for deeper human engagement, not headcount reduction.

For SaaS growth leaders evaluating where to apply AI internally, the place to start is the operational layer that already drains your senior team. The work that has to happen but doesn't compound. That is where agents earn their keep, and that is where the time savings show up in client retention and pipeline.

Run client success on coordinated AI infrastructure with Understory

Understory runs paid media, outbound, and creative for B2B SaaS companies as one coordinated team, built on the same AI agent infrastructure we use to run our own client success operation.

Book a call with Understory to see how coordinated allbound execution and operational AI work together in practice.

Frequently asked questions

What is an AI agent for client success?

An AI agent for client success is a purpose-built automation that handles a specific operational workflow tied to managing client accounts, like monitoring outbound deliverability, drafting post-call follow-ups, or responding to positive replies. It's not a general-purpose chatbot. The useful versions are connected to the tools where the work actually happens (Instantly, LinkedIn Ads, HubSpot, Fireflies) through MCP or API integrations, and they're scoped to one job each. Multiple agents working together cover the operational layer that drains senior time without contributing to strategy.

How long does it take to deploy an AI agent system like this?

Deployment time depends on how many workflows you're automating and how clean your existing tool stack is. A single agent tied to one workflow, like an AI reply agent for outbound, can be live in two to four weeks. A connected system across paid media, outbound, follow-ups, and reporting takes longer because each agent needs internal piloting before client-facing exposure. We rolled ours out sequentially across roughly six months. Rushing multiple deployments at once is the most common failure mode and produces agents that drift quickly because none get enough tuning attention.

Will AI agents replace our SDRs or account managers?

Not if you build them correctly. AI agents handle the repetitive operational triggers that consume senior time without contributing to results: pulling cross-platform reports, drafting follow-ups, responding to positive replies, launching sequences when a buying signal fires. They free your team to spend more time on positioning, creative direction, campaign architecture, and the relationship work that retains clients. The teams that use AI to cut headcount tend to lose the depth of client engagement that justified their pricing in the first place.

What's the difference between an AI agent and workflow automation in Zapier or Make?

Workflow automations like Zapier and Make are deterministic: if X happens, do Y. They're excellent for high-volume, predictable triggers and remain part of our stack. AI agents add reasoning and context. The follow-up agent doesn't just send a templated email after a call; it reads the transcript, references prior conversations with the same prospect, and pulls in relevant builds from similar client work. Most useful systems combine both layers. Deterministic automations handle the predictable triggers, and agents handle the steps that require judgment or synthesis.