Moving Agentic AI From Experiment to Production in Enterprise Customer Success

Most customer success teams have moved past asking whether AI belongs in their tech stack. The harder question is what happens when AI has to work inside a heavily regulated enterprise with real compliance requirements, real budget scrutiny, and customers who have been burned by big promises before.

On this episode of The Customer Success Pro Podcast, Neil Sparkes, who leads the Customer Excellence Group for UK and Ireland at ServiceNow, shares what it actually takes to move agentic AI out of the pilot phase and into live production. Neil looks after every post-sale function for roughly 700 enterprise customers, and before SaaS he spent 17 years in financial services buying and running this kind of technology from the customer side. That dual perspective shapes everything he says about adoption, risk, and value.

Efficiency Gains No Longer Win Budget

The first thing Neil challenges is the way most vendors and most CS teams talk about AI value. Telling a customer that AI will make their team more efficient used to be enough. It is not anymore.

His point is blunt. Efficiency is easy to claim and almost impossible to prove. If you tell a CFO that AI will make a team twenty percent more efficient, the immediate question is whether that means twenty percent fewer people, or one fewer line item on the balance sheet. Boards under cost pressure want to know what is coming out, not just what is going in. This is playing out right now with AI tokens, where consumption is running ahead of measurement and finance leaders are receiving bills before anyone has established the return.

The answer Neil's team uses is benchmarking. Because ServiceNow has been running enterprise workflows for years, there is existing data on how long a process takes, how many touch points it involves, and what it costs. When an agent is layered on top, the before and after comparison is straightforward to demonstrate.

He also warns against perfectionism in the measurement itself. Teams can spend months debating whether a task takes five minutes or ten, when either number represents a real saving. If the direction is right, take the leap and start. Companies that stay in the modelling phase for a year lose ground while they argue.

Treat Every AI Agent Like a New Hire

The most useful mental model in the conversation is to think of an AI agent as a person you have just hired. The difference, as Neil puts it, is that you can stand up a thousand of them in an hour. What does not change is everything that comes after the hire.

Those agents still need access to systems and data. They still need something resembling onboarding. They need to communicate with the other agents in the business, which increasingly means Copilot, Claude, and vendor-specific assistants all operating in the same environment and needing orchestration. And they need a control framework, because the same governance question you would ask about an employee going rogue applies exactly to an agent making decisions on your behalf.

That framing drives where to start. Pick a workload that is low risk rather than low value. Neil points to incident triage as the clearest example: an agent reads the context, routes the ticket, spots patterns across multiple incidents, and can query a monitoring agent to identify a common cause, all in milliseconds. Nothing catastrophic happens if it gets one wrong.

Guardrails then define where autonomy ends. An agent can execute up to a set threshold, and beyond that it presents options and asks a human to confirm. In a corporate setting there is no room for the ambiguity people tolerate in a chat window, where a model gets something wrong and cheerfully apologises. The work still has to be checked, exactly as you would check the work of a new team member.

The AI Divide Is Widening, and So Is the Human Load

Neil describes a growing AI divide between fast adopters and everyone else. The adopters accepted some risk early, and AI is compounding their advantage. The concerning part is that the laggards often do not perceive how far behind they have fallen.

There is a second risk running alongside it, and it lands on people rather than systems. AI has made producing information almost free, so teams now send more of it than ever. Leaders then use AI to summarise what AI generated, and the volume of context switching keeps climbing. Neil calls the underlying habit AI laziness, or a lack of marking your own homework. If agentic workflows are not built to genuinely absorb the work, the human at the end of the chain simply absorbs more.

The conclusion is not to slow down. It is to build agents that remove workload rather than generate more content, and to keep the value conversation anchored to outcomes a CFO recognises.

Key Takeaways

Stop selling efficiency and start benchmarking hard cost and revenue impact, because boards have no tolerance left for soft claims. Treat every agent like a new employee that needs access, onboarding, guardrails, and supervision. Begin with low risk, high volume workloads such as incident triage and expand as confidence builds. Define the exact point where an agent must stop and ask a human to decide. Watch the information load your AI is creating for your team, and design workflows that remove work rather than multiply it. Above all, start, because directional evidence beats a perfect business case delivered a year late.

🎧 Listen on your favorite platform:

🔹 YouTube

🔹 Spotify

🔹 Apple Podcasts