Churn in a world of AI agents

Will retention look like software or like an employee?

In software we’ve all gotten used to a world of high NRRs. The median NRR for public software companies is currently sitting at ~110% with some as high as 120-130%. In this world, if your landed ACV is $100K then after five years you can expect your customer to be paying north of $160K. It’s a pretty great business to be in.

In the future, Agentic AI systems may look more like employees than software platforms. Rather than a tool that’s used by an employee, the agent will just do that employee’s job.

If that’s the case, will their retention look like that of employees as well? 

Retention: Like SaaS or like an employee

With a software product, you generally know what you’re getting. New features may be added over time. Maybe you run into bugs. Sometimes those bugs will be big bugs but generally they won’t be. If you need to switch, it’s probably going to be a pain because you need to train all your people on the new tool and probably move over some data. That introduces a real barrier and is why retention is so strong in software.

Now let’s imagine an AI agent. This agent is a customer support agent and, like a customer support rep, it works with your existing customer support tools. It knows how to navigate Zendesk or Intercom and access certain internal systems that are required to answer customer questions. It essentially replaces the job of a human or, more likely, a hundred or a thousand humans.

Now let’s also imagine that, like some employees, the agent starts making some mistakes. The reasons don’t matter but performance starts to suffer. It’s stochastic after all and performance can’t be predicted with 100% certainty. 

If an employee’s performance was suffering, they would be fired and replaced. Is this what will happen to AI Agents if their performance isn’t up to par? If the Agent is just navigating an existing set of software systems, how hard would it be to swap in their replacement?

Why agent retention might look more like SaaS

One model of AI Agents is that it looks like your employees - it traverses various systems and takes action on your behalf. That’s the world described above.

Another, however, is that agent-like functionality lives within an existing system of record. Rather than your customer support agent sitting on top of Zendesk, your customer support agent IS Zendesk. In this version of the world, switching Agents means switching your system of record which is much more painful.

In this latter state, agent retention probably looks much more SaaS-like than not.

There’s a long way to go

We’re still in a place where AI Agent technology is unreliable which means we have time to figure out the business side of things. We first need to figure out how to chain together multiple steps of reasoning without magnifying the error rate to unacceptable levels. That’s a tough problem that I don’t expect to be solved over night.

Once that is solved (and I believe it will) then business model questions will become more important.  

To be clear - agentic AI systems are obviously worth pursuing even with some question marks around retention. Software that can do complex, multi-step work is a thing of science fiction. Making that a reality could add tremendous value to consumers and businesses.

To truly win this category, whoever figures out the technology first will need to solve the retention problem whether that is going deeper into the stack or building other hooks into their end customer. Otherwise they risk being replaced by the second, third, or fourth mover that is hot on their heels.