Why Fully-Automated Customer Service Derails the Road to Product-Market Fit
With the rise of “AI Agents”, the latest startup’s flavor-of-the-month seems to be the promise of “fully automated customer service (CS) teams” to improve the efficiency of the company and reduce the size of CS teams as the company scales. While on the surface, this seems like a no-brainer, and I’ve advocated for efficiency many times, including here, here, here, and here, in the long run, full automation creates more problems than it solves, introducing more friction on the path to product-market fit. Here are the top 3 reasons why claims of “we’ve fully automated our Customer Service team with AI Agents” are a yellow flag of poor product-market fit.
Fully Automated Agents don’t Support a Diverse Customer Base
As someone who helps accelerators and investors with product testing (and diligence) of AI startups, who’s also not male and not a native English speaker, I’ve run into (too) many scenarios where these agents can’t understand me – and hang up! (I’ve talked about this at LATechWeek, in my last Substack article, and in this workshop on what I look for when doing diligence.)
Fully Automated Agents Make (Catastrophic) Mistakes
Most companies (let alone startups) don’t have the right logging and alerting in place to even know when mistakes happen, or when code breaks at 3 AM! When AI Agents make mistakes, the consequences for the company and its brand can be severe. Companies get sued for less!
We’ve all seen Tay (Microsoft’s chatbot) make the news! Regulated industries such as Legal, Finance, or Medical devices are also in the news due to the consequences of using AI Agents.
If I were mentoring a company developing AI Agents, I’d recommend rolling out tiered support, starting with suggesting replies (as a co-pilot) that the AI Agent is really sure about getting correct (for example, by focusing on the very simple, repetitive things), and have humans accepting/modifying the results as a training/validation step; more on this here. AI Agent support that also includes the option to talk to a human Customer Service agent.
Another way to think about this, is, if a startup were preventing plumbing leaks in an apartment, they’d start, for example, by flagging floods (which we’ll define as leaks over 50 gallons), before moving on to detecting smaller leaks, and eventually to identifying drops of water – but only once they’re able to detect all of the larger water intrusions to a desired (small) error rate. When the detector is not sure, it kicks it off to a human for verification.
Amazon’s Contact Us page is a great example of this, as it pre-populates orders and suggestions of what a customer may be needing help with, with customers validating the reason(s) for support. I’ve helped multiple startups implement this; one got acquired!
Companies automating Customer Support get one more key thing wrong about their customers and product(s).
