Considerations for the Acquisition of a Chatbot

Here at MoP, it’s clear we’ve been perturbed about the recent boom in terrible AI applications. Sometimes I tell people that this newsletter is like Agent Scully from the X-Files running a blog on paranormal events. Yeah, we’re the skeptics, but we still show up to work every day with partners who think their compute comes from a parallel universe. 

With that in mind, now we’ll use our critical approach to guide you through a decision that plenty of companies are struggling with right now: Should we build a chatbot? 

We’ve put together seven questions you should ask yourself to help you figure it out.

1. Can You Avoid The “Empty Field” Problem?

Digital life pre-2023 was mostly navigated through a series of buttons, drop downs, checkboxes and form fields. In that context, a wide open  text box can feel jarring, if not frightening. 

So we don’t overwhelm our visitors, it should be clear what they can use the chatbot to accomplish. A chatbot should excel at specific tasks, rather than cooing, “ask me anything!” People intuitively know that prompt is a lie. 

For example, here’s what happened when MoP tried to find a video (that we’d already watched) about throwing dinner parties in small apartments on Instagram: 

We were prompted with “Ask Meta AI or Search.” We wanted Search and got an encyclopedic chatbot. Then we gave up.

2. Is it Obvious When This Thing is Wrong? 

LLMs sometimes return inaccurate results. That cost can be worth the benefit when the stakes aren’t super high. (Examples of high stakes: medical or time-imperative safety information). Beyond that, when your LLM is wrong, the nature of the error should be easily understood. Maybe the AI misinterpreted the question, or there’s an obvious category error. 

But if your chatbot is meant to help folks research obscure subjects they aren’t expert in, consider whether you are sowing the seeds for your own very public demise.

3. Have you Streamlined Human Interaction, or Simply Replaced Humans? 

Let’s say you want to build a company HR bot. That could be useful – after all, often folks just want to know about the corporate holiday schedule or find a company policy. That would also free up time for HR to better deal with complicated issues that require a human touch. It would be far less useful, however, if the HR bot justified layoffs that made it harder to talk to a real person in HR, which would have the long term effect of reducing compliance. 

Another counter example: The suggestion that talking to a chatbot can be a replacement for talking to the person. “Oh, you can ‘meet’ Taylor Swift!” I promise you, friend – no one cares, and if they somehow do care, they’ll despise you by the end. Or for a more concrete example, I will simply quote a recent headline from 404 Media: “Meta's AI Profiles Are Indistinguishable From Terrible Spam That Took Over Facebook”

4. Do You Have a Unique Competitive Advantage? 

Why would someone come to YOUR chatbot to ask these questions? If there’s not a clear reason why YOU should be trusted to convey the information, your chatbot is simply not going to do the thing. 

5. Do You Know Whether the Thing Works, and Will You Know If it Stops Working? 

You’ve gotta find a way to clearly test the effectiveness of a chatbot over the long term in an enduringly affordable way. 

How can you tell if it’s making people happy rather than mad? If its usefulness degrades over time, do you have a systematic method for catching that? 

It would be cool to not write a post called “Exploring the Mausoleum of Dead Chat Bots” 10 years from now in our stately country manor as our new landscaping team trims the South koi pond. 

6. Think About It: Are You Making Your Product Easier or Harder to Use? 

If the thing you’re offering is systematically complicated, you may be able to simplify it with a chat interface! But if your product is already easy to use, don’t make it more complicated! 

Remember, no matter how cool the demo looks, there is some friction to making people type stuff, and that may be more or less friction than your product already has. 

7. Have You Named Your Chatbot ‘Erica’? 

Okay, this is Erica’s pet peeve since that’s the name of Bank of America’s chatbot and no, she does not want to ask Erica for help accessing her account. 

But seriously, it’s a little creepy that they all seem to have women’s names – Click Here and a WOMAN will be at your SERVICE! 

Or, alternately… 

P.S. If someone has notes on the meeting that ended with a consensus that this ^^ was a good idea, please do send them our way! I’m just curious! It’s not for lulz at all! I have a legitimate academic interest! hello@machinesonpaper.com