Robot or Exoskeleton?

Before building with AI, it's useful to ask whether you're building a robot or an exoskeleton. The main problem is less about the particulars of building each kind of system and more about our tendency to overlook the exoskeleton option by assuming robots are somehow more valuable.

Yes, we're using an analogy here. Robots in the real world autonomously perform a task. In AI, the same is true: "AI robots" autonomously perform a task or workflow. The robot's code does what deterministic code does best, while the AI model(s) in the robot handle fuzzy judgment, input variance, and synthesis. The combination of deterministic and non-deterministic elements can be very powerful!

Exoskeletons in the real world augment human power, precision, or control. In AI, the same is true. But because "AI exoskeletons" are often deployed internally within organizations, we have a less clear and colorful picture of what they look like or how they work. This is the good ole availability heuristic in action. We undervalue the AI exoskeleton pattern because it's less visible, and it's less visible because it's often deployed in settings that are not in public view. Out of sight, out of mind.

What an AI exoskeleton looks like

What does it look like to wrap an AI exoskeleton around a human task, workflow, or job? Well, it's humble, but here's one example we built...

An outbound agency already had a custom lead generation system. It was useful, but error-prone; the generated outreach messaging was often based on out-of-date context about the prospect, the discovered email addresses were often for the lead's last job rather than their current job, and sometimes the system-generated outreach was a bit too robotic-sounding. So letting this system start emailing leads on full autopilot (making the system into an AI robot) would have resulted in errors and brand damage.

Here's what the exoskeleton approach looked like: 1) quickly modify the system's lead dashboard by augmenting the primary view of leads -- which had a complexity and information density comparable to a product like apollo.io -- with a minimalistic lead view with useful keyboard navigation, and 2) in that minimalistic view, make it easy to open the lead's LinkedIn profile in a new browser tab. Then, quickly develop a browser extension that, when viewing a person's LinkedIn profile, opens a drawer with just the essential information from the lead generation system's database and provides buttons/shortcuts for the essential, commonly-needed outreach actions.

With this workflow, a human can easily and quickly see if there are discrepancies between the information in the lead generation system's database and the lead's LinkedIn profile. The latter almost always has the most up-to-date possible facts about the lead. Presenting this information as a drawer that integrates with the LinkedIn UI prevents the human operator from having to flip between browser tabs. Additionally, the browser extension can extract information about the lead directly from the lead's LinkedIn profile, so if the human operator notices that the generated outreach message has outdated facts about the lead, the operator can click a button in the extension's UI to regenerate the outreach message using live data from the lead's LinkedIn profile. …

There are several more quality-of-life features built into this particular AI exoskeleton, but this is enough for you to get a sense of how it works.

Choosing the tradeoff

The design thinking here is simple: how can we use AI to reduce the mechanical complexity and cognitive burden that a human faces in working through a backlog of leads that need semi-customized outreach while also investing a minimum amount of engineering effort and getting the solution up and running and used by real users as quickly as possible?

Yes, it would be possible to build an AI robot that accomplishes the system design goals. And to be clear, I'm not saying that an exoskeleton is always better than a robot! It depends on the context, and in this context, an exoskeleton was the better tradeoff. And there are some contexts where either an exoskeleton or a robot might be a total non-starter. (I don't think any of us want an LLM performing air traffic control operations!)

In general, you'll want to think about these specific issues as you are working through the high-level "robot or exoskeleton" question:

• What's the risk budget in the context where the AI system will operate? The combination of high importance and small risk budget will tend to push us towards an exoskeleton approach.
• How well can the engineers involved collaborate with the humans involved? With an AI exoskeleton approach, you might need a relatively higher potential for such collaboration.
• What is your time-to-value budget? (In reality it's probably less than you think or wish it was! 🙂 ) A very compressed or unforgiving timeline here might point you towards the exoskeleton approach, or at least an AI exoskeleton as a high-learning-opportunity phase 1 within a larger initiative.

That's probably enough to get you thinking about what I think is an often-overlooked pattern.