Canonical URL
Do not index
Do not index
Stardog Voicebox is a fast, accurate AI Data Assistant, 100% hallucination free. But what does that mean? Here I analyze some categories of AI products, especially AI data products, and answer a simple but important question.
What’s the difference between a chatbot, copilot, and an assistant?
The first thing to say is that an AI product category isn’t the same as the product’s User Interface. UI isn’t the product. UI is how we experience the product. The product is the value delivered to the user. Salesforce has a Web UI and so does Google Mail. Are they in the same product category? No, obviously not.
Three Product Categories: Chatbots, Copilots, & Assistants
I’ve arranged them in this order since in that way we proceed from general to specific, both in terms of scope and value.
What is an AI Chatbot?
A chatbot is a text or turns-based conversational product in which human users interact with non-human agents (i.e., software programs) to accomplish some human end, which may be personal, professional, or group-oriented. For example, learning conversational Greek—
The most notable example of an AI chatbot is OpenAI’s ChatGPT. Others include Anthropic’s Claude and Google Gemini.
A chatbot is the most general product discussed here, since it can support the widest range of user value-creating tasks, that is, it has the widest scope. A wide scope may be associated with a large TAM but that isn’t always the case. Scope width is neither good nor bad, inherently, it’s just one way to think about product categories.
What is an AI Copilot?
The two most well-known copilots give us a big hint as to what a copilot is; consider Microsoft and Github’s (which, er, is owned by Microsoft!). A copilot is also a conversational product that may or may not be turns-based but is narrower in scope than a chatbot. Copilots are chatbots scoped to a more limited task; for example, writing code or building a presentation at work.
Here we can see Github CoPilot writing unit tests for a
parse_expenses() function in Python.
So, then, the value prop of a copilot is similar to a chatbot but in a limited scope. You can use Github’s CoPilot to write code, make tests pass, or accept pull requests. But you can’t use it to learn conversational Greek or to write an essay about the monophonic chansonniers of the Provencal troubadour.
Copilots are, we might say, tool or domain-scoped chatbots. They provide a range of services but often in the context of specific tooling or some larger product line; for example, Microsoft Office.
What is an AI Assistant?
Finally, my favorite AI category, not least since this is the one where we’ve been investing in Stardog Voicebox. An AI assistant is, again, a conversational, turns-based product which is used to interact with either a single tool scope or a more limited range of tasks, now more fairly thought of as job-to-be-done. We call Voicebox an AI Data Assistant since its purpose to provide answers to data queries and analytics. An AI Scheduling Assistant—I like the one in Vimcal quite a lot—is scoped to jobs-to-be-done around scheduling: “cancel my day”, “reschedule my 3 o’clock”, and the like.
In an AI Data Assistant, it’s often the case that the conversational scope is also restricted to questions-and-answers from some particular domain, organization, database, or knowledge graph. The scoping here is intentionally different than in a foundational chatbot, which is trying, more or less, to know a little bit about everything. By contrast, an AI assistant is trying to know everything about one part of the world.
In that way, you can think of chatbot versus assistant in much the way that Isaiah Berlin described the fox and the hedgehog, in his book of the inverse name, which he took from an old Greek aphorism owing to Archilochus: the fox knows many things, but the hedgehog knows one big thing.
ChatGPT is a fox; Stardog Voicebox is a hedgehog.
This pleases me since I read Berlin in graduate school. Given that I live in Virginia hunt country, where the fox flourishes and often runs across my yard at dusk, and there’s a big fat groundhog (no hedgehogs in North America!) named Bobby who lives in the trees behind my office, well, I thought knowing that might please you a little, too.
Knowing versus Doing
But there’s another way to differentiate these three AI product categories: in addition to what they know, they are different in what they can do. Conventionally, chatbots chat and then often grow over time to include tool use or just conversational interface to other tools or services. Slack is a great example with its nearly endless plugins and integrations.
But AI assistants are more balanced to doing than knowing, typically. Maybe you can query scheduling data with an AI Scheduling Assistant, but its main value prop is scheduling events, not knowing them. Same goes for Github CoPilot; it writes code and tests and does reviews, primarily.
Stardog Voicebox is unusual, then, because it’s an operational assistant for data and its main value prop is knowing things and telling those things to the people who need to know them. In that way, Voicebox is like a copilot for data analytics, which may give you some ideas as to how we’re thinking about roadmap, including fun things like new kinds of AI—predictive, causal, anomaly detection—so that Voicebox can answers more types of questions for its users.
Conclusion
I will look further in the next post in this mini-series, which considers three primary data interaction types—query, search, tool invocation—among the three AI data products we considered here.
That will let us build a typology or a “9 box” that has categories on the x-axis (chatbot, copilot, assistant) and interactions types (query, search, tool invocation) on the y-axis, which will give us 9-boxes of intersection and a good, sharp tool for understanding how the AI product landscape is developing.
