We all talk about product-market fit (PMF) and it being a measure of how well a product meets the needs of a market and satisfies its intended purpose. When you achieve PMF your target customers are buying and using your...
The Copilot and Agentic Paradigms are Blurring
After the releases of ChatGPT and GPT-4, as developers dove into building applications powered by foundation models, two broad application paradigms were discussed: copilots and autopilots. The “autopilot” terminology quickly morphed into the concept of an “agent,” which is how it is referred to now. Eighteen months later, the distinction between these two paradigms is already blurring, and it is unclear whether we will still be using this terminology in a few years.
A copilot can be thought of as an assistant. A copilot requires interaction from the user, but in exchange it augments the user’s productivity, usually by retrieving or summarizing information, making suggestions, and performing analysis. To take advantage of the power of LLMs, copilots usually interact with users via natural language. An agent is distinguished from a copilot by its level of autonomy. It is expected to perform a task without human input by engaging in planning, tool use, and self-evaluation of its own work in order to troubleshoot problems and fix errors.
While early generative AI applications were usually one or the other, with copilots being much more common due to less maturity in agentic application design, nowadays most applications I see are a hybrid of both paradigms. Whether the end user is a sales rep, a software developer, or a finance professional, startups building generative AI applications are determining that the winning value proposition combines both copilot and agentic elements. The precise mix varies by use case, and this is a general observation rather than a prescriptive rule, but it is something I am seeing over and over again across a heterogenous mix of industries.
Early on, I observed that startups were leaning heavily into the copilot-centric idea of an assistant that a user converses with in natural language. What they found is that using a conversational interface to retrieve information or pursue hidden insights in the customer’s data represented a behavior change that did not come naturally to users and was too ad hoc in its frequency. Instead, they found that customers most strongly resonated with the idea that AI could simply take over work that they needed to do as part of their regular work cadence. The “magical” AI experience was coming into the office and seeing that the AI had already done the tasks on which one would have otherwise spent the first two hours of the day. Or that the AI had already fixed a problem before the user needed to sink time into investigating it.
However, the copilot modality still has value, for at least three reasons. First, ad hoc inquiries pursued through a conversational interface can still have a role in enhancing productivity. Second, there are many use cases in which the best approach is for the AI to make a suggestion that the user can take as a starting point and modify as needed. GitHub Copilot has firmly demonstrated that such solutions can have a significant and highly measurable ROI. Third, the copilot interface can often serve as the mechanism for configuring and triggering more agentic actions. I expect a common paradigm to be one in which a user works with a copilot to craft a “plan” for the agent to execute.
Perhaps the biggest takeaway here is that copilots and agents are not mutually exclusive and will often be part of the same solution. The ultimate goal is to solve a business problem, using generative AI in whatever way delivers the most value. For this reason, I’m not sure the terms “copilot” and “agent” will be enduring ones as AI evolves.
If you are building generative AI applications for the enterprise using copilot and agentic approaches, we’d love to hear from you.
(p.c. Canva magic media)
