Iron Man's AI Assistant Might Just Be the Future of Work

Task-Executing Assistants

With LLMs being able to reasonably mimic the diction of an average person and task-executing agents seemingly right on the horizon, AI assistants are gaining popularity. As more of our internal communications and corporate data are exposed via streamlined APIs, these assistants can be prompted to access key data to perform (at first) basic tasks on behalf of users.

There are two main strategies we have seen to date. First, there are AI-native assistants that start with basic tasks, such as scheduling meetings. These products are largely activated via email prompts and are, for the time being, more limited in scope in what they can do for users. Second, there are virtual assistants who leverage AI tools to perform a greater range of tasks more quickly. While both strategies start from different places, we believe they share similar goals: to gather idiosyncratic data on users’ preferences to ultimately offer them a highly tailored AI assistant that can act as an extension of the user themselves. 

The two sets of strategies also target different demographics. People who are accustomed to interacting with assistants and want white-glove service might opt for the “safer” human-in-the-loop option (Athena charges thousands of dollars a month). Those who have historically not had access to an assistant or are more price-sensitive might opt for the fully automated offerings (Skej charges $10 a month).

Ultimately, we think that these businesses can be powerful conduits for the maturation of agent technology. Tool connectivity will shift the power towards companies who directly own the customer relationship, something that stands to benefit Task-Executing Assistants. As consumers become accustomed to AI assistants completing tasks that are feasible with today’s architecture, these assistants can eventually “earn the right” to handle more extensive tasks in the future, such as booking travel, ordering gifts, or other ancillary tasks that people may want to offload.

What’s Interesting About This Direction

• It taps into existing workflows: People are used to interacting with email and chat. Using those as mediums to generate tasks is a straightforward method for deploying this product.
• They are perfect conduits for continued agentic advancement: What consumers seek are good Swiss Army Knives. As AI models and tools mature and become interchangeable, the execution of individual tasks will also improve. But to usher this in, they seek a flexible device that can determine which model or tool performs each task best. The corkscrew or small knife within the Swiss Army Knife might change, but they’ll keep that orchestration layer itself.
• Text as the initial medium of communication should limit the surface area of potential mistakes: By using text as the way these AI assistants are interacted with, these companies can control the instructions fed into the “goal” agents in a way that should limit hallucinations and mistakes.

Open Questions

• What are the killer use cases outside of scheduling that these companies can grow on top of?
• What sort of leeway will customers give these assistants if they make mistakes? How quickly can the companies iterate to ensure people don’t churn?
• How will the AI assistant companies monetize task completion?

Context-Aware Systems

Advancements in transcription and voice AI have given rise to a new class of tools that process audio data from meetings, calls, and conversations. These tools primarily focus on summarization and transcription, making spoken content more accessible and searchable.

Companies like Gladia are developing infrastructure to support these capabilities, offering real-time, multilingual transcription services for various applications. Similarly, Hume AI is exploring the integration of emotional intelligence into voice interfaces, aiming to make interactions with AI more natural and personalized.

In the realm of verticalized solutions, Winn.AI provides a sales-focused assistant that automates CRM updates and offers real-time guidance during meetings. On the other hand, Amie emphasizes design and user experience, integrating calendars and to-do lists into a seamless interface.

While these developments are promising, it’s important to note that the field is still in its early stages. The majority of tools have not yet moved beyond enhanced (and pretty awesome) summarization and transcription. The potential for richer, context-aware systems exists, but realizing this vision will require further innovation and user adoption.

What’s Interesting About This Direction

• Underutilized Audio Data: Historically, conversations have been transient, with valuable insights lost after the fact. Capturing and processing this data can unlock new opportunities for understanding and collaboration.
• Improved User Interfaces: Recent tools offer more intuitive and user-friendly interfaces, making it easier for users to interact with and benefit from audio data.
• Emerging Infrastructure: The development of robust infrastructure, like Gladia’s transcription services and Hume AI’s emotional intelligence capabilities, lays the groundwork for more advanced applications in the future.

Open Questions

• What are the compelling use cases that will drive adoption beyond transcription and summarization?
• How will users respond to systems that attempt to interpret and respond to emotional cues?
• What are the privacy and ethical considerations in capturing and analyzing conversational data?
• How can these tools integrate seamlessly into existing workflows without causing disruption?

We’re incredibly excited about what the future of AI-enabled assistants has in store. The products that exist today are the worst they’ll ever be, and yet are already incredibly powerful. If you are building anything that intersects with the theses we outlined above, please reach out! Molly, Nick, Naseem