How are companies ACTUALLY using AI today?

An attempt to separate reality from buzzwords

As AI stories continue to dominate news cycles, sometimes “AI” can feel like an overquoted buzzword. The term Artificial Intelligence covers a wide swath of technologies / techniques (e.g., Deep Learning, NLP, Computer Vision, etc.) which can each be used for a wide variety of use cases solving an endless number of problems. Given this broad definition, sometimes the term can be misleading, or even meaningless.

What actually matters, though, is what AI is being used for, and what problems it solves. Given the timely announcement of Graft’s fundraising round and product launch yesterday, I thought I might share my mental model around how I bucket the various use cases of AI below (of note, this is a different framework than what the layers of the Modern AI stack look like, which we discussed at length last month).

To simplify, I think it’s easiest to begin with the most straightforward, and move into the most abstract:

Bucket #1: Productivity Tools

Starting with the most obvious category, AI is expected to make human beings significantly more productive. We can see this happening in both consumer settings (e.g., using ChatGPT to help plan an itinerary for a trip to Italy) as well as in enterprise settings (e.g., using ChatGPT to generate relevant marketing copy).

These tools rely on an AI’s ability to both understand existing information and generate new information. As a concrete example, Radical PortCo Hebbia, is an enterprise search business. It enables knowledge workers to quickly upload a large volume of files (e.g., a private equity dataroom), query that dataroom with natural language (e.g., “what are median sales per square foot”), and then return a simple answer in plain text. In this workflow, Hebbia’s AI first understands all of the uploaded information, and then generates a novel summary answering a user’s question.

Graft represents another offering in this category: it consolidates best-in-breed offerings across the AI landscape and allows individuals to use them via a simple platform, with no coding or ML expertise required. In other words, it allows non-technical users to leverage modern AI in their workflows (e.g., customer analytics, pricing, demand forecasting, etc.), allowing business analysts to generate deeper insights, faster.

TL;DR Productivity tools allow people to do more with less by allowing AI to abstract away a lot of the brute force / labour intensive tasks.

Bucket #2: Creativity Tools

Next on my list are creativity tools, which are forms of AI that help to solve the “zero to one” problem (which is often both frustrating and time consuming). While there is some overlap with the prior category, I generally think of these as a different bucket of tools. Often, creativity tools are used to inspire human creation, or create completely novel things altogether (e.g., AI art being designed by Stable Diffusion / Midjourney).

While some of these exist in the realm of text, most tend to apply to more visual applications, such as image generation and editing. Again, looking at a concrete example, Adobe’s Firefly allows artists / creatives to create new images from scratch (similar to Stable Diffusion or Midjourney), and / or quickly edit existing ones using simple text prompts (e.g., “remove the clouds in the background of the image”).

While AI-generated images often aren’t quite good enough for production today, they’re often used as a starting place for artists, or used to inspire new pieces altogether. Furthermore, by making reducing the technical barrier to create art, they also allow less technical users to create visualizations or their own, or tweak existing images to their liking (sounds like never learning how to use photoshop will finally pay off for me!).

TL;DR Creativity tools reduce friction in creative processes, both by inspiring new creations and by enabling more users to explore their creative visions.

Bucket #3: Embedded AI

As we move deeper along the complexity curve, we arrive at “Embedded AI” solutions. If AI is “embedded” into an existing product, it means that it is improving the functionality of that product. As we look at the recent explosion of LLM usage (e.g., via Cohere or OpenAI), much of this is driven by embedding language capabilities into existing products and services.

Looking at Oracle’s recent partnership with Cohere as an example, Oracle is working with Cohere to embed language AI directly into their ERP and Database offerings, so that customers can leverage LLM functionality directly in the products themselves.

One could imagine LLMs being used to improve the entity matching in a database, thus improving the accuracy and cleanliness of the various entries (e.g., LLMs can ensure that a minor naming typo doesn’t allow large amount of data to fall under a different entity). As an LLM is embedded into existing product offerings and businesses, these products are “powered” by an LLM, similar to how large software businesses are “powered” by AWS, GCP, or Azure on the back end.

TL;DR Embedded AI makes existing software applications more powerful by allowing them to leverage AI-powered insights inside of their offering.

Quick aside on Probabilistic AI vs. Deterministic AI

There is an important nuance to break out here as different “types” of AI are used by businesses. At the risk of oversimplifying, the term “Modern AI” (think LLMs) typically refers to “probabilistic AI”, in which AI generates a probable possible answer to a subjective question.

In contrast, the term “Classic AI” or “Classic Deep Learning” typically refers to “deterministic AI”, in which AI generates a determined answer to an objective question. While LLMs are the talk of the town these days, classic deep learning / deterministic AI has been powering software products for years (e.g., Netflix’s content recommendation algorithms and Uber’s ETA estimations both leverage deterministic AI).

As one can imagine, both types of AI have their uses, and both can be embedded into products in different ways. Typically, probabilistic AI is used to solved unstructured data tasks (e.g., free form text) and deterministic AI is used to solve structured data tasks (e.g., spreadsheet data). The best companies will find ways to embed both types of AI into their product offerings (e.g., by improving demand forecasting and then accurately explaining what changes were made in plain text).

Bucket #4: Autonomous AI Agents

Lastly, in what is potentially the buzziest category in all of AI, we have Autonomous Agents (sometimes referred to as personal assistants). While these don’t truly exist in production yet, the promise of Agents is that they will eventually go out and execute on commands themselves, directly. While productivity tools make existing workers more productive, agents offer the possibility of removing human intervention in some workflows altogether.

One consumer facing example of an agent use case revolves around booking a vacation. One can already ask ChatGPT to plan an itinerary for an Italy trip, but that simply returns a text-based itinerary. An agent would ACTUALLY go out and book the full trip for you (as in, literally book the flights, enter your passport number, check you in on time, pay for the flights on your credit card, deal with ambiguity in real-time as decisions need to be made, etc.), without the need for you to be involved whatsoever (could AI agents ironically represent the return of travel agents, albeit in digital form?). In other words, you can just show up.

In enterprise settings, the promise of agents gets even more exciting. What would a Partner at law firm pay for an infinitely scalable army of associates who can complete deep and thorough research for them 24/7 with minimal direction? I’d imagine, a lot.

The jury is still out on whether or not AI Agents end up being entirely controlled by Big Tech providers, horizontal startups, or end up verticalized by industry, but they will undeniably be the holy grail of AI R&D over the next few years. Even in the face of fierce competition from Big Tech, there are already a lot of promising startups looking to tackle this emergent category.

The impact of real and reliable AI Agents on the global economy would be undeniably massive, and likely requires a reworking of the both the global internet and our current economic infrastructure… but that’s a topic for another day.