A structured way to think about use cases for AI products
Enhance -> Replace, Correctness -> Creativity
This is the first of a three-part series on building AI products.
Use case matrix for AI products ← this one
UX/UI for AI products
Business models and pricing for AI products
Scroll to the bottom for more resources and companies mentioned.
AI is at the peak of the hype cycle. While experimentation and tinkering is a good thing, it’s important to start with real problems to solve rather than using AI for the sake of building an AI product.
So here’s a matrix to offer a more structured way to think about meaningful use cases when building AI products, especially in B2B SaaS. The two dimensions are:
How much of a workflow can AI enhance or replace?
Does the outcome optimize for correctness or creativity?
Correctness —> Creativity
The artist may have a number of goals and motivations, but there is only one intention. Rarely if ever do we know the grand intention, yet if we surrender to the creative impulse, our singular piece of the puzzle takes its proper shape.
- Rick Rubin, The Creative Act
“Correctness” represents how much the AI fulfills the human’s intention, whereas “creativity” represents how much AI comes up with new intentions you didn’t even know you had.
Correctness naturally matters more in regulated industries such as healthcare, financial services, law, and self driving. The model needs to learn from existing data on best practices to complete a task. The success of the outcomes is clearly defined and we don’t want to deviate from the defined outcome.
Creativity is obviously valuable for functions that involve “creative” work, where the outcomes are open ended. This could be general knowledge work, design, programming, entertainment, and more. Hallucination can be a feature, not a bug.
Enhance —> Replace
On the Enhance side of the spectrum, the human drives the workflow, AI is embedded in the workflow to level up human's creativity and productivity. It’s about getting unblocked right away to stay in flow-state longer.
On the Replace side, AI completes the full workflow, the human gets moved to the end of the loop for review or verification. These are any job that has a workflow that can be outsourced, ie. has steps that can be clearly defined.
4 categories of use cases for AI
⚙️ 1. Recommendation
These products enhance the current workflow with a high level of correctness. This is the most obvious and common category of use cases often adopted by regulated industries. These were early application of AI (before LLMs), optimizing or automating workflows that have defined best practices and outcomes. But now with LLMs, the level reasoning and orchestration can be so sophisticated that more and more of the workflow can be replaced.
So what? If you own specialized data and best practices in a regulated industry, there’s probably low-hanging use cases in this category to be a wedge to your AI strategy. However, long term, use cases in this quadrant will likely to move to the “Professional services” quadrant over time.
Examples:
Subtle medical. Generative AI models to improve quality and speed of radiology imaging.
Sardine. Predictive fraud prevention.
Coactive AI. Analytics from unstructured image & video data
🤖 2. Assistant
These products enhance the current workflow with openness for creativity. This is currently the most popular category of AI products. Wrapping a LLM with UX components to assist users with cognitive tasks is a low-hanging fruit. It’s a natural use case for productivity/collaboration tools, where the LLM can quickly unblock knowledge work and precision is not super important.
So what? GPT embedded in an established workflow tool’s UI to level up productivity will become table-stakes for productivity tools. There’s opportunity to reimagine a new paradigm of UI for LLMs to differentiate the product. But “wrapper” products on their own are not defensible as OpenAI continues to release native products such as custom GPTs.
Examples:
Lindy. Create AI employees that can work together to perform any task. No code is need, just natural language to set up the agents and connect apps they’d need for the job.
Shopify Sidekick (Magic). Co-pilot for e-commerce.
Microsoft Co-pilot. Level up productivity within the Microsoft suite.
👩⚕️ 3. Professional services
These products can replace a workflow with a high level of correctness. With significant data in a particular field to train a ML model to consistently and reliably imitate the best human behavior, the product can deliver an end-to-end service that currently requires costly human professionals. This can be healthcare, law, customer support, transportation, and more.
So what? This group requires significant upfront investment in data and model training since correctness is important. It’ll also have the biggest second and third-order economic impact, because it’s selling the “work” product rather than productivity gain. The service is the product, rather than the tool that a human works with. It can be priced relative to the cost of humans.
Examples:
Even up. Turns medical docs and case files into AI-driven demand packages for injury lawyers.
Medical billing - ? This is ripe for AI to replace. There’s immense data on billing codes and encounter notes to create a model to replace billers, but I haven’t come across a product/company yet. Let me know if you have.
Ambience healthcare. AI medical scribe captures patient-doctor conversations summarized into encounter notes.
Ambient AI. Proactive physical security, detects emerging incidents with threat signatures.
💡 4. Conceptualization
These products can replace a workflow with openness for creativity. The model is allowed to generate new ideas with less constraints, and then complete a full unit of work. Imagine, running an AI town with generative agents to conduct social experiments, or using a video generation tools to produce an entire movie or TV show.
So what? This group is going to be the most disruptive and innovative. With reinforcement learning with human feedback (RLHF), experimentations in this group can give rise to new “intentions” which can then be refined by humans to turn into new services that never existed before. There are more inconceivable use cases here to be discovered and new businesses and business models will be born.
Examples:
AI town. A fascinating experiment with generative agents that act in believable human behaviors. The research paper here.
Runway. Generative video
Synthesia. Text to video
Galileo AI. UI generation
Insitro. Predictive model for expensive drug R&D
FigJam AI. Brainstorming/white boarding. (This can also belong to the “Assistant” category but the specific feature of generating a board puts it in this category.)
What does this all mean?
More products will move from “enhance “towards “replace”.
As context windows get bigger, costs get lower, ability to codify knowledge and practices get easier, more complex work can be carried out by LLMs at faster rate and lower cost by orders of magnitude.
Jevons Paradox makes an optimistic argument about “replacement”: when demand is elastic, as price drops for a resource, the demand and consumption of the resource will also increase. So even with more AI replacement of human work, people will consume more and there’ll be greater increase in productivity and new job opportunities.
It'll be easier to become generalists.
The barrier to specialist knowledge is significantly lowered by LLMs. In the context of a tech company, roles will cross boundaries/blend into each other. eg. engineers, designers, product managers, marketers, can all do some of each other's work.
Experimenting with AI is a good thing, but also try to start with problems to solve
On putting AI products out there for people to tinker:
“We don’t know exactly what the future looks like, and so we are trying to make these tools available and the technology available to a lot of other people so they can experiment and we can see what happens.”
Reminder to start with use cases that actually solves user problems (from the FigJam AI launch):
… all of these user problems existed in FigJam independent of AI. AI just happens to be an effective way to solve them.
Does this categorization of use cases for AI products make sense to you? What new use cases are you most excited about?