Conviction Areas

Generative vs. Genuine: Why Today’s Generative AI Isn’t Tuned for B2B

URL Copied

2022 was a year that personified AI. Many of us had moments last year playing with ChatGPT and DALL-E where we saw the future. Some applications were more practical, such as students getting help with their homework or software engineers de-bugging their code, while some were more fun, such as asking the model for a new mac n’ cheese recipe or chatting with dead celebrities.

While these applications have strong potential for consumer audiences and new marketing techniques—as evidenced by the fast growth of companies like Jasper AI—we don’t believe that generative AI alone will have the same impact in most business applications. Why?

Generative AI today isn’t genuine; it isn’t consistently accurate, it isn’t outcome-tied, and it isn’t context-specific. While those limitations aren’t likely to be blockers for B2C, B2B applications have a higher hurdle to clear.

That said, generative AI will play a critical role in solving business problems. Paired with context-specific models and human influence, this new technology will create a step change in our ability to change the way the world works.

Generative AI has broad, generic value… and some risks

Generative AI is a foundational tool to help us with drafting—whether words, images, or code—and it has the reciprocal value of helping us condense and summarize. It will help us to save time and generally improve our communication, whether in our personal or work life. But if we rely on it too heavily, we’ll run the risk of sounding generic and, often, wrong.

Since 2016, we’ve called this potential risk of AI, “crutch software." Our Coaching Networks thesis has been based on the premise that “humans are the only mutation engine in the age of AI.” Generative AI will test our abilities and willingness to create and expand on its baseline or just rest on its recommendations.

With this context, let’s dive into the particular limitations of generative AI in business applications.

1. Generative AI is not accurate

LLMs such as OpenAI’s GPT-3 are trained to sound like human writers, but they are not trained to be accurate like human writers. That’s why ChatGPT can write a compelling, well-worded essay on the history of the Napoleonic Wars, but it can’t ensure that all of the facts in the essay will be accurate. Further, since LLMs are trained on a dataset at a specific point in time and then released, they are unable to incorporate more current data (e.g., for GPT-3, any content after 2021 isn’t included). Thus, asking it to write an essay on the history of the current war in Ukraine would yield poor results.

Generative AI today is just mimicking the trillions of words that it has consumed. Because models are trained to match the distribution of text on the entire internet—and not everything on the internet can be trusted as accurate—not everything generated by generative AI can be trusted.

A December 2022 piece by Mashable titled “The ChatGPT chatbot from OpenAI is amazing, creative, and totally wrong” detailed many of the instances where ChatGPT either got answers to basic-knowledge questions wrong or used its imagination to make things up completely. While most of these examples are low-stakes, you can understand why models like GPT-3 are riskier to use in business settings. This is why The Atlantic argued to “use [ChatGPT] as a toy, not a tool.”

Current business applications of generative AI are mostly tuned for marketing (copywriting/cold emails) and advertising purposes, use cases in which occasional factual inaccuracies are typically tolerable. But for most business use cases, accuracy is critical. In order for businesses to feel confident in using generative AI for most use cases, more context and human assistance will be required.

2. Generative AI is not outcome-oriented

Generative AI is output-oriented, not outcome-oriented, which works well for consumers but not for businesses. In other words, ChatGPT can spit out taglines for a new beverage brand, but it can’t tell you which one performs better. This is because the interaction with the model is a one-way street; it lacks the ability to continuously learn based on outcomes. When it comes to B2B, businesses need more than a generator; they need AI that is iterative and driven by outcomes specific to their industry.

Promising generative AI apps for B2B will anchor on ROI-based outcomes. For example, our portfolio company Ironclad is using AI to help draft and edit contracts more efficiently. This not only helps lawyers move more quickly; it helps them improve business outcomes. Their platform is being built to coach drafters on which clause formulations will drive faster deal close rates. By marrying LLM suggestions with their own proprietary data, Ironclad is building a defensible, outcome-focused product.

3. Generative AI is broad, not deep

In order for generative AI to move the needle in many business use cases, the AI needs to be trained on company-specific data. While off-the-shelf language models are mostly trained on publicly available data, today, they lack broad access to the context and IP needed to be effective for B2B. E.g., without the context-specific data created within Ironclad’s workflow software, an LLM can’t ascertain which clause is likely to close a contract the fastest.

We don’t anticipate the volume of proprietary data needed to make generative AI truly effective for most B2B applications being incorporated into LLMs in the near future (or ever). This is because enterprise organizations will likely be unwilling to share their most valuable IP with LLMs. Businesses’ competitive advantage lies in their proprietary data, which is why enterprise organizations are wary of LLMs having direct (or even indirect) access to their IP. Similar concerns originated in last year’s class action lawsuit against OpenAI, Microsoft, and Github regarding CoPilot. Corporations and large enterprises will be wary of contributing their proprietary data to these types of broadly distributed models out of fear of losing their competitive advantage. Instead, B2B applications will need to find another approach.

That approach could be a marriage between external LLMs and internally developed models and data sets. B2B applications could start by querying an LLM to generate an initial output. This could then be fed into a fine-tuned, internally developed model to refine the output for the specific B2B use case. These “Small Specific Language Models” (pardon the play on words) will leverage the knowledge graph that applications build from the proprietary data they are generating or have access to. By chaining together open-source LLMs and proprietary SSLMs (or building their own end-to-end stack), B2B applications will be able to deliver strong, defensible ROI.

Next-gen AI apps: From generic to genuine

So what does the future of B2B software look like in the age of generative AI? We believe it is leveraging the generic value of LLMs and coupling it with the people and intellectual property of your customers to generate accurate, outcome-tied, and context-specific results.

Every function in every company in every industry will be transformed over the next few years. Your job is to help customers extend and protect the core intellectual property at the heart of their business while integrating the new-found power of generative AI. We think it means leveraging the unique attributes of LLMs to build an outcome-tied app, but the powering model won’t be solely OpenAI or any of the other foundational models. Those LLMs will provide only a portion of the value. The future of generative AI for B2B will lie in a combination of these types of language models paired with proprietary data from human intervention and ingenuity.

We’ll leave you with this general architecture diagram to get you thinking about where you might want to build. Stay tuned for more details in our next piece. In the meantime, we’d love to hear from you if you’re building in the space. Please reach out to us at gordon@emcap.com or jake@emcap.com