Investing in Early-Stage Enterprise AI: Starting with the Application-Layer
Venture capital interest in AI for the enterprise has shown no signs of slowing down— in fact, of the last 15 Seed and Series A B2B deals that we evaluated in the past week or so, 13 of them were in AI. This comes as no surprise as the media has been buzzing with news of enterprises across all verticals conducting AI experiments or launching their own AI products — GitHub’s Copilot and Intel’s vPro platform products have made headlines in the past 24 hours from the time I’m writing this.
Witnessing the beginnings of enterprise AI adoption and experiments firsthand (via Bertelsmann, our fund’s sole LP) has made some of the VC hype around this technology feel very real. It certainly feels eons more real than during the web3 days of ’21 and ‘22, when we sat tight during slow sales cycles and initiatives that fell from the purview of executives who were once so keen on obfuscating the middleman (and often their own orgs).
But being a real(ly exciting) technology with legs and feet and being appropriately valued by the markets are two very different things. I’d argue that venture investors who focus on B2B have been sufficiently swept off their feet by the undeniable “wow” factor of the content coming out of these foundation models. The issue is, however, that when the intrigue comes from the unbelievable content deriving from models like OpenAI’s Sora or MuseNet, the “wow” effect is a consumer stated one and not coming from the enterprise level.
So, before we project any residual consumer-level excitement onto the enterprise (which of course does not nearly care as much as we (I) might about creating short-form videos that look like our dreams), I want to level set. Given the sheer amount of venture capital that has gone into both the foundation layer and the space in general, I think that it’s time to swing back the pendulum and focus on where we might see real traction take shape in the coming months in B2B.
What are the genuine applications of AI for the enterprise? There seem to be very many both in the near and long term. In my view, only once we have a clear thesis around where usage will occur can we predict where the value will accrue across the many facets of AI and make bets accordingly.
It seems like useful AI applications will fall into one of these four buckets of organization-level productivity via complex language generation and understanding:
- Monotonous task reduction. Putting an end to high-touch, repetitive tasks that should be no or low-touch. (note that this bucket likely sees more PLG / employee-level adoption first)
- Known output generation. Generating content or any deliverable that is pure execution on what is known. In other words, you already know how the slide deck or contract should look — you just need to create it. (seeing substantial activity in procurement and legal here)
- Knowable but unknown output generation. Creating the idea from a dataset that the employee can know but is difficult to derive insights from manually. Beyond that, creating and delivering the content that supports this idea. (seeing substantial activity in advertising and ecomm enablement here)
- Unknowable, unknown output generation. Creating the idea from a dataset that the employee can’t possibly know. Beyond that, creating and delivering the content that supports this idea. (seeing substantial activity in healthcare and pharma here)
Bullet numbers two and three above are the buckets in which I see the near-term value accruing at the early stages and where I’m spending most of my time today. Some of the strongest and earliest monetizable products that I’ve seen to date include vertical-specific contract generation in bullet two and highly-optimized ad creative in bullet three. SMBs and enterprises today are willing to spend on “pre-tested” ad creative given how quickly orgs are able to analyze the results of these campaigns and verify the efficacy of the technology.
In general, it seems that organizations will take time to leverage and trust technology entirely in creating product output, and 5+ year exit timeline horizons map to loose predictions around time-to-scale within an org for bullets two and three.
Bullet number one will likely look like what pessimistic B2B investors would describe as a “feature not a product” and scale through PLG. I think that employees are overripe for co-pilot tools and that these businesses can be rightfully invested in at later stages.
Bullet number four will take by far the longest to scale in my view but is interesting today for companies within industries that appropriately reward discovery (e.g., healthcare and pharma, legal).
I’ve included a visual representation below about how I’m thinking about the application layer — namely, I think about it on a “knowability scale” where the content output is a varying degree of known or unknown. I’ve included eight startups that I’m watching closely within each of these four loosely defined buckets of application-layer enterprise AI.
As I spend more time in the space, I affirm that B2B AI applications are just B2B applications, and traction matters more than anything else at the end of it all. The proof really is in the usage stats.
If you’re building something that employees are using or probably / definitely want to use, let’s chat. If you’re building something that I might want to use, as someone who writes a lot of emails, maps markets, and scours the web, I’d love to try it out.
Also, if you’re in NYC and work at a company that is interested in using or purchasing AI software, I’m co-hosting a meetup with Inspired Capital for heads of marketing at big orgs. Feel free to send me a message if you’re interested in joining at the end of March — I’d love to hear your perspective.
