There are Now Three Businesses in Vertical AI

One of the things that feels off to me right now is that we’re mapping old categories onto new AI businesses and acting like that tells us something definitive. It probably doesn’t.

Some of these new vertical AI companies look like software but operate differently. Some rely on humans in ways that, five years ago, people would have dismissed. Some are really services businesses, just powered in a new way. None of these approaches are necessarily better or worse than the others, but they require a rethinking of how we assess these types of businesses.

So before we rush to judgment, I think it helps to start with a simpler framework: there are three broad types of businesses in vertical AI. And founders should probably know which one they are. This pattern carries across sectors, whether you’re building in wealth management or legal or architecture. This distinction is based on the need for the models to learn and to get feedback, and the learning comes from. Traditionally software didn’t have this distinction, you were always delivering the service (the last three letters of SaaS).

Vertical AI companies need a human-in-the-loop (HITL) to refine the outcome and improve the experience. No doubt, forward-deployed-engineers (FDEs) can help in implementation, orchestration, and fine-tuning of a model. Nevertheless, there are three types of vertical AI companies at this point: those where the customer is the HITL, those where there is a HITL but abstracted from the customer via the product, and AI-services where the product is the HITL.

To make this clearer, let’s give three examples in the Scale portfolio:

Customer is HITL. GC AI is the leading solution for in-house legal teams, used by over 1,500 companies worldwide and expanding rapidly. Users can review or draft legal documents in a conversational interface like ChatGPT. As part of this flow, those users can provide feedback on the agents by providing follow up questions, comments, or the classic thumbs up or down experience. The customer is not aware, but they are the HITL in the product that is refining and improving the performance of the AI and agents.

HITL Bundled into Product. Monto is the largest provider for Accounts Receivable teams to get paid quickly but automating invoice submissions into payment portals. Users can connect their ERP to Monto, and immediately see their invoices get submitted into a complex long-tail of complicated payment and procurement portals. On the back end, Monto’s agents can extract the information from invoices, log into portals, authenticate the invoice to a purchase order to submit, and then track the payment flow. Despite the advances in AI, payment portals, invoices, and purchase orders are littered with exceptions that Monto must handle. To deliver complete coverage, Monto has a very limited HITL process that handles exceptions and uses those to further train the agents. It is likely that most customers do not even know that Monto has this built in, as the experience is seamless.

AI-Powered Services. Range is a category creator in wealth management, delivering tax planning, estate planning, and financial planning to consumers that was previously only available to the ultra-wealthy while not charging any management fees. Customers have wealth advisors and tax planners they can interact with that are real humans. Trust is incredibly important in this category, so the delivery of AI via a service makes the product that much more compelling for the customer. It also allows the company to address a comprehensive service, while allowing their advisors to leverage AI to deliver most of the experience. The educational savings plan a customer perceives to have been developed by the advisor they’re interacting with may have been completely generated by an agent, and simply reviewed by the advisor.

‍The Variables Involved

There are different variables to consider in choosing between these business models:

Speed of delivery. When doing diligence, I usually ask about this, as it is the tell-tale for what you have built. If it needs to be instant or is instant, then the customer is almost always the HITL. If the product is not instant but delivered in minutes, then HITL is usually bundled into the product. Interestingly, a lot of business processes agents will take over do not need to be instant, which is something more startups should consider and exploit. AI-services are almost always in the timeframe of hours or even days.

Cost. This point is intuitive to most, which is that the cost of delivery is cheapest if your customer is the HITL, and most expensive if you are delivering an AI-service as you need to employ the most amount of humans to deliver the product. What isn’t considered as much is the frequency of engagement as the driver of cost. Said differently, if your customer engages at a low-frequency you can bundle in HITL or do AI-services, but if your customer wants high frequency of engagement, you either need to charge a lot for those interactions (as they are expensive), or have them be the HITL.

Cost of Error. We all deal with hallucinations in ChatGPT because the experience is so superior most of the time that we have a good tolerance for false positives. However, certain business processes have an incredibly thin margin of error, and therefore a HITL either needs to get bundled into the product, or an AI-service needs to be developed. In architecture, during the concepting phase error tolerance is high so you could develop a product where the customer is HITL. But for final plans, you would need to deliver such a product through AI-services as the cost of a building collapsing is just too high.

In vertical SaaS, we only had one business model to worry about, which was differentiated by customer size and number of professional services bundled into delivering the product. While we had some attempts at tech-enabled services emerge, almost none of these were meaningfully successful. The relative value to cost of delivery the experience was too tough to balance.

Too Early to Call a Winner

We are in the early years of vertical AI and there is a large volume of debate on which business model will be supreme. No doubt, the breadth of what the foundational models can do will ultimately have the most impact on these outcomes. If agents can really do everything in the future than this whole conversation is moot.

However, my early hypothesis is that this will not be the case, and therefore we may have a broader set of success stories than in the previous software regime.

I suspect that some of these business models will have more success, will be easier to scale, and will have varying profitability structures and therefore valuations. I think it is too early to call one the better or best, which is why we at Scale have chosen to invest across the spectrum as we find the right form for the right vertical market.

If any of this resonates, or if you think I missed something, I’d love to hear from you.

**An astute reader might point out that Bundled HITL into the Product has been shrinking rapidly as we’ve seen the models improve and may shrink into irrelevancy. I might agree and observe that the companies that can grab customers today while keeping up with what the foundational models can produce will evolve into Customer as the HITL. This may also end up hyper-dependent on vertical market segment.

***Any review process may be able to command premium pricing and build moats through proprietary data generated by operators — which could create premium pricing structures negating the cost argument.