Where Is the Long-Term Value in Your AI Startup?

By Roy Lowrance, AI advisor at Blumberg Capital

As part of my advisor work for Blumberg Capital, I review business plans from AI-native startups. These are often based on a variant of this idea: find a business process that can be made more effective or efficient by automating the processing of information in complex documents, then create an automation platform that improves outcomes. Many of these businesses can be separated into two businesses: one that parses the required information out of the documents and puts it into a usable form, and another that uses that information as part of a workflow-support application. Understanding where long-term value is created can help founders make better products, partnerships, and fundraising decisions.

Consider a representative example. The X company proposed to parse emails received and sent from attorneys. They will identify the matters discussed in the emails and pull out relevant information such as open tasks. That information is passed to a workflow system. Initially, the system provides summaries of actionable work items to the attorneys. Later, it can propose actions and documents that the attorneys can approve. This example can be generalized to include any type of worker that processes documents.

Capability #1: Information Extraction

Notice that there are two difficult-to-build technical capabilities required for the startup to succeed. First is parsing the documents to extract relevant information. That requires understanding the domain. In the example, it is legal work, but we see many other domains. It also requires extracting relevant information from documents. This is relatively straightforward with emails, more difficult with PDFs, and perhaps hard enough to require deep learning with image and audio files. It also involves creating an ontology for holding the relevant data and populating instances from the documents. This capability is one that our firm X would consider purchasing or renting, if it were available.

Founder takeaway: Ask yourself whether document understanding is your long-term competitive advantage, or whether it will eventually become infrastructure that your company builds upon.

In X’s domain, legal work, there are many firms needing this capability, and eventually a company–say LegalDocumentParser (LDP)–will be built to parse all the relevant documents–not just emails, but contracts, court proceedings, and so forth. X will want to buy from LDP if it covers more document types of relevance to X’s supported workflow or if LDP’s parsing is much more accurate. Both of these conditions are likely eventually if X’s domain is large enough. If X’s domain is sufficiently large, the LLM providers will make this capability available.

Strategic question: Which capabilities should founders own, and which are likely to become platform services over time?

Capability #2: Workflow Automation

The second capability is building the workflow engine. That requires understanding how people currently do the work and their willingness and ability to adapt to new workflows. For example, in the X example, attorney’s will probably accept lists of prospective actions for them to consider, but whether they will accept having those proposed actions loaded into a work-tracking system is much less certain. Possibly they will accept reviewing work products developed by AI agents–that is a growing trend–but perhaps in the longer term, if creating work products shifts entirely to AI, employers may conclude that it deskills the workforce and adjust accordingly.

Founders are presently seeing two sources of value when startups formulate their business plans. The two parts of the business are bundled together. However, over the medium to long term, these capabilities may evolve into two distinct businesses with different sources of value and competitive advantage.

Founders can consider two development scenarios for companies like X. In one scenario, X purchases the parsing and knowledge extraction tool, so that X’s value depends upon delivering superior workflows. In another scenario, X splits its business into two parts: one for each capability, seeking to create value through two sets of IP and related competitive moats.

Questions for Founders

• Where does your long-term competitive advanage reside: information extraction, workflow automation, or both?
• If foundation models continue to improve, which layer of your business becomes more valuable?
• Which capabilities should you own, and which are likely to become infrastructure?
• How would your product strategy change if best-in-class document understanding become widely available?

Read this post and more from Roy on his blog