We believe every white-collar role will have an AI copilot. Some of these roles will be fully automated with AI agents.
While incumbents are often slow to respond to changes in technology (there’s a reason why the average tenure of an organization in the S&P 500 has fallen from 35 years in the 1970s to less than 20 years today!), the most natural place for these copilots and agents to live is the incumbent workflow or system of record (e.g., sales agents launched from Salesforce). The system of record (SOR) is where the data agents need to complete specialized tasks lives, and it’s also a natural launchpad for any new user interface to reside (e.g., prompting the agent).
So if a startup wants to build a large copilot or agent company, how can it overcome the natural advantage of the incumbent?
1. Insert at the data collection stage *upstream*
Data for loans or insurance policies is still often collected via email and PDFs. A startup could “AI-ify” this workflow and own the data before it gets to the incumbent SOR. For example, a virtual loan officer or insurance agent (like Cascading AI’s “Sarah”) could own the initial back-and-forth customer document collection and appointment scheduling. Similarly, virtual sales development representatives (SDRs) like 11x.ai can gather all the information about a potential customer and own the initial correspondence before a record is even created in the incumbent SOR.
Startups should look for tasks like data input and entry, scheduling, and back-and-forth correspondence as wedges.
2. “AI-ify” a painful workflow performed *outside* of the incumbent
There are few things more tedious than the Know Your Business (KYB) onboarding process in banking, which involves document checking, internet searching, and back-and-forth correspondence between businesses and financial institutions. Companies like Parcha.com will auto-parse every document that’s uploaded, extract the needed information, and follow up with the customer for missing information. Healthcare is another industry with plenty of painful workflows. Tennr will take in every medical document hitting a fax machine, extract patient and diagnosis details, and even run insurance pre-qualification to streamline patient visits to medical practices.
By solving a painful workflow, startups can become the repository of data and earn the right to automate further workflows.
3. Integrate disparate data sources to create a *new multimodal* system of record
Significantly more data exists and is relevant to the job to be done than what is currently held within incumbent SORs. For example, sales data doesn’t just exist in Salesforce or Hubspot: there are also emails and Slack messages, sales enablement materials, product usage data, customer support records, news and financial reports etc. By integrating these data sources, a newco could pull from more comprehensive data than the incumbent. For example, companies like Pylon aim to be the customer’s SOR for fast-growing B2B companies by providing a single view of customer issues.
With LLMs, startups can build new SORs that can be entirely unstructured and multimodal, constantly ingesting text, image, voice, and video data to create the most up-to-date context.
Copilot and Agent opportunities abound across many professions
A recent study by OpenAI and the University of Pennsylvania found that with access to an LLM, about 15% of all worker tasks in the U.S. could be completed significantly faster at the same level of quality. When incorporating software and tooling built on top of LLMs (i.e., Vertical SaaS), this share increases to between 47% and 56% of all tasks.
Inspired by this research, we pulled employment data from the U.S. Bureau of Labor Statistics for 2023 and identified the top 50 roles where 50% or more of the tasks could be performed by AI. That said, we think the longtail of jobs that we couldn’t include have just as much potential for AI copilots and agents, if not more, including the country’s 48k brokerage clerks, 44k switchboard operators, 37k word processors, 25k actuaries, and 52k medical transcriptionists, to name a few.
While this list is not mutually exclusive or exhaustive, we hope it sparks your imagination!
Working on turning one of these roles into software? We’d love to hear from you.
Thank you to Amit Kallakuri for his support in this data analysis.
