For the past 25 years, application software startups have had a singular focus: increasing company and employee (including developer) productivity. This looked like building software that increased productivity at the employee level, increased collaboration across employees and teams, and/or enabled better oversight and management at the leadership level. More often than not, this software has been priced on a per seat basis, in essence benchmarked against the cost of the headcount itself and increasing that headcount’s productivity.
Enter Large Language Models (LLMs). The first tranche of products and startups leveraging LLMs has kept within the mental model of selling software to achieve step-function improvements in end-user productivity. The "Copilot for [x]" trend reflects this mental model. While there are fantastic startups innovating to improve employee productivity, LLMs create an opportunity for startups to look beyond this way of thinking and discover surface area that previously was out of bounds for selling software given the required GTM and pricing limitations of software. To do this, rather than sell software to improve an end-user's productivity, founders should consider what it would look like to sell the work itself.
Selling work opens up new vertical opportunities that wouldn’t have otherwise supported a software company. Take EvenUp as an example (who I have no doubt will dominate their vertical). If you are a personal injury lawyer, a work product you create on behalf of a plaintiff is called a demand package. Essentially the demand package is a summary of the case, the medical costs of the injury (including lost wages), and then a recommendation on the settlement value from the defendant’s insurance company. Law firms have stretched lawyers, paralegals, or outsourced groups writing these documents.
If you were still in the mindset of selling software, you could imagine a software offering for personal injury law firms, sold on a per-seat basis, that uses AI to help people in the firm create a demand package (imagine a builder where you drop in the medical records during one step of the process). But then EvenUp would have been stuck in the paradigm of selling software – selling a 10% productivity improvement instead of 95%. Instead, EvenUp had the foresight to sell the work product itself: the demand package.
When you sell work, the sales cycle is different, it’s priced relative to the cost of a human performing the work instead of as a productivity improver, and the competition for a similar product (besides the company’s own human capital) is essentially outsourced groups internationally. An AI-driven product with the consistency and SLAs it can achieve should be vastly superior to an outsourced offering – a 10x and cheaper opportunity. Indeed, I’d guess a good test of the viability of a market opportunity to sell AI-built “work” is, crudely, whether there already exists a focused, outsourced group internationally to support it.
For example, just to spark some ideas, here are some of the BPO services for an outsourcing provider:
And here are some examples of Legal Process Outsourcing options:
Any of these, I imagine, are vulnerable to automation leveraging AI.
As I write this, I’m aware of the 3rd rail I am touching – the fear that AI will replace humans over time. Here, I look to EvenUp as an example. When the lawyers and paralegals are freed up from putting together demand packages, the customer (the plaintiff) benefits from better demand packages, and the people in the firm benefit from being able to shift their time to less mechanical things like client services, getting more customers, or of course the finishing touches on the demand package.
If you see any other examples of this, I’m all ears. And if you are building a company leveraging LLMs to sell work, I’d love to hear from you. My X DMs are open, and my Benchmark email address is sarah at benchmark dot com.
Also, it’s not enough to sell work — you must escape competition. To this end, I’d also suggest my prior post How to escape competition -- Building enduring application-level value with LLMs, so you don’t fall into the trap of providing a service that gets all the margin squeezed away.
This is so smart.
I just keep reading and rereading it.
So smart.
I think my experience at Notarize is relevant here, not because it is AI but because it offloads a core task (i.e. a mortgage closing) from someone. We "sell the work" by offloading the task (albeit to other people, not to AI) and allow our customers to repurpose employee time to other accretive tasks or to alter their staffing models. Much of our ROI story is rooted in that outcome. The primary challenge though is to actually get customers to adopt above a threshold required to change either their internal operations or staffing models. The beginning of a customer's adoption curve is the worst - they're forced to run bifurcated processes (actually costing more) AND the actual people you aim to offset are often critical in managing that transition. At Notarize, I've cribbed some thoughts from the medical and hospital industries which believes that a 30% adoption rate is required to see the ROI of actual process change, beyond which they can actually adjust the "standard of care" and make that new/better process the new normal. Getting to 30% is really hard and we've obsessed over doing just that, which is to convince our customers to make us the standard of care for mortgage closings, auto sales, you name it. I think it will be especially hard for AI in some of the industries you outline above, particularly legal/compliance/etc. Why? Everyone says LLMs drift, but that will surely be solved. Many of the things you described are considered the practice of law and people will need to adjudicate UPL claims. Fun!! I think the real issue is regulatory headwinds. Specifically, regulators are terrified of algorithms/machine learning/AI systems instituting global systemic bias. Take a property appraisal, which everyone agrees should be digitized. Regulators would rather local consumers connect with local appraisers to disperse and decentralize the bias - an African American consumer may get an African Appraiser and "win" on one and another might get a racist white appraiser and "lose" on another. That is better to them than a one nebulous system making judgements they cannot assess. And the government has no ability, mandate, or agency even to test these models. So how is a large bank who is constantly sued for unfair lending practices able to adopt these systems? If anything, recent advancements are showing any/all of these issues can be solved... but for AI to advance as you outline, it needs to think much more deeply about bias and how to instill the confidence required to take over from the humans who are obviously slow, but easier to regulate and more random in their outcomes.