Logistics Automation Winners Will Be Services, Not Software

A framework for AI-enabled services and why we’re building one in logistics.

It’s been a year since Sarah Tavel wrote about AI startups selling work, not software and as a firm, we are all in on this concept.?

The first wave of AI application startups went after horizontal services such as customer support, sales, etc. Now we are seeing vertical players emerge as AI-enabled business process outsourcers (BPOs) in industries where incumbent BPOs struggle to adopt automation because of an innovator’s dilemma. We think the logistics industry is a textbook case for this kind of an opportunity.?

Since the beginning of the year, we have been building an AI-enabled BPO for the logistics industry as one of our Primary Labs incubations. We will have more to share about that company in the new year, but the lessons from launching this business and the market’s incredible receptivity to its value proposition have accelerated our thinking about the realities of “selling work, not software” in industrial end markets. Our goal in this piece is to bring some of those realities to life for others building and investing in similar companies.??

Why logistics

Quick (we promise) primer on logistics and the role of BPOs here. Logistics companies sit between brands that sell physical goods and transportation companies that can move those goods. Instead of contracting directly with transportation companies, brands work through logistics companies that are responsible for orchestrating the movement of their products. In exchange, the logistics companies take a percentage of what is ultimately paid to the transportation provider. The orchestration work that logistics companies do is mostly back office repeatable tasks like customer quotes, data entry, appointment scheduling, order tracking, billing, and so on. This work accounts for about 60% of a logistics company’s net revenue and is largely outsourced to BPOs, some of which do hundreds of millions in revenue selling to US logistics companies alone. In total, there is over $100B spent globally on back office logistics work, most of which consists of repeatable tasks that are outsourced to lower cost labor markets.?

Why we’re sitting out of the AI App Wars

We decided not to invest in any of the AI apps selling into logistics providers and incubate an end-to-end services business instead. It was a hard choice. We know that there are going to be successful companies in this space. We also fully agree with this perspective from Sam Altman, which basically says that startups should act under the assumption that the models will always improve. This view could suggest that startups selling AI apps without human in the loop services as a backstop might be better positioned to win in the long term. If a model can’t fully do the job today, it will probably be able to do it tomorrow. That may be, but the reality today is that there is very little margin for error in industrial end markets. Though repeatable, each task is critical. Anything short of perfection just isn’t good enough in most cases.?

Point solution SaaS products have also been on the rise in the logistics space for over 10 years and most have struggled to scale beyond $10M of ARR. We think this is because they have to go after very niche tasks to actually deliver automation and, in doing so, pigeonhole themselves into smaller market opportunities where you have to squint hard to see a billion-dollar outcome. We think a lot of the AI applications in the logistics market will have the same problem and struggle to see how most of these companies can build a defensible, differentiated, and ultimately venture-scale-enabling product.?

Why BPO service, not software, wins

Now here’s why we’re excited about an AI-enabled service in logistics. Hopefully this framework is applicable to founders building similar businesses in other verticals:??

1. Service TAM > Software TAM: In logistics, spend on core systems of record claims about 1% of gross revenue. Meanwhile, back office tasks cost logistics firms about 5-10% gross revenue. By selling a service, we compete with the service providers who do those tasks today, not the software providers.
2. Low-friction GTM: Major logistics companies already have 8-9 figure contracts with BPOs, but limited loyalty to those partners. Selling into existing budgets for outsourced operations will reduce GTM friction and AI solutions with humans in the loop will earn trust at the enterprise level.The key is selling these customers a better and cheaper version of what they’re already buying, not a tool that they will have to spin up new internal capabilities to be able to realize the value of.
3. Incumbent innovator's dilemma: BPOs in the logistics space charge a markup of their human capital costs, not by output or delivered work/value. This means offering automation as part of their offering would cannibalize their revenue. Nevermind the fact that these companies generally don’t have a shred of product innovation DNA.
4. Aligned incentives: We monetize by output, which means we are incentivized to automate as many tasks as possible.
5. Sticky customer relationships: Incumbent BPOs have no tech stack connectivity with their customers, so it’s easy to swap them out. With a platform, we can actually integrate with our customers and allow them to plug into any AI-enabled service they want.?
6. Clear expansion opportunities: Since we aren’t reliant on a single application, we can start expanding into a new service or task before our product is great at automating that task. We will also have more data on different parts of our customers’ operations, which allows us to offer higher value analytics and optimization beyond workflow automation.?
7. Model agnostic platform: By owning the service and the user experience, not the model layer for our customers, we can be agnostic to the technology behind our service and consistently move to the best models and apps on the market.?
8. Data advantage: Since we’ll store each external facing workflow on one platform, we will ultimately reuse automation workflows across multiple customers, which reduces our COGS.?

Why customers want this?

We’ve been working through pilots with a handful of customers and have gotten a lot of great feedback on our offering. Like many industries, logistics is a category that is getting inundated with AI apps that promise to automate specific operational tasks. The term “AI-fatigue” is one we hear more and more often from customers. Customers complain that products over promise and under deliver. They ask, “Why should I pay you if I have to train your product?” Above all, they don’t have the time to keep up with the rapidly accelerating capabilities of AI and have no way of knowing if they are using the best of what’s available. In pitches and pilots, three things we do have resonated most with customers:

1. Service, not SaaS, means they don’t train the AI, we do. We differentiate ourselves from other AI products by talking about the “humans in the loop” we have as a backstop to our technology. We are responsible if the model doesn’t get it right, not them.?
2. Output-based is the right incentive. Customers want their opex to be as low as possible and the status quo of time-based pricing puts a floor on that. In output-based, both parties do better when we automate.?
3. Model and app agnostic. Since we charge by output, we are incentivized to use the best model possible, and build a platform that enables that. They outsource the noise of AI procurement to us.?

Of course there have been other learnings, but those are the big ones. We are in the early days of our journey here and expect to learn a lot more and will share what we learn as we build.?

Why we’re excited (and a little nervous)

We’ve been waking up in the middle of the night a lot recently, sometimes due to cold sweats, but more often to write down new ideas. The anxiety comes from a familiar source in this time of rapid AI innovation – the possibility of getting leap-frogged by something that we can’t imagine today. At the same time, our excitement comes from the same source – the models are getting better quickly. In the last two months, we’ve seen releases from OpenAI and Anthropic that radically expand the promise of what AI-enabled services can do. The implications of Strawberry and the race to scale the reasoning layer have a huge impact on the value a service like ours can provide to customers. Claude’s latest computer use release also fundamentally changes what we can do for customers without application layer integrations. We can’t wait to see what we’ll be able to do just one year from now!

If you’re building an AI-enabled service in the industrial space, or generally share the same excitement and anxiety, we would love to hear from you.??

Huge thanks to Gaby Lorenzi for the help in writing this up!