Most AI startups are service businesses, pretending to be product companies
Also, why Accenture makes more money from AI than OpenAI
A month ago, NYTimes and FT reported that Accenture and BCG make more money from Generative AI than OpenAI, Anthropic, etc, combined. This baffled many, but it makes perfect sense when you consider this: selling AI automation to enterprise is more about consulting than building software.
In fact, as LLMs get more powerful and AI becomes more democratized, selling AI will become more about selling time, than product. More work will shift to doing consulting and client service, and away from hands on keyboard coding. This will also be a reflection of extreme competition in the AI infra layer, as I wrote about here.
On the flip side, due to reliance on services, AI startups will be harder to scale than traditional software. The barrier of entry will also be lower especially as knowledge workers learn to build their own automations with low-code tools. As a result, the best AI startups will end up niche versions of Accenture or Palantir, rather than OpenAI.
!This is bad for VCs paying software multiples for what's basically agencies!, which can’t scale easily (for reference, Accenture trades at 27x earnings, while Harvey raised at ~60x sales).
So in this post, I'll dive deeper into 1) why services are unavoidable for AI startups targeting enterprise, 2) and why AI startups are hard to scale, let alone become venture-scale.
For our purposes, an “AI startup” is a company that offers streamlined AI agents, workflows, or chatbots, essentially trying to be the next ServiceNow’s and Zapier’s for different verticals. Common feature set includes 1) workflow builders, 2) workflow templates, and 3) a bunch of integrations and connectors to databases, etc.
Why AI Startups Are Stuck Selling Services
So here's a secret that enterprise AI startups hate to admit, especially those working on AI automation - that they are secretly stuck selling AI software consultancy, rather than software with high gross margins. By consultancy, we refer to creating more or less bespoke solutions for customers, without delivering platform revenue.
For early stage startups, doing services without selling anything repeatable is okay, as long as founders can pick up domain expertise. Services are also generally easier to sell initially than products, especially when founders don’t really know what they are building.
But here’s the problem: the consulting component never goes away when it comes to selling AI software, outside of very select categories (like coding agents). That’s because beyond very simple things (think Zapier “zaps”), rarely one workflow template works for another customer out of the box.
It’s almost impossible to build a “one-size-fits-all” type of AI agent / integration / workflow / chatbot, especially beyond anything trivial, which slows down sales and lowers margins. There are different stakeholders, database schemas, policies, and various other constraints that require solution engineering.
Customizations, integration assistance, and hands on keyboard support doesn’t go away. In other words, the “product” can only sold with “forward deployed engineering”.
Every enterprise customer thinks their use case and situation is unique (which has an element of truth). Even within the same industry, customer requirements are meaningfully different. The more ambitious the AI automation is, the harder it is to convince the customer why an out of the box automation can work for them.
Of course, AI startups will get better at solution engineering, and customers will be more willing to fast track POCs. But every engagement will still be meaningful different, which means the startup can only scale revenue with the # of solution engineers or consultants. They will still want to talk to a human.
But this hands on consulting component will never go away, because selling AI automation or productivity is essentially selling business transformation, which means customer success is important. Founders need to be careful about what “automation market” they are entering to avoid becoming a full-blown software consultancy.
Unfortunately, this also means most AI startup founders will find themselves working “in the job”.
The Lock-In Mirage
At this point, some may point out AI workflows and agents - once adopted by the enterprise - are hard to displace. After all, if your finance department outsourced all of its compliance work to AI agents, then your AI agents are hard to fire, right? Meanwhile, your company is charging the customer a fat spread in LLM inference costs versus what you are paying OpenAI (free lunch??)
But here’s the problem: as LLMs and AI agent infra gets better, customers get smarter, and the low-code tools get better, it will become easier for companies of all sizes to build their own automations (which is what they are already doing anyways).
Only time will tell whether AI agents and automation platforms have a strong lock-in effect. Unlike traditional SaaS products where data migration and retraining users create significant switching costs, AI workflows and agents are likely portable.
Moreover, as enterprise customers become more AI-savvy, they're less likely to see value in overpaying for inference on workflows they've already implemented. They might opt to keep the consulting services but ditch the "platform" altogether.
This "lock-in mirage" is a reality check for AI startups and their investors. It underscores the need for these companies to continually innovate and provide value beyond mere workflow implementation. Otherwise, they risk becoming glorified system integrators rather than the next big software unicorns.
Escape Routes: Point Solutions vs. Platforms?
To avoid this trap, I see startups taking two paths:
Developing point solutions that can be sold without heavy consulting component
Building AI agent / workflow builder infra
The problem with point solutions? They're easy pickings for incumbents to copy, releasing your product as an "AI feature." Not to mention, most valuable workflows already have point solutions. It’s already happening.
What about building a platform by accumulating point solutions, à la ServiceNow? I suspect that LLMs won’t lead to multiple ServiceNow-scale companies, but many smaller competitors that operate in niches and operating essentially as consulting companies.
The Bottom Line
Enterprise AI startups face a unique scaling challenge: they're caught between product and service. While they promise software-like margins, the reality of AI implementation demands intensive customization and customer hand-holding. This service-heavy component isn't just a growing pain—it's intrinsic to enterprise AI adoption.