From SaaS to Services as Software

There’s an old joke in venture capital: services businesses are where software companies go to die. Software scales infinitely: write code once, sell it a million times. Services businesses, by contrast, have to hire another human every time they take on another client. With AI, that calculus is breaking down. 

The scalability trap

The knock on services companies was never about quality. They delivered genuinely sophisticated, tailored work. But the problem was always unit economics. Every new engagement required proportionally more headcount, more coordination overhead, more management layers. You could not simply “deploy” another copy of your firm the way a SaaS company could spin up another server.

The conventional answer was productization. Esri is the textbook case: founded in 1969 as a land-use consulting firm, it eventually recognized that the spatial analysis it was performing for clients followed repeatable patterns enough to be encoded into software. The result was ArcGIS (then called ARC/INFO), a platform that let other organizations do their own GIS work, at scale, without hiring Esri’s consultants for every project. Esri escaped the services trap by turning its methodology into a product. The trade-off was that customers had to adapt their workflows to fit what the software could do. 

“The old model was ‘software as a service.’ What’s emerging now is something closer to ‘service as software’ — expertise that scales like code.”

Agents change the unit economics

AI agents are doing something strange to both sides of this equation simultaneously. On the services side, they function as infinitely deployable specialists that let services firms scale output without scaling (human) headcount linearly. An agent handling risk assessments or monitoring change over time can be cloned to handle ten times the volume tomorrow morning. Labor as the fundamental constraint of the services model starts to bend.

At the same time, agents are quietly undermining the value proposition of a lot of traditional software. If a workflow is simple enough to have been packaged into a SaaS product, it’s simple enough to describe in a system prompt. Outbound sales is the obvious example. The instructions are not complicated: research the prospect, write a relevant opening line, reference a pain point, ask for a meeting. Any reasonably capable language model can execute that loop today, and companies are discovering they don’t need to pay for a dedicated platform to do it - they can build it themselves in an afternoon.

Where agents should be deployed externally

Not every AI use case is a good candidate for building in-house, and GIS is a clear example of why. A large enterprise commissioning spatial analysis — whether for infrastructure planning or risk assessments — faces a dual expertise problem. Deploying agents effectively in this domain requires deep GIS knowledge: understanding of data models, projection systems, analysis methods, and how to interpret outputs in context. It also requires serious AI implementation skill: knowing how to design reliable agent workflows, where to insert human review, and how to avoid the confident-but-wrong failure modes that make unsupervised automation dangerous in high-stakes decisions. Few enterprises have both, and building them in parallel is a distraction from the decisions the spatial analysis is meant to support.

This is where specialized agents become genuinely compelling — and where the trajectory points to something more radical. Rather than enterprises trying to build this capability in-house, or commissioning consultancies whose expertise is locked inside individual practitioners, the architecture exists today for agents to own the entire pipeline end to end: from interpreting a business question, to sourcing and processing the right datasets, to running the analysis and delivering a defensible output — with no human in the loop unless the decision warrants it. The domain expertise that once made GIS engagements so expensive and slow to scale is no longer locked inside a person. It is encoded in the agent.

The new shape of competition

What emerges from this is a genuinely interesting inversion. Services companies now have a credible path to venture-scale economics and scale. And software companies are discovering that “scalable workflow” and “defensible moat” are not the same thing.

The firms that win the next decade will likely look more like hybrid organizations: deep human expertise in a domain, amplified by agents that let a team of twenty deliver what previously required two hundred. Services, finally, at software margins.

Johanna von der Leyen is CEO and co-founder of PangeAI, a geospatial AI company that enables energy, insurance, and natural capital organizations to query the physical world in natural language. PangeAI emerged from stealth in December 2025.

Want to learn more? Reach out: hello@pangeai.com