AI-Native Services
How replacing labor, not augmenting it, creates the next $10T category of software-like businesses
The Thesis
For decades, software made humans more productive. AI replaces large part of the work humans do altogether.
In Systems of Action I wrote:
“Systems of Action execute entire workflows with humans stepping in only to approve exceptions. By replacing human labor, services businesses—historically seen as non-scalable—begin to exhibit product-like margins. Law, accounting, and property management firms may soon become scalable and investable businesses.”
Previous technology waves expanded software TAM to roughly $1 trillion by digitizing workflows and automating coordination. AI goes further.
By making human judgment a digital and scalable good, it expands the market from software spend to labor spend.
We are moving from roughly $300 billion in SaaS to the trillions spent on human labor.
As Jensen Huang put it at Sequoia’s AI Ascent: “This is the first time in technology we’re not creating a tool. We’re displacing labor budget.”
Lightspeed captures the same idea concretely: Salesforce helps manage sales workflows, but an AI sales agent can do the selling. Genesys runs contact centers, but an AI agent can resolve queries directly. Excel helps analysts build models, but an AI analyst can do the analysis.
The pattern holds wherever work is structured, repeatable, document-driven, and rule-based.
This shift is giving rise to AI-native services: companies that present as service providers but operate with software economics, selling outcomes rather than seats or billable hours.
Two paths lead here:
AI-native services built from scratch with humans in the loop
SaaS companies adding agents and forward-deployed engineers
This essay provides a framework for understanding this inflection point.
I. Why Now: The Enabling Technologies
AI-native services are enabled not by a single breakthrough, but by the convergence of capabilities that allow software to perform real work reliably.
First, general-purpose models have crossed a threshold in handling service workflows. Large language models now excel at unstructured document processing, data reconciliation, semantic search, multi-step reasoning, and tool use. This matters because most services work is interpretive rather than creative: reading messy inputs, matching them to rules and precedents, resolving inconsistencies, and producing a decision or action.
Second, voice AI has quietly become good enough to unlock entire categories of service work. Advances in speech synthesis and understanding mean agents can handle phone calls without breaking flow or sounding robotic. Tasks like collections, intake, scheduling, and customer support move from “human-only” to automatable.
Third, agents are no longer confined to single applications. Emerging browser and desktop automation allows AI systems to operate across heterogeneous software environments much like a human worker would. Agents can navigate legacy systems, fill forms, reconcile data across tools, and complete workflows end to end without bespoke integrations.
Together, these capabilities remove the historical bottlenecks that kept services labor-bound. Software can now read what humans read, speak where humans speak, and operate where humans operate. The remaining challenge is no longer model capability, but workflow redesign, which is why AI-native services are now viable rather than theoretical.
In practice, this approach works best in workflows that are document-heavy, rules-driven, and repeatable, with humans stepping in primarily for judgment and edge cases.
II. The Low-Hanging Fruit: BPO
The BPO market, projected to reach roughly $525 billion by 2030, illustrates the opportunity. As a16z notes in Unbundling the BPO, incumbents like Cognizant, Infosys, and Wipro each generate $10–20 billion handling work enterprises prefer not to do internally: customer support, claims processing, reconciliation, and IT operations.
The experience is often broken. Turnaround times are long, errors accumulate, and staff lack the context or authority to resolve edge cases. Enterprises tolerate this because building equivalent internal capacity is worse.
BPO is the low-hanging fruit for AI-native services for three reasons. First, the budget already exists, creating a PMF advantage and reducing adoption friction. Second, much BPO work is repetitive and rules-based, making it especially vulnerable to automation that improves accuracy and speed. Third, outcome-based pricing aligns incentives far better than billable hours or fixed retainers.
BPO incumbents are structurally unprepared for this shift. AI-native operations compress margins, cannibalize labor revenue, and require product and ML capabilities that outsourcing firms largely lack today.
III. The North Star: Revenue per Employee
If AI-native services truly behave more like product businesses, that should show up in the numbers. The clearest early signal is revenue per employee.
This metric reveals whether automation is compounding or whether the business is still scaling like a traditional service firm. In practice, it should be evaluated against service benchmarks in the same vertical, not against SaaS peers. Past waves of “tech-enabled services” often promised operating leverage without delivering it; applying an exacting standard here is essential to avoid mistaking hype for progress.
IV. Getting Automation Right
Full automation is not required at the outset. In practice, it is often wiser to “trade margin for moat” by automating where leverage is highest while keeping humans in the loop elsewhere.
This is the central operating challenge for AI-native services. More than model quality or interface design, how, when and where automation is applied, determines whether a company compounds into a defensible business or stalls as a labor-heavy service.
Done well, this approach allows companies to understand true complexity, refine workflows with real data, build proprietary training sets from human-handled cases, and identify which steps can be automated versus augmented at each stage of maturity.
V. Beyond BPO: Property Management as a Case Study
Consider AI-native property management. The workflow spans tenant communication, maintenance coordination, lease administration, rent collection, and owner reporting. Some components are ready for automation immediately: tenant inquiries, maintenance triage, rent reminders, and reporting.
Others benefit from humans initially. Lease negotiations, complex maintenance decisions, disputes, and vendor management require judgment. Keeping humans involved while structuring every interaction generates training data for future automation.
Over time, this compounds. Eighteen months of human-led lease negotiations can yield tens of thousands of tagged transcripts, forming the basis for a negotiation agent competitors cannot easily replicate.
The margin trajectory reflects this sequencing:
Year 1: ~30% gross margin
Year 2: ~45% as core workflows automate
Year 3: 60%+ as agents handle most cases
The winner is not the fastest automator, but the company that sequences automation to build defensible data assets.
VI. Will SaaS Companies Become Service Providers? The Rise of the Forward Deployed Engineer
AI also pulls SaaS companies toward services.
As agents move from assisting users to performing work, success depends on how well systems reflect real operational workflows. This is where forward-deployed engineers become central.
Forward-deployed engineers embed with customers to understand how work is done on the ground, then redesign workflows around AI capabilities. They map business logic, define where automation is safe, integrate fragmented systems, and turn edge cases into structured rules the product can learn from. In doing so, they translate operational reality into executable systems.
This model shares core characteristics with services businesses: deep operational involvement, humans in the loop, accountability for outcomes, and early economics shaped by human effort. Forward-deployed engineering becomes a bridge between SaaS and services.
Decagon illustrates this pattern. Its agent product managers work closely with customers to deploy AI support agents in production, integrating systems, adapting workflows, and handling early failures so agents can resolve requests reliably. An older precedent is Palantir, whose forward-deployed teams have long embedded inside customer organizations, co-running mission-critical workflows and redesigning processes as reality changes.
What we once called “implementation” has become an ongoing operational function. As software moves from enabling work to doing the work, that function becomes load-bearing.
VII. A Blurry Line
As AI capabilities advance, the line between AI-native services and SaaS with agents will blur.
The distinction will hinge on the answer to one question: who owns delivery?
AI-native services sell completed work and bear responsibility for execution, quality, and failure.
SaaS with agentic capabilities leaves accountability with the customer and prices access rather than outcomes.
As AI systems improve, companies will move along this spectrum.
Some SaaS companies will be pulled towards services and operational responsibility as customers demand outcomes rather than tools.
Similarly, some AI-native services will push towards software models as automation absorbs more of their workflows.
Conclusion
By automating jobs rather than tasks, AI-native services capture labor budgets and achieve ACVs SaaS alone cannot.
Over time, the market will sort companies not by how they describe themselves, but by who performs the work, how incentives are aligned, and how value accumulates as it scales.
Attribution
This essay synthesizes ideas from several original contributions on AI and services, including:
Lightspeed, AI Will Eat Services
Andreessen Horowitz, Unbundling the BPO; Trading Margin for Moat
Point Nine Capital, AI-First Service Businesses
Bessemer Venture Partners, AI Systems of Action
Sequoia Capital, AI Ascent 2025
Emergence Capital, forward-deployed engineers
Any interpretations or errors are my own.
If you look at this from where value migrates — not where technology goes — the pattern is clear:
Whatever AI can fully replace will not be the right place to build lasting value.
Enduring value lives where responsibility, judgment, trust, and unique insight cannot be automated — only amplified.