AI agents are becoming power users of enterprise software and deciding which tools to buy next—and Microsoft knows it

AI is transforming how enterprise software gets bought—not by replacing users, but by becoming one. 

The debate around AI and the workplace often centers on labor displacement: Will it replace workers? Where will it fall short? And indeed, some “AI-first” experiments have produced mixed results—Klarna reversed course on customer service automation, while Duolingo faced public backlash for an AI-focused growth strategy. 

These outcomes complicate our understanding of Microsoft’s recent efficiency-driven layoffs. Unlike a premature overcommitment to automation (à la Klarna), Microsoft is restructuring to operate as “customer zero” for its own enterprise AI tools, fundamentally changing how the computing giant writes code, ships products, and supports clients. It’s a strategic shot in the arm—a painful one—that reveals what’s coming next: AI agents built not just to automate outcomes, but to make decisions about the tools, processes, and infrastructure used along the way. 

AI agent as orchestrator

In the past, enterprise software was chosen through a familiar dance: evaluation, demos, stakeholder alignment, and procurement. But today, AI agents are building applications, provisioning infrastructure, and selecting tools—autonomously, and at scale. Ask an agent to spin up a customer feedback portal, and it might choose Next.js for the frontend, Neon for the cloud database, Vercel for hosting, and Clerk for authentication as a service. No human has to Google options, compare vendors, or meet with salespeople. The agent simply acts.

Internal telemetry from Neon shows that AI agents now create databases at 4 times the rate of human developers. And that pattern is extending beyond engineering. Agents will soon assemble sales pipelines, orchestrate onboarding flows, manage IT operations—and, along the way, select the tools that work. 

Microsoft’s sales team re-org further hints at how this procurement will occur in the future. Corporate customers now have a single point of contact at Microsoft, rather than several salespeople for different products. In part, this may be because agentic AI tools will select vendors on their own—and copilots don’t need five sales reps. The agent won’t pause to ask, “Do you have a preferred vendor?” It will reason about the task at hand and continue on its code path, hurtling toward an answer.

Human-in-the-loop AI

This evolution from executor to decision-maker is powered by the human-in-the-loop (HITL) approach to AI model training.

For years, enterprise AI has been limited by expensive labeling processes, fragile automation, and underutilized human expertise, leading to failure in nuanced, high-stakes environments like finance, customer service, and health care.

HITL systems change that by embedding AI directly into the workforce. During real-time work, agents observe GUI-level interactions—clicks, edits, approvals—capturing rich signals from natural behavior. These human corrections serve as high-quality validation points, boosting operational accuracy to ~99% without interrupting the workflow. The result is a continuous learning loop where agents don’t just follow instructions, they learn how the work gets done. This also creates dynamic, living datasets tailored to real business processes within the organization.

This shift offers entirely new market opportunities. 

On the development front, traditional supervised learning models are giving way to embedded learning systems that harvest real-world interaction signals, enabling cheaper, faster, more adaptive AI. This further offers a massive new training set for agentic AI systems without incurring the cost of hiring human knowledge workers to shepherd the AI. With lower development costs, high fidelity, and better dynamism, the next generation of copilots will blend automation with real-time human judgment, dominating verticals like customer service, security, sales, and internal operations. 

Accordingly, these tools will require infrastructure for real-time monitoring, GUI-level interaction capture, dynamic labeling, and automated retraining—creating further platform opportunities.

Microsoft’s sense of urgency

While the internet abounds with zippy coverage of savvy employees “AI hacking” their workflows, the reality is most workers lack that kind of product-development acumen. (And same for their bosses.) Save for a small subset of the business world possessing rare tech fluency, most corporate outfits will see greater value in buying AI tools—those built, customized, and serviced by world-class talent to solve specific workflows.

Microsoft’s sense of urgency comes from its understanding that the question of “build or buy” is changing quickly. This “eureka” moment, technologically speaking, is what’s catalyzing an operator pivot at enterprise AI outfits. HITL represents a move away from read/write data integrations toward a richer, more dynamic GUI-interaction-based intelligence layer—one that mirrors how work actually gets done in the enterprise. 

We’re seeing the beginning of a race toward enterprise AI dominance among the goliaths of the tech world. Signals like OpenAI’s investments into application-layer experiences (shopping agents, its acquisition of agentic developer Windsurf) highlight a clear trend: Mastering human-application-interaction capture is becoming the foundation for scalable agentic automation. As companies like Microsoft, OpenAI, and others absorb critical data environments and restructure themselves to serve as “customer zero,” they’re treating AI as the new chief procurement officer of their own ecosystems. These companies see the value of selling shovels in a gold rush—and know AI is finally sharp enough to start digging.

Tomasz Tunguz is the founder and general manager of Theory Ventures. He served as managing partner at Redpoint Ventures for 14 years.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.

Read more:

• ‘Sovereign AI’ is political branding. The reality is closer to digital colonialism
• Trump has inadvertently shown Europe it needs to build a full-stack AI industry—and avoid a risky reliance
• The AI cost collapse is changing what’s possible—with massive implications for tech startups
• When AI builds AI: The next great inventors might not be human

This story was originally featured on Fortune.com