How 80,000 companies build with AI: Products as organisms and the death of org charts | Asha Sharma

Asha Sharma, who oversees Microsoft’s AI platform, infrastructure, and foundation models, shares a perspective informed by observing thousands of companies navigating the AI landscape. The conversation reveals several fundamental shifts in how products are built, how companies are organized, and where the next wave of innovation lies.

From Product as Artifact to Product as Organism

A core thesis is the evolution of products from static “artifacts” to dynamic “organisms.” Previously, product development was a linear process: have an insight, solve a problem, ship a product, and maybe iterate. In the AI era, this model is insufficient.

“All of a sudden these are these living organisms that just get better with the more interactions that happen. And in many ways, I think this is the new IP of every single company… products that think and live and learn.”

The key metric is no longer just shipment, but the “metabolism of a product team”—its ability to ingest data, digest it through a rewards model, and tune the product toward specific outcomes like price, performance, or quality. This continuous feedback loop is where value and defensibility are now created.

The Rise of Post-Training and the New AI Moat

This shift to living products is powered by the increasing importance of post-training over pre-training. While building a large foundation model requires immense capital expenditure (capex), research suggests that once a model surpasses 30 billion parameters, the economic and performance benefits of fine-tuning and reinforcement learning (RL) on the “loop” often outweigh the cost of further pre-training.

This is where companies build their competitive moat. By using proprietary data—or synthetically generated data—to fine-tune models for specific use cases, they create a virtuous cycle. For example, Microsoft’s Dragon product for physicians saw its character acceptance rate jump from 30-60% to over 83% after being fine-tuned on 600,000 annotated patient-physician interactions.

Asha predicts that investment in post-training will soon eclipse pre-training, creating a new ecosystem of infrastructure and platforms dedicated to this part of the stack.

The New Organizing Principle: The Loop, Not the Lane

To support these “product organisms,” the structure of teams must also evolve. The traditional model of siloed functions—product, engineering, design, research—creates too many touchpoints and slows down the feedback loop. In a world with hundreds of new models and technologies emerging weekly, this is untenable.

The new principle is “the loop, not the lane.” The focus is on the entire data and feedback loop, blurring the lines between functions. This has led to the renaissance of the “full-stack builder” or “polymath”—individuals who can operate across disciplines to increase velocity. The entire team becomes obsessed with the efficiency of the product, the rewards system design, and the user experience, dissolving traditional boundaries to metabolize learnings and iterate faster.

The Agentic Society: From Org Chart to Work Chart

Looking forward, Sharma envisions an “agentic society” where the marginal cost of a good output approaches zero, leading to an exponential demand for productivity. This demand will be met by agents—both embedded as tools and embodied as autonomous entities.

This will fundamentally reshape organizations:

• The Org Chart Becomes the Work Chart: Hierarchical structures designed for upward communication will give way to dynamic, task-based networks. When a new task arises, it can be automatically routed to the best-suited agent or human, with progress observed and fine-tuned in real-time.
• Expanded Skill Sets: Employees will have their own “agent stack,” analogous to a “bring your own device” policy. This will give them access to skills they never had before, exponentially increasing their capabilities and overall economic productivity.

The Future of Interfaces: Code-Native over GUI

Just as databases evolved from desktop apps to SQL and cloud management moved from consoles to Terraform, AI is driving a shift from graphical user interfaces (GUIs) to code-native interfaces.

LLMs interact more naturally with streams of text. This means the future of product development will focus less on the “canvas” and more on “composability.” Product makers need to rewire their thinking away from the visual UI and toward how an agent can read and interact with their product, enabling infinite scale and new forms of collaboration.

Planning and Strategy in the AI Era

Planning in such a volatile environment is a challenge. Microsoft has adopted a “seasons” framework to navigate it:

1. Define the Season: Ground the organization in the current secular trend (e.g., the “advent of agents”). This can last three, six, or twelve months.
2. Set a North Star: Establish a clear vision of what winning looks like and the key metric for that season.
3. Use Loose Quarterly OKRs: Set flexible objectives for the next quarter that align with the season’s theme.
4. Leave Slack in the System: Intentionally reserve capacity not just for unplanned work, but for investing in the “slope”—disrupting your own platform to stay ahead of the curve.

• Tags:
• Future of Work
• Product Management
• Large Language Models
• AI Strategy
• Organizational Design
• Agents
• Microsoft