From Systems of Record to Systems of Action

What a16z Big Ideas 2026 Signals and Why the Real Bottleneck Is Still Data

Every December, a16z publishes Big Ideas: a snapshot of where their investors believe technology builders are heading next. It’s less prediction and more pattern recognition, signals emerging across infrastructure, growth, and applied AI.

Big Ideas 2026 (Part 1) lands on a familiar conclusion from multiple angles:

AI progress is no longer limited by models. It’s limited by data quality, data coordination, and execution environments.

For builders and operators, especially those working with real businesses, not demos, this matters far more than the next model release.

Rather than summarize everything, this post focuses only on the ideas that intersect with a single question:

What actually has to be true for AI to work reliably inside real businesses?

1. The Unstructured Data Problem Is the AI Problem

Jennifer Li frames what many teams now experience firsthand:

“The limiting factor for AI companies is now data entropy.”

Enterprises (and SMBs) are drowning in unstructured and semi-structured data:

Emails
PDFs
Spreadsheets
Screenshots
Logs
Chat threads
Videos
Comments trapped inside SaaS tools

Meanwhile, 80% of corporate knowledge lives outside clean tables.

The result?

RAG systems hallucinate
Agents break silently
Humans remain the QA layer
AI pilots stall

The key insight isn’t just “clean the data.” It’s that data must be continuously structured, validated, reconciled, and kept fresh -or AI becomes brittle.

This marks a shift:

From one-time ingestion → ongoing normalization
From storage → truth maintenance
From access → verifiability

AI doesn’t fail because it’s unintelligent. It fails because it’s asked to reason over contradictory, stale, and contextless inputs.

2. Agent-Native Infrastructure Changes What “Work” Looks Like

Several a16z contributors converge on the same reality:

We are moving from human-speed systems to agent-speed systems.

Malika Aubakirova describes it bluntly:

A single agent goal may trigger thousands of parallel actions: queries, API calls, workflows -at machine speed.

Legacy systems weren’t built for this.

Rate limits trip
Concurrency collapses
Databases interpret agents as attacks
Control planes fail under recursive load

This is not just an infrastructure problem. It’s a coordination problem.

AI-native systems must manage:

State
Permissions
Context
Policy
Sequencing
Recovery

In other words, execution becomes the product.

The winning platforms won’t just “host agents.” They will orchestrate work. Reliably, repeatedly, and safely.

3. Systems of Record Are Losing Their Strategic Power

Sarah Wang articulates one of the most consequential shifts in enterprise software:

“The system of record is becoming a commodity persistence layer.”

Why?

Because AI collapses the distance between:

Intent → Execution
Insight → Action
Data → Outcome

When agents can:

Read across systems
Write back into workflows
Coordinate multi-step processes
Anticipate next actions

…the interface layer becomes the control surface.

Systems of record don’t disappear, but they lose primacy. They become inputs, not destinations.

Strategic value migrates to:

Who controls context
Who governs execution
Who measures outcomes

This reframes the entire stack:

Databases persist
AI executes
Humans supervise

4. Vertical AI Becomes Multiplayer (and That’s the Moat)

Alex Immerman’s “multiplayer” framing is subtle and important.

So far, vertical AI has progressed through:

Retrieval (find the data)
Reasoning (interpret the data)

Next comes: 3. Coordination across parties

Real work is not single-player. It involves:

Buyers and sellers
Managers and teams
Vendors and clients
Internal and external stakeholders

Multiplayer AI requires:

Permissioned context
Shared state
Role-aware workflows
Auditability
Feedback loops

This is where switching costs finally emerge.

When AI doesn’t just answer questions but coordinates labor, the collaboration layer becomes the moat.

5. We Are Designing for Agents, Not Screens

Stephenie Zhang captures a quiet but profound transition:

“We are no longer designing primarily for human consumption.”

As agents take over:

Retrieval
Interpretation
Summarization
Action

…the optimization target changes.

Success is no longer:

Screen time
Click depth
UI density

It becomes:

Machine legibility
Semantic clarity
Consistent business definitions
Deterministic outcomes

The best systems will be the ones agents can read, trust, and act on, not just the ones humans enjoy using.

6. ROI Replaces Engagement as the North Star

Santiago Rodriguez closes the loop:

Screen time is a terrible proxy for value in an AI world.

The future belongs to:

Outcome-based pricing
Time returned
Errors avoided
Revenue unlocked
Stress reduced

But measuring that requires something foundational:

You must be able to observe, verify, and attribute work done by machines.

No verifiability → no ROI story → no scale.

The Throughline: AI Needs a Business Operating Layer

Across infrastructure, growth, and vertical applications, a single pattern emerges:

AI doesn’t need more intelligence. It needs better grounding in how businesses actually operate.

That grounding requires:

Normalized data across tools
Shared context across workflows
Execution environments that coordinate agents
Feedback loops that verify outcomes
Interfaces designed for both humans and machines

This is not a point solution. It’s not another dashboard. It’s not “just integrations.”

It’s an operating layer that sits above systems of record and below human intent.

What This Means for Builders

If you’re building in 2025–2026, the bar is higher:

If your AI can’t explain why it acted, it won’t be trusted.
If your system can’t reconcile conflicting data, it won’t scale.
If your workflows can’t coordinate across tools and people, they won’t stick.
If you can’t quantify outcomes, you won’t win budgets.

The next generation of platforms won’t just assist work. They will run it -under supervision.

And the companies that get there won’t start with models. They’ll start with how businesses actually work.