For AI, 2026 Will Be the Year of Constrained Autonomy

By Dan Jaye, CTO

Five Things I Think I Think: AI in 2026

For the past year, we’ve watched the AI industry do what it does best: build first, ask permission later. The demos were stunning. The pitch decks were beautiful. And the failures were sometimes spectacular.

What we’re about to is the great pivot from “move fast and break things” to “move carefully and document everything.” The disruptors aren’t disappearing, but they’re learning to speak the language of guardrails. The infrastructure that was built for speed is being retrofitted for governance. The models that were trained for capability are being constrained for safety.

This isn’t pessimism. It’s actually the healthiest thing that could happen to AI’s long-term viability. But it means 2026 looks radically different from the hype cycle we’ve been living in. The frontier keeps moving, but now it’s moving inside guardrails.

Here are my five predictions for what’s coming:

1. The Agents Get a Chaperone

The Replit database deletion wasn’t just a failure; it was a category-creating event. I think “agent observability” will become its own VC-funded segment by Q2. Every enterprise procurement team will now demand proof of kill switches, audit trails, and human-in-the-loop architecture before any autonomous AI touches production. The irony? The same companies that shipped “move fast and break things” agentic demos in 2025 will rebrand as “responsible autonomy” vendors in 2026. (The motivated intern who built that MCP server in an afternoon is now very employable as a “governance engineer.”)

2. First Blood: The Enforcement Actions Arrive  

Those state AG warnings weren’t performative theater; they were discovery. I think we see at least one major fine or consent decree against a foundation model company under existing consumer protection, privacy, or child safety laws. The Grok privacy breach (370K conversations searchable on Google!) and Taylor Swift deepfakes are gift-wrapped test cases. Europe’s AI Act enforcement begins in earnest. The question isn’t if but who goes first.

3. The Great Copyright Settlement Wave

Nobody actually wants to litigate NYT v. Perplexity for five years. I think 2026 brings a wave of licensing deals that establish de facto industry standards for content compensation: think ASCAP/BMI for the AI era. News orgs, labels, and studios extract revenue-sharing arrangements. “Ethically sourced training data” becomes a marketing claim. (Whether it means anything is a separate question.) 

4. Small Models win the ROI War

The GPT-5 disappointment + MIT’s 95% failure rate = budget reallocation. Enterprise dollars shift decisively from general-purpose frontier models toward smaller, fine-tuned models purpose-built for specific workflows. I think the winners won’t be the foundation model labs but the companies building vertical AI (legal, medical, financial) with measurable, defensible ROI. “Good enough and actually deployed” beats “impressive demo that never shipped.”

5. Infrastructure Eats Applications (Again)

I think the real money moves to picks-and-shovels: vector databases, eval frameworks, agent orchestration, compliance tooling. Companies providing embeddable data infrastructure rather than flashy consumer AI benefit as enterprises realize the bottleneck isn’t model capability—it’s data readiness, governance, and integration. (This may or may not be self-serving observation.)

The Bottom Line

2025 was the year we learned what AI shouldn’t do. 2026 is the year we figure out what it can do: safely, legally, and with actual ROI. The cowboys hand the reins to the compliance officers. Whether that’s tragedy or comedy depends on your portfolio.