Look at the AI products that broke out in the last two years and you'll notice something. Almost all of them are co-pilots. They sit alongside a human, suggest text, draft code, summarize meetings, recommend the next sentence. They make the human faster, but the human is still doing the work. The AI is the assistant. The human is the agent.
This is about to flip. The next wave of breakout AI products will be the inverse: the AI is the agent, the human is the supervisor. The work happens autonomously. Humans only step in for exceptions, approvals, and edge cases. The economic implications of this flip are larger than anyone is currently pricing in.
The economic difference
Consider what the co-pilot wave actually did. It made knowledge workers maybe 20–30% faster at their existing tasks. That's meaningful — comparable to the introduction of spreadsheets or email — but it doesn't fundamentally change the business model of knowledge work. You still need the same number of humans, doing the same workflows, in the same companies.
Autonomous agents do change the business model. When one agent can run an end-to-end workflow that previously required ten humans, the per-unit cost of that workflow drops by 90% or more. That's not a productivity gain. That's a market reset. Entire categories of business — legal services, accounting, customer support, recruiting, basic medical triage — get rebuilt around the new cost structure.
The three thresholds for autonomy
The technical bar for autonomy is high but achievable. We look for agents that have crossed three thresholds before backing the company.
Reliability
The agent has to be reliable enough that a human doesn't need to verify each step. If a supervisor has to read every email the agent drafts, the human is still the bottleneck and the cost structure barely moves. We want to see agents that operate at 95%+ task completion rates without supervision in production.
Observability
When something goes wrong — and it will — a supervisor has to be able to figure out what happened in minutes, not days. The best agentic teams treat observability as a first-class product surface, not a debugging convenience. Customers will not hand over real workflows to a black box.
Accountability
Businesses are willing to hand over workflows when the vendor is on the hook for the outcome. That means SLAs, money-back guarantees on missed bookings, or outright outcome pricing. Agentic founders who hide behind "we're just providing a tool" disclaimers will lose to founders who guarantee the result.
The order of operations
What's interesting is the order in which industries are crossing the autonomy threshold. It's not the most prestigious knowledge work going first — law firms and investment banks are nowhere close. It's the workflows that are high-volume, repeatable, and expensive: appointment scheduling, lead qualification, invoice processing, basic compliance review, claims intake, dispatch coordination. These are unglamorous, but they're enormous markets, and they're where the first wave of autonomous-agent businesses will be built.
Workflows we believe cross the threshold first
- •Inbound voice answering for service businesses (HVAC, dental, med-spa).
- •AR/AP automation and collections for SMBs.
- •Lead qualification and routing for high-volume sales orgs.
- •Insurance claims intake and triage.
- •Recruiting outbound, screening, and scheduling.
- •Restaurant ordering, reservations, and supplier reordering.
We invest at exactly this layer — the agent that owns one specific workflow end-to-end for one specific industry. That's where the durable value is. Co-pilots are a feature. Autonomous agents are a business.
