The AI Agent Economy — Who's Building, Who's Buying, and What's Actually Working
By April 2026, "AI agents" had become the most overloaded term in tech. Every SaaS company with a chatbot called it an agent. Every startup with a for-loop around an LLM claimed to be an agentic platform. Investors allocated $12B to agent-related startups in Q1 2026 alone.
So it's worth asking bluntly: what's actually working?
Verqo's database tracks 101 tools in the AI Agents category — the largest single category in the directory. This article maps the real landscape: who's building agents, who's buying them, and which categories have crossed from hype into genuine traction.
What's actually working in AI agents:
- Coding agents (Cursor, Devin, GitHub Copilot Workspace) have the highest demonstrated ROI
- Customer service agents are winning in high-volume, low-complexity scenarios
- Research agents are reliable for structured data gathering, unreliable for judgment calls
- Personal assistant agents have traction with power users but high churn at the consumer tier
- The "general agent" category is still mostly hype — the real wins are vertical specialists
How We Define an Agent (It Matters)
Before surveying the landscape, the definition matters. The industry uses "agent" to mean at least three distinct things:
- Orchestrators: LLMs that call tools in sequence to complete multi-step tasks (most "agents" today)
- Persistent agents: LLMs that maintain memory, run continuously, and act on triggers without user prompts
- Autonomous workers: LLMs that are given a goal and a budget, and figure out the path themselves
Most products on the market are orchestrators. True persistent agents and autonomous workers are rarer, harder to build, and where the real value will ultimately concentrate.
This distinction explains why agent demos are spectacular and real-world deployments are more modest: demos show orchestrators completing a scripted task. Production agents need reliability across thousands of different edge cases they were never explicitly tested on.
The Four Categories With Traction
1. Coding Agents — The Category That Actually Shipped
Coding agents are the clear winner. This isn't close.
Why did coding agents win first? Three reasons:
The feedback loop is tight. When an agent writes broken code, you know immediately — it won't compile, tests fail, the linter screams. That tight feedback loop enables fast iteration on agent quality. Coding is one of the few domains where "did the agent do the right thing?" is measurable in under a minute.
The economic case is obvious. A developer costs $150-250K/year. An agent that saves 2 hours/day at $20/month pays back 100x. No CFO requires a committee meeting to approve that.
The context is bounded. A coding agent knows: the codebase, the tests, the linter rules, the PR conventions. The context needed to do the job is captured in the repository. Other agent categories (research, customer service) deal with fuzzier context, harder verification, and messier failure modes.
The coding agent ROI is compounding: Teams using coding agents aren't just writing code faster — they're handling larger surface area with the same headcount. A 5-person eng team with coding agents can maintain what previously required 8 people. That's not 40% more productive developers; it's 40% fewer developer headcount in the next hiring plan.
2. Customer Service Agents — High-Volume, Narrow-Domain Wins
Customer service AI agents have genuine production deployments at scale. Intercom, Zendesk, and Salesforce all have AI-first agent layers. Standalone platforms like intercom AI, freshdesk, and newer entrants are running millions of conversations monthly.
The wins are concentrated in a specific profile:
- High-volume queries with standard answers: "Where's my order?", "How do I reset my password?", "What's your return policy?"
- Narrow domain: Support for a single product, not a general assistant
- Clear escalation paths: Agent handles tier-1, humans take everything else
The failures are equally concentrated:
- Complex or emotionally charged situations: Agents struggle with angry customers, nuanced edge cases, or anything requiring genuine empathy
- Novel situations: First-contact for a bug the support team hasn't seen before
- Cross-system actions: Agents that need to update 3 different legacy systems with inconsistent APIs
Customer service agents that stay in their lane — high-volume, low-complexity, clear escalation — have demonstrably positive ROI. Those that try to cover the full spectrum fail in predictable ways and damage brand trust when they get it wrong in front of a customer.
3. Research Agents — Reliable for Facts, Unreliable for Judgment
Research agents (Perplexity, perplexity, browser-use-based agents, deep research tools) have found real use cases. The pattern: they're excellent at gathering, weak at synthesizing.
A research agent can:
- Pull all mentions of a company in news, SEC filings, and press releases in 5 minutes
- Aggregate competitor pricing across 15 websites
- Summarize recent technical papers on a topic
A research agent fails at:
- Deciding which of those signals actually matters
- Catching when a source is wrong
- Making the call that a number "doesn't feel right"
The enterprises getting value from research agents treat them like very fast junior analysts: give them structured tasks with clear deliverables, then apply judgment on the output. The companies that fail deploy them as autonomous decision-makers and discover the errors only after acting on them.
4. Personal Assistant Agents — Power User Traction, Consumer Churn
Personal assistant agents (
Power users (typically knowledge workers, entrepreneurs, developers) use these tools for 1-4 hours per day. They've integrated them into workflows, built prompting habits, and understand the limitations. Retention for this cohort is exceptional — over 80% monthly retention in most platforms.
Casual users churn brutally. Sign up to try "AI assistant," generate a few images, ask a few questions, never build it into a real workflow, and churn within 45 days. Consumer AI assistant churn is the dirty secret of the category — top-line MAU numbers look great; paid retention doesn't.
The products winning at PA: those focused on one valuable workflow rather than being everything. Notion AI (writing), Otter.ai (meetings), Superhuman (email) — vertical focus with AI embedded, not AI that's trying to replace everything.
The Categories That Are Still Mostly Hype
General Autonomous Agents
Every few months a demo goes viral: an agent that browses the web, books a flight, fills out a form, writes an email, and reports back. The demo is real. The reliability at scale is not.
70% success sounds reasonable until you realize it means 30% of tasks fail silently or produce wrong outputs. For a task that takes 30 minutes manually, a 30% failure rate is worse than just doing it manually.
General autonomous agents will be transformative. They are not ready for unsupervised business-critical workflows today.
AI SDR / Outbound Sales Agents
The pitch is compelling: AI agents that prospect, personalize emails, follow up, and book meetings autonomously. Reality is messier.
The best AI SDR tools work as drafting assistants — humans review and send. The fully autonomous versions are drowning in spam filters, generating personalization that's technically accurate but tonally wrong, and creating compliance risk in regulated industries.
Net result: AI-assisted sales development has clear ROI. AI-autonomous sales development is still a liability.
What the Landscape Tells Us
Looking across 101 tools in the category, the pattern is consistent:
| Category | Traction | Key Requirement | |----------|----------|-----------------| | Coding agents | ✅ Proven | Tight feedback loop, bounded context | | Customer service | ✅ Proven (narrow) | High volume, standard queries, escalation paths | | Research agents | ⚠️ Mixed | Structured tasks only, human review required | | Personal assistants | ⚠️ Mixed | Power users retain, casual users churn | | General autonomous agents | ❌ Mostly hype | Not production-ready at >70% reliability | | AI SDR / outbound | ❌ Mostly hype | Spam/compliance risk exceeds benefit |
The common thread in what's working: agents with a narrow domain, measurable output, and clear failure handling. The common thread in what's failing: agents solving for generality before solving for reliability.
The next 12 months won't produce "general agents that work." They'll produce 20 more narrow vertical agents that work extremely well. The general agent moment is probably 2028, when model reliability clears 95%+ on complex multi-step tasks.
Who's Buying
The enterprise buying pattern has emerged clearly:
- High-growth startups buy coding agents first, then productivity agents. ROI is immediate and measurable.
- Mid-market companies buy customer service agents. Cost reduction case is clear; risk tolerance is moderate.
- Enterprise moves slowly, buys research agents for internal use cases, moves cautiously on anything customer-facing.
- Consumers try personal assistants, retain if they build a habit in the first 2 weeks, churn otherwise.
Notably absent: large-scale enterprise deployment of autonomous agents for business-critical workflows. The CISOs, legal teams, and risk committees haven't signed off on it yet. That's not a product problem — it's a trust-building problem that will resolve over the next 2-3 years as reliability improves.
What This Means for Builders
If you're building in the agent space:
- Pick a vertical. The "AI agent platform" category is getting crowded fast. Coding agents, legal research agents, medical coding agents — specific verticals with specific feedback loops win.
- Solve reliability first. An agent that does one thing at 97% is worth infinitely more than one that does ten things at 72%.
- Build the human handoff. Every production agent deployment needs a graceful failure path. Products that treat failure as a bug (it should always work) lose. Products that treat failure as a product feature (graceful escalation, clear recovery) win.
- The data moat is real. Coding agents trained on millions of code changes know things about "what works" that no new entrant can match. Build your data flywheel early.
The AI agent economy is $40B+ by 2027. The winners will be vertical specialists that built reliability first.