The AI Diary 2025: The Year We Learned What AI Agents Really Are

Nov 12, 2025

2025 was supposed to be “the year of AI agents.” Instead, it became the year we learned what AI agents actually are — and aren’t.

The Paradox: Spectacular Failures Meet Remarkable Success

The numbers tell a contradictory story. On one hand, 40% of agentic AI projects will be cancelled by 2027, and a staggering 90% of agent deployments fail within 30 days. On the other, AI agent companies are experiencing 400% year-over-year growth, and specific implementations are delivering genuine business value.​

McDonald’s cut onboarding time by 65% without hiring a single new trainer using a voice-activated AI training simulator that guides employees through tasks in real-time. Walmart deployed a “self-healing inventory system” that detects demand surges, adjusts replenishment schedules, and reroutes products between distribution centers—reducing food spoilage by hundreds of thousands of dollars. Mercedes-Benz integrated the MBUX virtual assistant into select vehicles, allowing drivers to request highly contextual recommendations through natural conversation.​

These success stories are real. But they’re also misleading if you don’t understand the economics underneath.

The Two Business Models Emerging

Agent-as-a-Service (AaaS): A product-led model where the agent itself is the product — designed for specific tasks and sold as a software solution.

Agent Marketplaces: Distribution platforms similar to Uber, Fiverr, or Airbnb that follow traditional marketplace economics, connecting agent capabilities with business needs.​

The distinction matters because each model faces different economic pressures and viability challenges.

Why Most Enterprise Deployments Fail

B2B adoption of AI agents remains disappointingly slow, and the reason is straightforward: security teams are blocking deployments. According to AI researcher Andrej Karpathy, agents produce brittle and unpredictable results, lack basic reliability, don’t possess true reasoning capabilities, and don’t learn unless manually retrained.​

Current agents excel at basic repetitive tasks — formatting documents, creating to-do lists, assigning tasks, and eliminating administrative work. But they’re nowhere close to autonomous operations. Everything they do requires validation, checking, and refinement.​

The pattern in failed deployments is revealing: the 90% that failed were attempting to save money, while the 10% that succeeded weren’t trying to save anything at all.

The Harvey AI Exception: Expensive Agents That Work

Harvey AI stands as the most striking example of successful agentic implementation — precisely because it defies conventional software economics. The legal AI platform hit $100 million in annual recurring revenue in August 2025, representing 400% year-over-year growth.​

Their pricing model? $1,200 per attorney per month — ten times more expensive than traditional legal software. They serve 500 enterprise customers with 12-month contracts requiring a 20-seat minimum, and those customers are doubling seat count within 12 months.​

Harvey’s team includes 10% ex-lawyers dedicated to ensuring law firms hit usage thresholds for renewal. They position their product around preconfigured agentic workflows with clear expectations about what the agent will automate.​

How can the most expensive agents be the most successful? The answer lies in a fundamental shift in how AI agents compete in the market.

The Labor Budget vs. IT Budget Revolution

Traditional software — Salesforce, Slack, Zoom, Microsoft Suite — competes for the IT budget, which typically represents around 2% of a company’s total spending. AI agents represent the first technology in history that competes for the labor budget, which accounts for 60-70% of total company spending.​

Consider a typical law firm’s budget breakdown: out of every $100 in revenue, $45-50 goes to labor costs (salaries, benefits), while only $2 goes to technology. Harvey isn’t competing for that $2 technology allocation — it’s competing against the $45-50 labor budget.​

When a law firm evaluates Harvey’s $1,200 monthly price tag, they’re not comparing it to other software costs. They’re comparing it to a first-year associate earning $13,000 per month (approximately $150,000 salary plus benefits). When competing for labor budgets instead of software budgets, expensive becomes cheap.

The Hidden Iceberg of Agent Costs

The economics of AI agents diverge fundamentally from traditional SaaS models. In SaaS, once infrastructure is deployed, adding an extra user costs near zero because software can be replicated infinitely at minimal cost. Shopify, for example, acquires new customers at almost zero marginal cost.​

AI agents operate differently: marginal cost is very far from zero. Every action burns GPU compute and energy. Costs do not trend toward zero even at scale.​

Agentic systems can consume five to 20 times more tokens than simple AI chains because they involve loops, retries, and multi-step planning. Every routing decision, tool selection, and context generation can trigger multiple large language model calls.​

McDonald’s publicized their 65% reduction in onboarding time but not the deployment cost: $12 million across 200 locations — that’s $60,000 per location, more than a new hire earns in two years. The agent didn’t replace labor cost; it frontloaded it.​

Beyond compute costs lies an iceberg of invisible expenses:​

• Pre-deployment data work including curation and preparation of training data
• RAG knowledge-based construction requiring embedding generation, chunking strategy design, and semantic indexing
• Context optimization to eliminate data duplication (storage can increase by 25% without proper deduplication)​
• Cloud infrastructure costs ranging from $2,000 to $10,000​
• Integration costs, technical debt, and security measures

What Agents Are Actually Good At

The sobering truth: agents make sense when task predictability exceeds 90%, decision logic is simple, and zero errors are required. They excel at basic, repetitive, monotonous tasks but remain far from autonomous operations.​

Current agents don’t possess cognitive abilities. They don’t learn from past experiences. They handle structured, predictable workflows but struggle with messy, unpredictable business operations.​

For problems where each instance is unique, requiring reasoning across unstructured data with natural language interaction and continuous learning, agents may not make sense — or need to be deployed with highly variable cost structures.

The Emerging Opportunity: AgentOps

Despite the hype-reality gap, genuine opportunities are emerging. AgentOps — a new business function focused on orchestrating and managing multi-agent workflows — is slowly forming.​

Analysis of 3,000+ AI job postings reveals emerging requirements around agent orchestration and multi-agent workflows. At least 17 AgentOps tools already exist in the market, and major platforms are embedding AgentOps functionalities into their core offerings.​

Unlike DevOps, which tends to be centralized, AgentOps will likely be more distributed across organizations. Fixing, debugging, and rerouting agents will be more accessible to non-technical teams, with agents deployed across customer support, project management, finance, and accounting.​

The infrastructure layer for agent management represents a particularly promising opportunity for startups. Currently, no clear market leader exists — only fragmented tools that don’t match the quality available for DevOps.

What 2025 Actually Taught Us

2025 wasn’t the year agents replaced people. It was the year we learned what agents actually are: not cheap labor, but a new category of software that competes for labor budgets instead of IT budgets.​

The successful implementations weren’t trying to eliminate human workers — they were augmenting human capabilities in highly specific, predictable workflows. The failures came from treating agents as plug-and-play labor replacements without understanding the economic realities, technical limitations, and hidden costs.​

As we approach 2026, the path forward is clearer: agents that assist, empower, and support — rather than replace — are finding product-market fit. Those marketed as automated labor are struggling against the reality that agent technology isn’t there yet, and may never be designed for full autonomy.​

The real question isn’t whether agents will change everything. It’s whether businesses can identify the specific, high-value use cases where predictable workflows, clear decision logic, and augmented human capabilities create genuine ROI — even at premium prices competing against labor budgets rather than software subscriptions.

We’ve learned agents are expensive, limited, and nowhere near the autonomous intelligence promised. We’ve also learned that within those constraints, properly deployed agents can deliver transformational business value. The companies succeeding in 2025 understood that distinction. The ones failing did not.