Agents at Work: The 2025 State of AI in Production and What Is Actually Shipping

The AI conversation has a hype problem. On one side, accelerationists who believe we are six months from AGI and everything will be different by next year. On the other side, skeptics who have been calling it a bubble since 2022 and point to every model failure as validation.

Both miss what is actually happening: a quiet, unglamorous, enormously consequential deployment of AI into production systems across every sector — shipping real value, measurable in dollars, right now.

This is a dispatch from inside that deployment. Not the demos. The production systems.

What Is Actually in Production

Customer service AI is the most widely deployed category. Not chatbots in the old sense — scripted decision trees with fallback to humans. Genuine conversational AI that handles the full range of inbound customer inquiries for major companies in financial services, telecommunications, e-commerce, and SaaS. The deflection rates — the percentage of inquiries resolved without human involvement — are running at sixty to eighty percent in mature deployments. The customer satisfaction scores are, in most cases, comparable to human agents. The cost per interaction is eighty to ninety percent lower.

Code generation and review is the second most widely deployed category and the one with the most measurable productivity impact. GitHub's internal data on Copilot usage shows developers completing tasks fifty-five percent faster. McKinsey research on AI-assisted coding shows similar numbers. The less publicized finding is that the productivity gains compound: developers who use AI assistance consistently are not just faster — they are building more ambitious systems because the cognitive overhead of implementation has decreased.

Document intelligence — extraction, classification, summarization, and analysis of unstructured documents — is running in production at scale in financial services, insurance, legal, and healthcare. A major insurance company processing ten thousand claims per day with AI document extraction reports ninety-two percent accuracy and a seventy percent reduction in processing time. These numbers are not marketing. They are audit-traceable production metrics.

Sales development AI is earlier in its production curve but accelerating rapidly. AI systems handling cold outreach, lead research, personalization at scale, and follow-up sequencing are in production at hundreds of companies. The performance benchmarks at mature implementations consistently show eight to fifteen percent reply rates — three to five times the industry average for human-operated outreach — because the AI has eliminated the generic messaging problem.

What Is Consistently Failing in Production

Long-horizon autonomous task execution without human checkpoints. The demos of AI agents completing complex multi-step tasks end-to-end are impressive. The production reality is that current models hallucinate, lose context, and make consequential errors at a rate that makes fully autonomous long-horizon execution impractical for high-stakes tasks. The solution — human-in-the-loop at defined checkpoints — works but limits the autonomy that makes agents most valuable.

Complex reasoning under genuine uncertainty. AI systems perform remarkably in well-defined problem spaces with adequate training data. They struggle when the problem space is genuinely novel, when the relevant context is not in the training distribution, or when the task requires the kind of judgment that integrates ethical, relational, and contextual factors that are not reducible to pattern matching.

Real-time structured data reasoning. Asking an AI to reason about a live database, a current spreadsheet, or real-time market data without careful architecture produces unreliable results. The models were trained on static snapshots. Without explicit tooling to ground them in live data, they confabulate.

The Twelve-Month Forecast

Voice AI for inbound customer interactions is the category that will see the most dramatic production expansion in the next twelve months. The quality of AI voice has crossed a threshold where most callers cannot reliably distinguish AI from human in the first sixty seconds of interaction. The cost differential — roughly ninety percent lower than human agents — will drive adoption regardless of the philosophical debates about authenticity.

Multimodal AI in professional workflows is the capability that will unlock the next wave of enterprise value. The ability to reason across text, images, documents, video, and audio simultaneously opens use cases in quality control, medical imaging, legal discovery, and architectural design that were not viable with text-only models.

AI-native software development — not AI-assisted development, but development where AI is responsible for the majority of code generation and the human role is primarily architectural and review-based — will become standard practice for new projects within twelve months. The tools are already there. The workflows are being established. The productivity data will pull the rest of the industry.

What This Means for Operators

The production data is a forcing function. The businesses deploying AI in production are accumulating cost advantages, speed advantages, and data advantages over those that are not. These advantages compound. They are not linear.

The practical implication: identify the highest-volume, most-repeatable cognitive tasks in your operation. Those are the immediate deployment targets. Implement with appropriate human oversight. Measure obsessively. Expand what works.

The window between early adopter and competitive necessity is approximately eighteen months for most categories. In some — customer service AI, sales development AI, code assistance — it is already closing.

The demos are behind us.

Production is the present.

The question is where you are in the deployment curve.