Agentic AI dominated tech press for two years. Every vendor pitched agents. Every conference had agent panels. Every analyst forecast agent adoption growth in the high triple digits.

We’re now far enough into 2026 to assess what’s real and what was vapor. Here’s the deployment picture from the operations side.

What’s actually shipping

A few categories of agentic AI are now in genuine production use:

Customer service triage agents that classify tickets, gather initial information from users, and either resolve simple issues or route to humans with full context. These work well because the task is bounded and there’s plenty of training data from existing ticket histories.

Code generation and refactoring agents running inside developer environments. The agentic version (where the AI navigates a codebase, makes coordinated changes across files, and runs tests) has become reliably useful for specific task types — adding tests, refactoring single concerns, generating boilerplate.

Internal knowledge agents that search across an organization’s documents, summarise findings, and produce drafts. These work because the failure mode (a slightly wrong summary that gets corrected by the human reading it) is acceptable for the task.

Research and procurement assistants that pull together data across multiple sources for human decision-makers. Same dynamic — the human is still the decision-maker, the AI is doing legwork that scales.

What hasn’t shipped at scale

The vision of fully autonomous agents that handle complex multi-step business processes end-to-end is not where production is today. The systems that try this are failing at one of three points:

Long-horizon planning. Agents lose track of what they were doing across 10-20 step processes. Even with planning frameworks, accumulated errors compound over long sequences. The result is agents that handle the first few steps competently then drift.

Tool integration reliability. Agents that need to interact with several external systems (CRM, ticketing, email, finance) hit edge cases constantly. The integration plumbing wasn’t built for AI traffic patterns. When something fails, the agent often doesn’t know how to recover.

Trust and verification. For high-stakes processes (financial transactions, customer-facing communications, regulated decisions), the verification cost of letting an agent run autonomously is similar to the cost of having a human do the work in the first place. The economics don’t favor agents until accuracy and explainability improve significantly.

The architectures that work

The agent deployments that work share a structural pattern: bounded autonomy with human checkpoints. The agent handles the parts that are clearly delegated, presents results at decision points, and a human decides whether to continue. This is much closer to “AI-augmented workflow” than to “autonomous agent.”

This is unglamorous compared to the AGI-adjacent vision, but it’s the architecture that actually delivers operational value. Organisations adopting agents successfully have largely accepted this framing.

For organisations evaluating where to start, working with business AI solutions on a bounded use case typically delivers more value than chasing a moonshot autonomous workflow project.

The vendors winning vs losing

Vendors that built thin wrappers around foundation models with the word “agent” in the marketing have mostly underperformed. The category was crowded, the differentiation was minimal, and customers learned quickly that the demo and the deployment were different things.

Vendors winning are the ones investing in the unsexy parts: integration libraries, observability, error handling, audit trails, multi-tenant security. These don’t generate keynote excitement but they determine whether agents deploy successfully.

What’s coming next

A few trends are visible in late 2026:

• Better composition of small specialised agents rather than monolithic generalist agents
• Standards emerging around agent-to-agent communication protocols (still messy)
• Tighter integration between agent frameworks and identity/access management infrastructure
• More work on agent evaluation methodology that goes beyond benchmark scores

The early hype phase is over. The real engineering phase is well underway. The interesting question is which categories see meaningful productivity gains in the next 18 months and which remain “promising in pilots, hard at scale.”

My current bet is that customer service, software development, and internal knowledge work will see meaningful gains. Marketing automation, complex sales workflows, and most enterprise process automation will continue to underdeliver against initial expectations. Time will tell.