Agents are practical now

For the past year, “AI agent” has meant a glossy demo. A bot books a flight, files an expense report, and somehow never gets stuck. The real change in 2026 is less cinematic and more useful. Companies are not betting on fully autonomous software employees. They are building narrow systems that can search a knowledge base, draft a reply, classify a ticket, enrich a lead, or move a task to the next queue, then hand the result to a person or a rule-based approval step. That is not a retreat from the agent idea. It is the first version that fits inside an actual business process (openai.com, developers.openai.com). The tooling has shifted to match that reality. OpenAI’s Responses API and Agents SDK are built around tool use, traces, handoffs, and full records of what happened during a workflow, which is exactly what a company needs when an agent is touching customer data or triggering an action in another system (openai.com, developers.openai.com). In March, OpenAI went further and described the missing layer in blunt terms: production agents need somewhere to put files, a way to run commands, controlled network access, and retries that do not collapse the whole job when one step fails (openai.com). That is a very different picture from the old prompt-and-pray era, and it leads straight to the next constraint. The constraint is that models still need supervision. OpenAI’s own description of tool use is careful on this point: the model proposes an action, but it does not execute anything by itself; the platform runs the call and feeds back the result (openai.com). Microsoft’s Copilot Studio has moved in the same direction. Its new agent-flow approvals combine AI review with manual approval stages so routine cases can move quickly while harder cases still land with a human who can say yes or no (learn.microsoft.com). The industry’s practical insight is simple. The useful unit of automation is not “replace the worker.” It is “complete the next safe step.” That is why the first business wins are so narrow. In customer support, the work starts with triage. A system identifies intent, pulls policy, drafts a response, and hands off if confidence is low or the case is sensitive. Zendesk’s 2025 CX report described the direction clearly: companies are adopting AI copilots first, not because autonomy is fashionable, but because agents want help on repetitive work and companies see better retention and acquisition when AI is used in a more human-centered way (zendesk.com). Even Zendesk’s own framing of “autonomous service” is really a story about staged escalation, privacy, and reliability rather than a machine running wild (zendesk.com). Sales shows the same pattern. The best use case is not a magical closer. It is relentless follow-up on leads that humans never had time to touch. Salesforce says it began using Agentforce internally in July 2025 to automate outreach, qualification, follow-ups, and meeting booking for neglected leads. By October 1, it said the system had re-engaged 68,000 leads, sent 156,000 emails, booked 800 meetings, and helped close dozens of deals (salesforce.com). Those numbers matter less as proof of universal success than as proof of where the value lives. It lives in the backlog. It lives in the queue that nobody owns because everyone is busy. That is also why the agent boom suddenly feels real. Google Cloud’s 2026 agent trends report argues that the business shift is happening now, not at the distant horizon of AGI, and describes agentic AI as systems that pursue goals with “extensive human guidance and oversight” across office workflows (services.google.com). The surprising part is how modest that sentence is. After two years of giant claims, the breakthrough idea is ordinary workflow design: connect the model to the right tools, keep it inside a narrow lane, log every step, and put a human at the choke points. The future-looking demo has turned into something much less glamorous, like a lead that gets an email in under a minute or a support ticket that arrives pre-triaged with the right article already attached (salesforce.com, learn.microsoft.com).