We've been asking companies a simple question: how many of your employees have actually deployed an agent to automate part of their day? The answers tell the story.
"You mean chatting with Claude?" "I have no idea what an agent is." "The engineers love agents!"
That gap is the real story of AI adoption right now. McKinsey's November 2025 State of AI puts 23% of organizations scaling agentic AI and another 39% experimenting. Those numbers describe the company, not the people in it. At the employee level, the share who've shipped their own agent is a rounding error.
It's not because the technology isn't ready. The frontier models can already do the work. The reason most employees haven't deployed an agent is that nobody has handed them a repeatable mental model for what "deploying one" even looks like. So most attempts fizzle. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027.
Agents are supposed to be a natural extension of your team, accomplishing tasks that would otherwise require a human. The formula for getting one into production is pretty simple. Five decisions, in order.
1. Pick the trigger
The first decision is the most important and the most underrated. An agent is not a chat window. It's a process that runs when something happens. A new meeting note appears in Granola. A row lands in a spreadsheet. A webhook fires. A cron job kicks off every morning at 8am.
Picking the trigger pins down what the agent is actually for. "Summarize emails" is not a workable spec. "When a new Granola note is created, log the call in HubSpot and draft a follow-up email in Gmail" is one. Vague triggers produce vague agents.
2. Define the goal
Once the trigger is set, write down what the agent should accomplish when it fires. Two or three sentences for the end goal, plus a few rules of thumb for how to get there. "Always reply in the customer's language." "Never quote a price." "If the request mentions a refund, escalate."
Resist the urge to write a 2,000-word prompt. The instructions you'd give a smart new hire on day one are almost always the right level. If you have to explain every edge case in advance, you're building a flowchart, not an agent.
3. Pick the model & loop in a human
Different jobs need different lanes. A summarization agent doesn't need the same reasoning model as one negotiating a refund. A high-volume tagging agent doesn't need the same context window as one reading a 50-page contract. The model layer is changing fast enough that the right choice in March can be the wrong choice in May. Build agents on a platform that lets you swap.
In the same step, decide where a human belongs. The default for any agent touching real systems should be: a person approves before anything irreversible happens. Sending an email. Updating a CRM record. Charging a customer. Deloitte's research shows teams that redesigned workflows around human-AI collaboration saw a 30% productivity lift, compared to just 5% when AI was bolted onto unchanged processes. Human-in-the-loop is not a tax. It's the thing that lets the agent earn trust fast enough to eventually remove the loop.
4. Build the first draft
With the trigger, goal, model, and approval points decided, build the agent. The first version should be the simplest thing that could possibly work. Resist the temptation to wire in five tools, three retrieval systems, and a self-critique loop on day one. You don't yet know what the agent gets wrong. Get something running end-to-end before you optimize anything.
This is also the right time to think about credentials. The agent will be calling external APIs on your behalf. Those keys should be encrypted at rest, injected at runtime, and never visible to the model itself. If your platform's design lets the AI see raw API keys, the blast radius of a single bad agent is your entire stack.
5. Test and iterate until it's reliable
This is where most teams quit. The agent runs on three test cases, looks impressive, and gets called done. Production inputs are messier. The agent encounters a customer writing in two languages mid-thread. It hits a malformed webhook payload. It tries to escalate to a Slack channel that was archived last quarter.
Run the agent on real inputs. Watch it fail. Tighten the goal or the rules of thumb. Add a tool when a missing capability shows up twice. Move a step behind human approval when the agent makes a class of mistake you can't accept. Keep going until the runs are boring. Boring is the signal you're done.
What "deployed" actually means
The first agent doesn't need to be complex. It needs to be useful. An agent worth running in production is one you've watched fail enough times to trust.
That's how agent adoption starts. Not with a company-wide transformation deck, but with one person deploying one useful agent into one real workflow. The full step-by-step walkthrough lives at generalinput.com/guides.