Your Next Coworker Is an AI Agent. Here's How to Onboard It Well.

If you watched someone at your company use AI today, there's a good chance it would look a lot like a Google search. They'd type a question, read the answer, maybe copy a paragraph into an email, and move on. That's useful, and I don't want to knock it. But it's a small slice of what the technology can now do, and the gap between that everyday habit and what the most advanced teams are doing has become enormous.

OpenAI published the largest study to date of how people use ChatGPT, based on a sample of more than a million conversations. It found that about 49 percent of messages were people asking for information or advice, and three kinds of conversation (practical guidance, seeking information, and writing) made up nearly 80 percent of all use. Only 27 percent of messages were related to work at all. In other words, for most of the roughly 700 million people using it every week at the time, AI was a very good answer machine.

Meanwhile, a handful of companies have quietly moved on to something different. For them, the chat window has become the place where work actually happens, and the other software they pay for has turned into a set of tools the AI reaches into on their behalf. That shift has a name, agentic AI, and I think it's the most important change in how businesses will operate over the next few years. So I want to explain what it is in plain terms, what happened this year to push it forward, and what it takes to bring it into a company the right way.

What makes an agent different from a chatbot

A chatbot answers. An agent acts. That's the simplest way I know to put it.

When you ask a chatbot for help writing a follow-up email, it writes a draft and hands it back to you. You still open your inbox, find the thread, paste it in, attach the file, and hit send. An agent can do those steps itself. It can look up the thread in your email, pull the contract from the shared drive, write the message, attach the document, and ask you whether it's ready to go out. It works across several tools in a row to finish a task, and it only comes back to you when it needs a decision.

The second difference is memory. A traditional chatbot forgets you the moment the conversation ends. The newer agents keep track of who you are, what your company does, how you like things formatted, and what you were working on last week. Over time they behave less like a piece of software and more like a colleague who has been around long enough to know how things work.

The moment agents went mainstream

For a while, agents were mostly a talking point at tech conferences. That changed quickly at the start of this year.

In November 2025, an Austrian developer named Peter Steinberger released a free, open-source project that, after a couple of name changes, became known as OpenClaw. The idea was simple and a little startling. Instead of opening an app to use AI, you'd run an agent that lives inside the messaging tools you already use, like Telegram, Signal, or Discord. It runs around the clock, remembers what you've told it, and can take action on your computer and accounts. People gave their agents names, added them to group chats, and started assigning them work the way they'd assign it to a new hire.

It spread incredibly fast. By early March, the project had more than 247,000 stars on GitHub, which is roughly the software world's equivalent of a viral video. In February, Steinberger announced he was joining OpenAI. And almost overnight, every major AI company started racing to offer its own version of an always-on AI coworker.

OpenClaw also taught a lesson that I think business owners should take seriously. The same power that made it exciting made it risky. In February, security researchers at Koi Security audited 2,857 add-on "skills" in the project's public marketplace and found 341 that were malicious, many of them built to steal passwords and cryptocurrency wallet keys. Laurie Voss, a respected engineer and former chief technology officer at npm, called it "a security dumpster fire." The takeaway I'd draw is a practical one. An agent that can do anything you can do on your computer needs someone deciding, carefully, what it's actually allowed to do.

The "plus one" model

So what does working with an agent look like inside a normal company? The picture I keep coming back to is a plus one. Every person on the team gets an AI agent that works alongside them, knows their responsibilities, and takes on the parts of the job that don't need their judgment.

Here's an example. Say part of your role is giving the CEO a weekly report on the company's financial picture. Today that might eat most of a day. With a plus one, the work gets split. You ask your agent to research what happened in the markets that week and draft the context section. You pull the actual numbers yourself from your accounting system, because that's sensitive information and there's no reason your agent needs direct access to it. Then you hand those numbers to the agent and ask it to build the presentation. A task that used to take a full day can come together in about half an hour, and you've kept control of the part that matters most.

The part people tend to miss is what happens with everyone else in the company. If your coworkers know your agent handles the weekly report, they can send their pieces straight to it. The agent organizes what comes in, prepares it, and briefs you when you have time to look. You skip the back-and-forth over formatting and file versions, and your conversations with colleagues can focus on the substance.

We use this at Avolis ourselves. Our team has an agent that handles sales research. When anyone meets a business we might work with, they ask the agent in our team chat to look into the company. It comes back with a thoughtful profile covering what the company does, its size, and where our work could help. Then our sales team can role-play with it to practice handling the questions and objections that company is likely to raise. It's one of the simplest things we've built, and it gets used every day.

Where adoption actually stands

It's easy to read the headlines and assume everyone has already made this jump. The data says otherwise.

McKinsey's latest State of AI survey, published in August, found that nearly 90 percent of organizations now use AI regularly in at least one part of the business. But only about 20 percent report scaling AI agents across the organization. The split by company size is the number I'd pay attention to. Among large enterprises with more than a billion dollars in revenue, 40 percent say they're scaling agents, up from 27 percent the year before. Among smaller organizations, the figure sat flat at 22 percent. The biggest companies are pulling ahead, and the small and mid-sized businesses we work with are the ones at risk of being left behind.

The results gap is even starker. Only 6 percent of companies in McKinsey's survey qualify as "high performers," meaning AI has made a significant, measurable difference to their bottom line. Using AI and getting value from it are turning out to be two very different things.

None of this means the technology is stalling. Gartner predicts that 40 percent of enterprise software applications will include task-specific AI agents by the end of this year, up from less than 5 percent in 2025. Consumer-friendly agent products are growing fast too. Manus, one of the best-known general-purpose agents, reached an annual revenue run rate of about $125 million by the end of 2025, less than a year after launch. The agents are arriving whether a company is ready for them or not. What varies is how prepared the people are to work with them.

Why this is a people project

Every agent deployment we design starts with a set of questions that have nothing to do with which model to use. They're about people and responsibilities.

What does the agent do, and what do you do? The first step is breaking each role into its actual tasks and deciding which ones an agent can own, which ones it can help with, and which ones stay entirely human. That map is different for a dispatcher, an estimator, and a bookkeeper, and it's worth getting right.

What can it see? An agent is only as safe as its access. Some systems, like payroll, banking, and anything with client financials, usually shouldn't be connected at all. Others can be connected with read-only permissions. Deciding this up front prevents most of the problems people worry about.

Where does a person sign off? An agent should never be the one making important decisions on your behalf. Sending money, committing to a price, or replying to an upset customer are all moments where a person reviews the work before it goes out. Good agent design builds those checkpoints in from the start.

Who else can work with it? The biggest productivity gains come when coworkers can hand work to each other's agents. That requires shared rules about what's appropriate to send, how requests get routed, and how everyone stays informed.

Who helps people get comfortable? Most people's first experience with an agent is a little unnerving. It's doing things on their behalf, and they're not sure yet whether to trust it. Training built around each person's real work, plus someone they can go to with questions in the first few weeks, is what turns a pilot into a habit.

These questions are where most of the real work of AI adoption lives, and they're why I'm skeptical of any pitch that says you can install an agent and walk away.

A simple way to try it yourself

If you haven't worked with an agent yet, the best way to understand this is to watch one work. Pick a general-purpose agent tool and connect it to a couple of low-risk accounts, like your calendar and a test email inbox. Then give it one instruction with several steps: gather some information, turn it into a short PDF report, and draft an email to a colleague with the report attached.

Watch what happens. It'll figure out which tools it needs, pull the information, analyze it, build the document, write the email, attach the file, and then stop and ask whether you want to send it. Seeing that play out from one request does more to explain agentic AI than anything I could write here. And notice that last step, where it waits for your approval. That's the part worth protecting as you scale this up.

Why we named the company Avolis

People ask about our name, and the answer explains a lot about how we work. Avolis comes from two words: Avalon, the mythical paradise, and polis, the Greek word for city.

I think there are two broad directions AI can take us. One looks like a better version of work, with more balance, a bigger economy, and people spending their time on the parts of their jobs that need a human. The other is the version from science fiction, where the machines run things and people are along for the ride. Real life will probably land somewhere less extreme than either picture, but I believe the better outcome is very achievable. It depends on how companies adopt this technology right now, and whether they keep people in the driver's seat and invest in their skills as the tools get more capable.

That's the outcome we're building toward. Every agent we help design is meant to give someone on your team back hours of their week while keeping them in charge of the work that carries their name.

If you'd like to understand where your company stands today and what a plus one could look like for your team, a good place to start is our AI Readiness Evaluation.

Sources

• Aaron Chatterji et al., "How People Use ChatGPT," NBER Working Paper 34255 (September 2025): https://www.nber.org/papers/w34255

• OpenClaw, Wikipedia: https://en.wikipedia.org/wiki/OpenClaw

• The Hacker News, "Researchers Find 341 Malicious ClawHub Skills Stealing Data from OpenClaw Users" (February 2, 2026): https://thehackernews.com/2026/02/researchers-find-341-malicious-clawhub.html

• The Register, "DIY AI bot farm OpenClaw is a security 'dumpster fire'" (February 3, 2026): https://www.theregister.com/2026/02/03/openclaw_security_problems/

• McKinsey & Company, "The State of AI" (August 2026): https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

• Gartner, "Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, Up from Less Than 5% in 2025" (August 26, 2025): https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025

• Manus (AI agent), Wikipedia: https://en.wikipedia.org/wiki/Manus_(AI_agent)