AI agents for business: what an agent is and how to build one
Everyone talks about agents. Few can say what one is. Our definition is short: an AI agent is a job that runs on its own. A trigger starts it, skills do the tasks in order, connectors give it reach into your systems, and a person approves before anything leaves the house.
That makes an agent the top of a ladder, not a separate IT project. You build it from skills you already trust, for a job you have already run by hand. This guide shows how, in Claude, now with Cowork built in, and in ChatGPT Work.
- 01
Skill
One task, described in detail. The how.
- 02
Connector
Access to the system the task needs. The where.
- 03
Job
Skills in order, run by a person. The proof it works.
- 04
Agent
The job on a trigger, with a checkpoint. The time you get back.
What an AI agent is
A chat assistant answers when you ask. An agent does a job when something happens: the first working day of the month, a new ticket, a signed contract. It works through several steps, uses your systems along the way and hands over a finished result.
Anthropic separates workflows, where the steps are set in advance, from agents, where the model decides the next step itself. In a business, most of the value sits in the first kind: jobs with clear steps that someone does every week. Start there. Every agent we recommend has five parts.
- A trigger
- A time or an event that starts it. Without one it is still a chat.
- Skills
- One per task, written and tested before the agent exists. See our guide to Claude skills.
- Connectors
- Access to the CRM, inbox, accounting or project tool the job needs. See all MCP connectors.
- A checkpoint
- The step where a person approves, before anything is sent, booked or published.
- An owner
- One named person who answers for the result and fixes the skills when the process changes.
Agent, automation or chat?
Three things that often get the same name. They suit different work.
| What it does | Good for | |
|---|---|---|
| Chat assistant | Answers when someone asks | One-off questions, drafts, thinking out loud |
| Rule-based automation | Moves data when a rule fires, with no judgement | Copying a form into the CRM, sending a receipt |
| AI agent | Runs a job with judgement: reads, writes, summarises, decides the next step | Reports, triage, reviews, follow-ups |
How to build an AI agent, step by step
The order matters more than the tool. Skip a step and the agent is the first thing to break.
- 1. Pick one job
- Weekly or monthly, clear steps, a clear result, and someone who is tired of doing it. The monthly report, not "marketing".
- 2. Write the skills
- One skill per task in the job, each with steps, rules and an example.
- 3. Run it by hand five times
- A person runs the skills in order and fixes what goes wrong. This is where the missing rules show up.
- 4. Connect the systems
- Give the AI access to the data the job needs, on the access of the person it runs for.
- 5. Put the checkpoint in
- Decide who approves what. Nothing leaves the house before that person has said yes.
- 6. Add the trigger
- Now, and only now, let it start on its own: a schedule or an event.
- 7. Measure it
- Time saved per run and how often the approver has to fix something. If fixes do not go down, go back to step 2.
Claude Cowork: the agent is already in your chat
Cowork is the part of Claude that does a whole job instead of answering a question: it gathers the inputs from your files, connected tools and the web, works through the steps and comes back with a finished result. Since 16 September 2026 it is no longer a separate mode. Chat and Cowork are one Claude, and Claude decides itself whether to answer or to do the job, in the same conversation. It is rolling out to Pro and Max first, on web, desktop and mobile. By default Claude asks before it takes an action; you can let it run and check in only when needed.
Two things turn that into an agent. Skills and plugins make Claude do the job your way. Scheduled tasks let you describe a job once and have it run every day, week or month on its own, with your connectors and plugins, even when your computer is asleep. Scheduled tasks are on every paid Claude plan: Pro, Max, Team and Enterprise.
ChatGPT agent is now ChatGPT Work
Look for ChatGPT agent today and you will not find it. OpenAI retired agent mode and replaced it in July 2026 with ChatGPT Work: long tasks across your apps and files, with documents, spreadsheets and slides as output and scheduled tasks built in. It is OpenAI’s answer to Claude Cowork.
For teams there are workspace agents, shared agents a whole team uses, on Business, Enterprise and Edu plans. OpenAI calls them an evolution of GPTs, and custom GPTs themselves retire on 11 December 2026: migrated to a plugin, a GPT’s instructions become a skill. Agent Builder, OpenAI’s visual tool for developers, shuts down on 30 November 2026. If you built something there, plan the move now.
The direction is the same at both companies: skills for how the work is done, connectors (OpenAI calls them apps) for reach, and the AI doing the job in the same place you chat.
Claude Cowork vs ChatGPT Work
The two have grown close. Both do whole jobs, both run on a schedule, both read skills in the same open format.
| Claude (with Cowork) | ChatGPT Work | |
|---|---|---|
| Where the work happens | In any Claude conversation, on Anthropic’s servers | In ChatGPT, on web, mobile and desktop |
| What it works with | Your files, connectors, plugins and the web | Your files, apps (OpenAI’s word for connectors), plugins and the web |
| Your way of working | Claude skills and plugins | ChatGPT skills, in the same open format, and plugins |
| On a schedule | Scheduled tasks on every paid plan | Scheduled tasks built in |
| For a whole team | Skills and plugins shared by admins on Team and Enterprise | Workspace agents on Business, Enterprise and Edu |
Why AI agents fail
When agents fail, the model is rarely the problem. It is nearly always one of four things, and every one of them is fixed lower down the ladder. Gartner expects over 40 % of agentic AI projects to be cancelled by the end of 2027. We wrote about why in why AI agents fail in production.
- Nobody owns the result
- An agent without an owner degrades quietly until someone notices a wrong report.
- The process was never clear
- If two people do the job differently, the agent does a third version. Write the skills first.
- No checkpoint
- An agent that sends or books without approval will one day send the wrong thing.
- No access to the data
- Without connectors the agent fills the gaps itself. That is where invented numbers come from.
Agents for a whole company
One person’s scheduled task in Claude is a good start. Twenty agents across a company raise three questions IT will ask on day one: which agents run, on whose access, and what did they do?
Deployed OS answers them. Company skills and connectors live in one place and reach Claude, ChatGPT and Copilot. Every person connects with their own login, so an agent never sees more than the person it runs for. Admins decide which groups may let the AI write, and every call is logged.
AI agents by role
Six roles, each walked up the ladder: the skills to write first, the job they make, and the agent that runs it.
- MarketingAI agents for marketing6 skills · The monthly marketing report · The report agent
- SalesAI agents for sales6 skills · The meeting follow-up · The follow-up agent
- FinanceAI agents for finance6 skills · Preparing the month-end close · The close prep agent
- HRAI agents for HR6 skills · Onboarding a new starter · The onboarding agent
- Customer serviceAI agents for customer service6 skills · The morning queue · The morning queue agent
- OperationsAI agents for operations6 skills · The weekly project status · The status agent
AI agents: common questions
A job that runs on its own. A trigger starts it, skills do the tasks in order, connectors give it access to your systems, and a person approves the result before anything leaves the house.
A chatbot answers when you ask. An agent starts by itself when something happens, works through several steps in your systems and hands over a finished result.
Pick one recurring job, write a skill for each task, run the job by hand a few times, connect the systems it needs, add an approval step, and only then put it on a schedule in Claude or ChatGPT.
The one your team already works in. They have grown very similar: both do whole jobs, run on a schedule, use skills in the same open format and connect to your tools. With Deployed OS the same skills and connectors work in both.
OpenAI retired ChatGPT agent, also called agent mode, and replaced it with ChatGPT Work in July 2026. Agent Builder shuts down on 30 November 2026, and custom GPTs retire on 11 December 2026.
It is when the agent runs on the access of the person it works for, writes only where you allow it, stops for approval before anything is sent, and every call is logged. Start read-only and add writing step by step.
AI that carries out multi-step tasks instead of only answering questions. In practice: the agents described on this page.
It depends on the size of your team. Get in touch and we will walk you through it.
Start with one job.
Tell us which task eats your team’s week. We’ll show you the skills that do it, and the agent that runs it.
