Everything is an AI agent now. The chatbot on your dentist’s booking page is an agent. The Zapier zap that drops a form entry into a spreadsheet got promoted to agent sometime in 2025 without anyone telling it.
In the first week of November 2026, Perplexity launched always-on Teammates. The same week, CrowdStrike put out a kill switch for AI agents, so we’ve officially reached the stage of the market where the off button is a product.
Gartner said that of the thousands of vendors selling agentic AI in 2025, about 130 were the real thing. Gartner gave the rest a name I’ve been using ever since, agent washing.
And I like them. My business runs on them. Fifteen people on my team work inside Claude all day, my laptop time sits at about 3 hours a week, down from 90, and my clone has published 750+ videos I never filmed.
That setup took me 15 months to build. A good 14 of them went on working out which of my agents did a job and which were a chat window with a name badge. This is the post I wish I’d had in month one, and it starts with what the word even means.
What is an AI agent, in plain English
An AI agent is software that takes a goal, works out the steps itself, uses tools inside your systems to carry them out, and keeps going until the job is done or a rule tells it to stop. A chatbot waits for your next message. An agent gets on with the job.
The cleanest line I’ve found comes from Anthropic’s December 2024 engineering post on building agents. A workflow is a model and some tools wired along a path you decided in advance. An agent is a system where the model gets to “dynamically direct their own processes and tool usage,” which is a long way of saying it picks its own next step.
Anthropic’s advice in the same post was to find the simplest thing that does the job and add the complicated parts only when you’ve earned them, advice I’d frame on the wall of every founder I meet.
Most of what gets sold to you as an agent is a workflow. Honestly, that’s fine.
A workflow that sends your invoice reminders on day 15 is doing more for you than an agent with opinions about your invoices. I covered that whole layer in my AI business automation guide. The trouble starts when you pay agent prices for workflow behavior.
So here are the 3 parts a real one has.
1. The model
This is the thinking part. In my business it’s Claude, loaded with 100,000+ words of my own writing so it decides things the way I would. The model on its own is a very well-read intern with no hands. I ran my pick between the two big ones through 30 real marketing tasks in my Claude vs ChatGPT comparison if you want the long version.
2. The tools
Tools are the hands. Access to your inbox, your CRM, your calendar, your publishing queue, your bank feed if you’re brave. Without tools the model can describe what it would do and nothing more. With tools it can send the email, move the deal and book the slot. This is also the part that should make you slow down, because a model with your bank login is not a demo anymore.
3. The loop and the stop rule
The loop is what makes it an agent. Do a step, check the result, decide the next step, go again. My chief of staff agent reads my inbox, turns it into 3 action items and hands them over, 6 times a day, on a loop that runs every 2 hours.
The stop rule is the piece most builds skip. Mine stop and ask me whenever money, a client or a publish button is involved. Everything else, it handles.
Agent washing, and the 4 questions I ask before I buy one
You’ve seen the demo. A founder shares a screen, types “plan my week” and something cheerful happens. Gartner’s June 2025 release described exactly what you’re watching, chatbots, assistants and old RPA bots relabeled as agents “without substantial agentic capabilities,” and Gartner’s count of real vendors came in at about 130 out of thousands.
I mean, you’ve already been through this once with “AI-powered,” haven’t you? So on the sales call, ask these 4. Watch how fast the energy in the room changes.
| Ask this | A real agent | A renamed chatbot |
|---|---|---|
| What can it do inside my systems without me? | Names the tools it connects to and the actions it takes in them | Talks about “insights” and “recommendations” you then act on yourself |
| Who decides the next step? | The agent, inside limits you set | You do, one prompt at a time |
| When does it stop and ask me? | Has a written rule, usually around money, clients and publishing | Never stops because it never starts anything |
| Can I see what it did yesterday? | Shows you a log of every action | Shows you a chat history |
That fourth question is the one that sorts the room. Gartner’s survey of CIOs from Q2 2026 found 71% expect to deploy agentic AI within the next 12 months, while 21% have any governance for it in place. That gap is the kill switch market. You don’t want to be in it.
AI agents examples from inside my business
Everyone’s list of AI agents examples starts with a customer service bot, so here’s mine, the 5 that run First Movers AI as of November 2026. I wrote the full cost and breakage report in my AI employee breakdown. The short version fits in a table.
| Agent | What it does without me | What I still do |
|---|---|---|
| Writer | Drafts emails, scripts, posts and pages in my voice from a Voice DNA file built on 100,000+ words of my writing, in 10+ formats | Read it and approve it |
| Clone | Fronts my YouTube channel, 750+ videos published to 300,000+ subscribers, none of which I filmed (how I built my clone) | Approve the script before it records |
| Chief of staff | Reads the inbox and turns it into 3 action items, 6 times a day | Act on the 3 items |
| Builder | Builds and fixes the systems in Claude Code when something breaks | Say what broke |
| Follow-up | Drafts the replies to leads and keeps the conversation warm in my voice | Approve the send, then take the call |
The honest cost sits in the right-hand column. I still spend about 6 hours a day on my phone, 5 days a week, reviewing what Claude produced, steering the YouTube strategy and talking to the team. The laptop only opens for builder work, which is where the 3 hours a week comes from.
Three of those agents have broken on me badly enough to write up. The follow-up one is the breakage worth knowing about before you build your own, because its pre-written replies committed me to calls I hadn’t agreed to. The chief of staff’s summary loop had its own problem around the same time, it flagged everything as urgent. The fix for both was a procedure. Nothing sends without me, the 2-hour loop became the checkpoint where I read the full draft, and the writer, builder and chief of staff were split into separate roles.
After selling my 100-person agency in 2021 I started again with 2 employees on payroll. That rebuild passed $100K a month in month 11.
For your business, the first agent usually lands in one of these 5 jobs. Pick the one with a number already attached to it.
| Job | What the agent needs from you | Where it breaks |
|---|---|---|
| Lead follow-up | Your offers, your prices, 3 of your best past replies | Promising a discount you never agreed to, so give it a stop rule on price |
| Inbox triage | A list of who matters and what counts as urgent | Treating your biggest client like a newsletter, so review week one in full |
| Content drafting | A Voice DNA file, your past posts, your banned words | Sounding like everyone else, which is a context problem, covered in my guide to humanizing AI content |
| Invoice chasing | Your accounting feed and a tone you’d sign your name to | Chasing a client who paid yesterday, so connect the feed before the sends |
| Research and monitoring | The questions you ask every Monday and where the answers live | Confident summaries of things it misread, so keep links to the sources in every report |
What AI agents can do today, according to the people who tested them
Carnegie Mellon built a fake software company and gave its office jobs to AI agents, then published the results as TheAgentCompany in December 2024, updated September 2025. The best agent finished 30% of the tasks on its own.
Would you hire a person who finishes 30% of what you hand them? You would if they cost a few dollars an hour and you knew in advance which 30% it was going to be. That second part is the whole skill.
Then there’s Gartner’s other number from that June 2025 release, the one that made every vendor’s week worse. Over 40% of agentic AI projects will be canceled before the end of 2027, for 3 reasons, escalating costs, unclear business value and inadequate risk controls. All 3 are decisions a person skipped before the build started.
Klarna’s 700 agents, and the part the case studies leave out
In February 2024 Klarna put out a press release that every AI agents post on the internet quoted for a year, including the first version of this one.
Its assistant handled 2.3 million conversations in its first month, two-thirds of all customer service chats, doing the work of 700 full-time agents. Resolution time fell from 11 minutes to under 2, repeat inquiries dropped 25%, and the company projected a $40 million profit improvement for the year.
Fifteen months later, in May 2025, Klarna’s CEO told Bloomberg the cost-cutting push had gone too far and announced a recruitment drive so that customers could always reach a real person. Gartner saw it coming across the whole industry. Its June 10, 2025 prediction was that half of the organizations planning big customer service cuts would abandon those plans before 2027.
My read on it, for what it’s worth, is that Klarna’s agent did its job. The 2.3 million conversations were real. The piece that got cut was the stop rule, the moment where the agent hands a frustrated human to another human. Remove that from a customer service agent and what you’ve built is a very fast wall. Keep the agent, keep the handoff. Those were never the same decision.
Is agentic AI the same as AI agents?
Agentic AI and AI agents are near enough that vendors use the two interchangeably, which should tell you something. Agentic AI is the adjective. It describes any AI system built to act toward a goal on its own, which is a property. An AI agent is the thing itself, one piece of software doing one job in your business.
When someone sells you “agentic” anything, your next sentence is “show me the agent.” What does it connect to, what does it decide, when does it stop. Same 4 questions, different brochure.
How to put your first AI agent to work in 30 days
This is the 4-week sequence my team runs with clients. My own first run took 15 months, mostly spent learning this list the hard way. If you’ve already got some automation running, my guide to AI workflow automation covers the plumbing side and this covers the agent on top of it.
| Week | What happens | What you measure |
|---|---|---|
| Week 1 | Pick one boring job that already has a number attached, such as hours spent on follow-up each week. Write its job description the way you’d brief a new hire, including what good looks like and what it must never do | The baseline number, written down before anything is built |
| Week 2 | Load the context. Your voice file, your offers, your past examples, your banned words. Connect one tool, usually the inbox or the CRM. Set the stop rule in writing | How many times the agent stops and asks. Too few is scarier than too many |
| Week 3 | Run it beside the person who does the job today. Sit with them and one real task from their own week, and stay there until what comes out is good enough to send | Edits per output, counted, not felt |
| Week 4 | Let it run the job with daily review. Widen the limits only where the log earned it | The baseline number again, and the hours the person got back |
If you’d like my team to run those 4 weeks with you, and build the agents in the process, the AI consulting build starts with a free call where we map your tools and pick that first job together. Either way, the sequence is the same. Job first, context second, stop rule third, then the agent. The 3 reasons on Gartner’s cancellation list all live in the first 3 steps.
What AI agents cost to run
The subscription side is small. My own AI staff runs on a Claude Max plan at $100 a month on monthly billing, plus clone tools at $50 to $150 a month at list price and a software line of about $150 to $250 a month in total, notetaker included. The build is the real line item, since the agents themselves run on flat subscriptions.
A done-with-you build from my team is $15,000 and a done-for-you build is $40,000, each one 2 months long with a running cost of about $100 a month once it’s handed over.
I’ve broken down every pricing model on the market in how much AI consulting costs. What you own at the end of a build is covered in my done-for-you AI guide. Read the ownership part twice. An agent you rent is a subscription with extra steps.
The best AI agent in your business will have the most boring job
The agent people show off at dinner is the one that planned their trip to Lisbon. The chief of staff, the one that gave me my weeks back, reads my inbox and writes 3 lines, 6 times a day. I have never once been asked to demo it.
That’s the shift I’d love you to make before you spend a dollar. Go hunting for the job that bores you most. The impressive agent can wait.
Boring jobs have clear inputs, clear outputs and a number you already track, which is exactly the shape an agent completes at a rate a lot higher than 30%. My clone is the one with 300,000+ subscribers. The chief of staff, the one that reads my inbox, is the one I’d rebuild first if I lost everything tomorrow.
FAQ on AI agents
What is an AI agent in simple terms?
An AI agent is software that takes a goal, decides its own steps, uses tools inside your systems to carry them out and keeps going until the job is finished or a rule tells it to stop. A chatbot answers one question at a time and waits for the next one. An agent reads your inbox, drafts the replies, handles the routine ones inside its limits and asks you about the rest.
What are some AI agents examples for a small business?
AI agents examples for a small business include a lead follow-up agent that replies to inquiries in your voice, an inbox triage agent that turns the day’s mail into a short action list, a content agent that drafts posts from a voice file, an invoice-chasing agent connected to your accounting feed, and a research agent that answers the same questions every Monday. First Movers AI runs 5, a writer, a clone, a chief of staff, a builder and a follow-up agent.
Is agentic AI the same as an AI agent?
Almost. Agentic AI is the broader term for any AI system built to act toward a goal with some independence. An AI agent is one specific piece of software doing one job that way. Vendors use the two words interchangeably in marketing, so the useful question is what the agent itself connects to, what it decides on its own and when it stops to ask a person.
How much does an AI agent cost?
An AI agent costs a model subscription plus whatever you spend to build it. Claude Max is $100 a month on monthly billing. First Movers AI’s full set of agents runs on about $150 to $250 a month in software. A professional build costs more up front, $15,000 for the done-with-you tier and $40,000 for done-for-you at First Movers AI, with running costs of about $100 a month after handover.
Will AI agents replace my team?
Rarely in full. The companies that tried it loudly have walked some of it back. Klarna announced in February 2024 that its AI assistant was doing the work of 700 agents, then in May 2025 its CEO said the cuts had gone too far and began hiring people again so customers could always reach a human. Agents take over the repeatable parts of jobs. The judgment, the relationships and the handoffs stay with people.
Do I need a developer to build an AI agent?
No. A first AI agent for one job, such as inbox triage or lead follow-up, can be built inside Claude or ChatGPT with a written job description, a context file and one connected tool. A developer or a build team becomes worth paying for when the agent needs to touch several systems at once, when the stop rules get complicated, or when you want it handed over with your team trained to fix it.
Where to start if you’re new to this
Say it’s a Tuesday evening in a few weeks. You open AI Labs and pick the course on building your first Claude project. Before the hour is up you’ve got a context file with your voice in it and one job description written out. Wednesday, you ask the question that’s bugging you in the live session, or you ask the Ask Julia tutor at midnight because it doesn’t sleep either.
That’s what AI Labs is for. Sixty-five courses, 2 live sessions a week on Tuesdays and Wednesdays, the Ask Julia tutor trained on my frameworks, all for $197 a month at the introductory rate. It’s where I’d send you before you book a call with anyone, me included.
The market now sells agents with one hand and the off switch with the other. Build yours around a boring job and a written stop rule, so the switch stays in the box.
End of content.
Final word count: 3,156
