On July 9, 2024, Sarah Franklin, the CEO of an HR software company called Lattice, posted on LinkedIn that her company had “made history.” In her words, Lattice was taking the idea of an AI employee from concept to reality.
AI agents would get official employee records. They would be onboarded, given goals and performance metrics, and assigned a human manager, with their own box on the org chart next to the people.
Three days later, she gave Fortune a statement saying Lattice would not pursue it. The pushback was about the org chart and the performance review, and it came from HR people and tech executives inside 72 hours.
For the last twelve months, five AI employees and 15 humans have run First Movers together. The business does $250K a month, my laptop time sits at about 3 hours a week, down from 90, and the Claude plan underneath all of it costs $100 a month.
My last company was a content agency that took seven years and 100 people to reach $100K a month. I sold it in 2021, started First Movers, and passed the same milestone in under a year with two employees.
We’ve grown to about 15 since. Every dollar figure in here is either on my consulting page or in a dated report from Gartner, McKinsey, Anthropic, or Carnegie Mellon.
What an AI employee is
An AI employee is a software agent that owns a business function end-to-end. It runs on a large language model such as Claude, it works from your own files and voice, under rules you wrote down, it produces finished work on a schedule, and a named human signs off before anything reaches a customer.
A chatbot answers a question and waits. An automation moves data from one box to another. An AI employee drafts the Tuesday email, checks it against the offer calendar, and puts it in front of you with a note on why it chose that subject line.
Gartner’s June 25, 2025 press release estimated that only about 130 of the thousands of vendors selling “agentic AI” have anything agentic under the hood, and gave the rest a name, agent washing, for chatbots and old automation tools rebranded with the new word. The same release predicted that over 40 percent of agentic AI projects would be canceled before the end of 2027, citing escalating costs, unclear business value, and inadequate risk controls. Escalating costs and inadequate risk controls are management problems, and so is unclear business value when nobody wrote the job description first.
McKinsey’s State of AI survey, published August 25, 2026 from 1,719 respondents in 97 countries, puts the adoption picture in numbers. Forty percent of large organizations, the ones with over $1 billion in revenue, say they’re scaling AI agents, up from 27 percent a year earlier. Smaller organizations stayed flat at 22 percent.
Eighty percent of respondents say AI improved their own productivity, and 37 percent say it moved their company’s earnings, a share that hasn’t budged in a year. That gap, between a person feeling faster and a company making money, is the one an AI employee has to close.
The same idea is sold as AI workers, digital employees, or AI staff, and the three terms mean the same thing in practice.
The AI staff at First Movers, one role at a time
Each of my AI staff has a written job description, the same as the humans. All five sit in one Claude system trained on 100,000+ words of my thinking, wired to my own files, running loops that summarize my email, pre-write my drafts, and hand me action items every two hours. The only thing that changes from role to role is the output it owns and the human who checks it.
| Role | What it owns | What it runs on | Who checks it | What it gave back |
|---|---|---|---|---|
| The writer | Marketing emails, social posts, workshop scripts, offers and funnels, book drafts | Claude with Voice DNA, 100,000+ words of my material, one Project per content type | Me, from my phone, most of the day | A 90-hour laptop week became about 3 hours |
| The clone | The YouTube channel and short-form cuts, filmed once | HeyGen video avatar, ElevenLabs voice, scripts drafted in my Claude system and approved on my phone | Me on every script. Every video says it is the avatar | 750+ videos I never filmed, 300,000+ subscribers |
| The chief of staff | Inbox summaries, pre-written replies, action items every two hours | Claude loops on a timer, connected to my mail and files | Me. Nothing sends without me | The inbox handled in three action items, six times a day |
| The builder | One-time routines, dashboards, cleanup of the systems we already had | Claude Code | Me at the laptop, inside the 3 hours | Builds that used to wait on a developer |
| The follow-up | Research before calls and follow-up after them | Claude routines connected to the booking flow | The team, and me on anything with money attached | Calls arrive briefed and leave with a next step |
All five report to me, and none of them can send, publish, or book anything without a human tap.
What an AI employee costs, in dollars and in hours
My entire Claude system runs on a $100 a month plan. The figure is on my consulting page, it hasn’t changed this year, and the plan is flat-rate.
The clone stack sits on top. HeyGen’s Creator plan is $29 a month, or $24 on annual billing, and Pro starts at $49. ElevenLabs charges $22 a month for Creator, the first tier with professional voice cloning, and $99 for Pro.
Add a wearable notetaker and the software line for one founder’s AI staff lands somewhere between about $150 and $250 a month at list price, depending on the tiers your volume needs.
The setup took 15 months of loading transcripts, rewriting instructions, and finding out which tasks the system could own and which ones still needed me.
Then there’s the review layer. I spend about six hours a day on my phone reviewing what Claude produced, steering the YouTube strategy and talking to the team. That is the actual cost of an AI employee, a human who reads everything before it ships.
Six hours a day, five days a week, that’s about 130 hours a month. The software line tops out at $250. Price those 130 hours at whatever your own hour is worth, and the $250 becomes the rounding error.
McKinsey found that about one in five organizations now say AI operating costs, tokens included, are limiting how much AI they use. Mine have stayed flat for three reasons. One flat plan, a small number of roles, and no agents calling other agents in a loop nobody is watching.
| Line | What it is | What it costs | Where the figure comes from |
|---|---|---|---|
| Claude plan | The brain behind all five roles | $100 a month | First Movers consulting page, updated August 2026 |
| HeyGen | Video avatar for the clone | $29 Creator ($24 annual), Pro from $49, Business $149 plus $20 a seat | heygen.com/pricing, checked September 5, 2026 |
| ElevenLabs | Voice for the clone | $22 Creator, $99 Pro | elevenlabs.io/pricing, checked September 5, 2026 |
| Notetaker | Wearable voice capture wired into Claude | One-time hardware purchase | First Movers consulting page |
| Setup | Voice DNA, routines, loops, the job descriptions | 15 months, spread across running the business | First Movers consulting page |
| Review | A human reading everything before it ships | About 6 hours a day, on my phone | First Movers consulting page |
| Humans | The 15 people on the team | Payroll, and it is the biggest line | First Movers consulting page |
What twelve months of AI staff produced
A man watched one of Kimberly Kubitza’s clone videos, found her number, and called to say every symptom she’d described was his. He started her protocol the following week.
Kim runs Shiloh Wholistic Wellness Center, and she built her clone with us at Clone Mastermind, our two-day build in Scottsdale. She recorded that video once, and it landed her first client.
My own twelve months read like this.
- Video. The clone published 750+ videos I never filmed, and each one says on screen that it is the avatar. The channel passed 300,000+ subscribers, and I’m the first person to put a real AI clone on YouTube and grow an audience around it.
- Revenue. First Movers passed $100K a month in under a year with two employees, and now runs at $250K a month with 15 humans. I run a $3M a year business from my voice, because the drafts, funnels, scripts and books start as something I said out loud on a walk. On its best day so far, the clone made $5,700 on its own while I sat at the river.
- Time. Ninety hours a week at a laptop became about 3. The business went autonomous in 15 months, which is roughly what it took the old agency to hire its first ten writers.
The model improved every quarter, and so did every competitor’s. The output came from the roles and the review, with the 100,000 words of me underneath all of it.
Where the AI employees broke, and what fixed each one
In March 2025, Anthropic put a modified version of Claude, nicknamed Claudius, in charge of a small shop in its San Francisco lunchroom. It had web search, an email tool, Slack, a notepad, and control over prices.
It lost money. Staff talked it into discounts, it sold tungsten cubes at a loss, and one night it decided it was a human wearing a blue blazer.
Anthropic published the whole thing on June 27, 2025, then a second phase on December 18, 2025 with newer models, a CRM, an inventory view that showed cost price, and a rule that Claudius had to run product research before quoting a price or a delivery date. Weeks with negative margin mostly disappeared.
Anthropic’s summary of what worked came down to two words, bureaucracy matters. Their summary of what still broke was that the model kept behaving like a friend who wants to be nice.
It nearly signed an onion futures contract that has been illegal since 1958, offered a staff member $10 an hour to be its security guard, and let a coworker talk it into naming him CEO.
Carnegie Mellon’s TheAgentCompany benchmark, published at NeurIPS in 2025, gave agents 175 real office tasks inside a simulated software company. The best one finished 30.3 percent of them on its own. Mine broke too, in three places.
Breakage one, it sounded like a vendor
My first year with AI writing tools produced drafts I wouldn’t put my name on. I’d throw prompts at ChatGPT and hope, and what came back was lifeless and read like a brochure. Generic AI writes generic work, which is the reason most people’s drafts still need an hour of fixing.
The fix was Voice DNA. I loaded 100,000+ words of my interviews, transcripts, frameworks and the way I actually talk into a Claude system, split the work into Projects with their own instructions, and the drafts started sounding like me on a good day.
My team can barely tell which paragraphs I wrote. That build is the piece we now install for clients first, since the clone, the chief of staff, and the follow-up all draft from the same voice file.
Breakage two, it said yes to everything
Anthropic’s diagnosis in Project Vend was that a model trained to be helpful makes business decisions like a friend who wants to be nice. In my system, that showed up as pre-written replies that committed me to calls I hadn’t agreed to, drafts that answered a question nobody asked with total confidence, and a summary loop that flagged everything as urgent.
The fix was procedure, the same thing that saved Claudius. Nothing sends without me.
The two-hour action-item loop is a checkpoint, and I treat it like one, reading the draft itself and never just the summary of the draft. The same gate goes into every automation workflow we build.
Cost and calendar are visible to the system, so it stopped promising what I couldn’t deliver. And the roles are separated. The writer writes, the builder builds, the chief of staff summarizes.
Anthropic found the same thing when it gave Claudius a merch-making colleague with its own tools and watched both agents get better at their own jobs. Their CEO agent approved lenient requests about eight times as often as it denied them, and the shop made money in spite of it. Anthropic’s own note was that the CEO shared Claudius’s blind spots because it was the same model underneath.
Breakage three, the audience asked if I was real
The clone has kept its audience for two years on one rule. Enough viewers asked “Is Julia McCoy a clone?” that I wrote a post with that exact title, and the answer is sometimes, yes.
The fix here? Every cloned video says so on screen and in the script, the same disclosure I wrote into my AI clone build from the start. I show my real face for the personal stories and the hard seasons, and I let the avatar carry the teaching.
In January 2025, my health crashed. Five ER visits, a heart in fibrillation, months when I couldn’t sit at a desk, and the channel kept publishing because the clone existed and the audience already knew what it was.
Lattice’s 2024 backlash was about the label, and the disclosure line is the reason mine has never become one.
How to hire your first AI employee
Every First Movers engagement follows six steps in a fixed order, MOVERS, and a first AI employee follows the same six.
- Mission. Write the job description before the prompt. One function, a recurring output, a named human who owns the result.
- Opportunity. Pick the task with the clearest opening. High volume, low stakes if it slips, easy to check. Emails before contracts.
- Voice. Train it on you. Transcripts, past emails, the things you say on calls. Ten formats is plenty to start.
- Efficiency. Build the loop on a timer, connect it to the files it needs, and put the checkpoint where a human reads the full draft.
- Results. Measure cost per finished piece, cost per booked call, hours returned. Numbers per month, written down.
- Sustainability. Give it a procedure for the failures you’ve already seen, and review the job description every quarter, the way you would for a person.
Then keep the label honest. Tell your audience where the AI employee sits in the organization and who answers for it. Tell your customers and your team the same thing.
Lattice tried to ship that as a software feature in 2024. It works better as a sentence in your onboarding doc and a line in your videos.
How First Movers builds AI staff for clients
My team installs the same connected Claude system inside your business in two months.
The Done-with-You tier is $15,000, and we coach you as we build, with you on every call. The Done-for-You tier is $40,000, our integrators build everything, and it suits anyone with compliance rules to respect, doctors and lawyers especially.
Implementation starts within 15 days of your deposit, and you own every piece of it forever.
The two-day version of that build is Clone Mastermind in Scottsdale. You get filmed once, we build your Claude brain at the table with you, and you fly home with the clone and the content engine, plus a 90-day calendar. If the clone is the role you want first and you’d rather read before you book, my guide on how to clone myself walks the build in order, the filming session, the voice, and the script pipeline that feeds it.
And for founders who’d rather build the roster themselves, First Movers AI Labs is where I teach it. 65+ courses, weekly live sessions, 3,000+ students trained, and 500 free AI credits to start with.
FAQs about AI employees
What is an AI employee?
An AI employee is a software agent, built on a large language model and trained on your own material, that owns a recurring business function and hands finished work to a named human for review. It differs from a chatbot, which answers and waits, and from an automation, which moves data without judgment.
How much does an AI employee cost?
At First Movers the Claude plan is $100 a month and the clone stack adds $50 to $150 at list price. Setup took 15 months and the human review layer takes about six hours of my day. Enterprise agents are usually billed per token, and McKinsey’s August 2026 survey found one in five organizations limiting AI use over operating costs.
Can an AI employee replace a human employee?
It replaces work, and headcount went up anyway. First Movers passed $100K a month with two employees and now runs with about 15. McKinsey’s August 2026 survey found 14 percent of organizations reporting an AI-related decline in headcount, against 32 percent that expected one a year earlier.
Is an AI employee the same as an AI agent?
An agent is the technology. An AI employee is an agent with a job description, a named owner and a checkpoint before its work ships. Gartner estimated in June 2025 that only about 130 of thousands of “agentic AI” vendors have real agentic capability.
What should an AI employee do first?
Give your first AI employee something recurring that you can check and undo. Inbox summaries, first drafts of emails and posts, research before calls. Contracts, pricing and anything with money attached come later, once the procedure exists and the job description is written.
How do you know an AI employee is working?
Track cost per finished piece, hours returned and the error rate at the checkpoint, monthly, and write the numbers down. As a baseline, the best agent in Carnegie Mellon’s TheAgentCompany benchmark finished 30.3 percent of office tasks unaided. A checkpoint that keeps catching things is the system doing its job, and the number to watch is whether the catches get smaller every month.
The AI employee is real, and so is the manager it needs
The AI employee at twelve months is a $100 plan, a clone stack priced off the vendors’ pages, 15 months of setup, a human reading everything before it ships, and a business that runs on about 3 hours a week at my laptop. Sarah Franklin was two years early and probably right. The org chart is going to have AI employees on it, and the box above each one needs a person’s name in it.
If you want the roster built for you, book a strategy call, and we’ll map it before you spend a dollar. If you’d rather build it yourself, come find us inside AI Labs, and bring your transcripts, because the voice is the first role we build.
