What Is an AI Employee, Really? Examples for a 20-Person Business

"AI employee" is one of those phrases that sounds either exciting or alarming depending on who says it. Vendors use it to mean anything from a chatbot to a fully autonomous system that supposedly runs a department on its own. Neither picture helps an owner of a 20-person business decide what to do on Monday.

Here is a more useful definition, and what it looks like in a business that size.

What an AI employee actually is

An AI employee is a defined job, not a digital person.

It has four parts:

1. A job description. The steps, written down: what to read, what to check, what to produce and in what format. In Claude this is usually a skill; in ChatGPT, a custom GPT or project with instructions.
2. Access. The systems it can read, such as the inbox, the CRM, the accounting system or the shared drive, and the few it is allowed to write to.
3. Knowledge. How your business does things: your products, your terms, your tone, your rules. Without this, the AI produces answers that would suit any business except yours. This is the job of a second brain.
4. A manager. A named person who owns the result, checks the drafts and corrects it when it gets something wrong.

The technical word for this is an AI agent: an AI tool that can take several steps on its own, reading something, deciding what to do next and producing a result. "AI employee" is just the business framing. It has a job, a schedule and someone it reports to.

What it is not: a replacement for judgement, a person who can be held accountable, or something you switch on and walk away from.

Examples by function in a 20-person business

These are the kinds of jobs that suit an AI employee: recurring, mostly reading and drafting, with a clear "done" and a person who can check the result. The exact list depends on your business.

Marketing

• The weekly performance report. Pull last week's website, ads and email numbers, compare them with the week before, and write a one-page summary of what changed and what is worth looking at. The marketing lead reads it, adds a judgement, and decides what to do.
• The content calendar. Draft next month's posts from the plan, in the brand's voice, and lay them out on the calendar. A person approves each post before it is scheduled.
• Search monitoring. Check weekly which pages are gaining and losing search traffic and flag the ones worth updating.

Customer service

• First-draft replies. Read incoming enquiries, sort them by type, and draft a reply for each using the business's policies and tone. A person reviews and sends. The AI never sends on its own.
• The FAQ that writes itself. Notice questions that come up again and again and propose an answer for the help page.

Finance

• Invoice chasing. Each week, list unpaid invoices past their due date, draft a polite reminder for each, and flag the ones that need a phone call instead. A person approves before anything is sent.
• Spend review. Sort the month's transactions, flag anything unusual or uncoded, and prepare the questions for the bookkeeper.
• Cash flow notes. Prepare a short commentary on the month's numbers for the owner, from the accounting system's reports.

Operations

• Meeting notes into actions. Turn the notes from each meeting into a list of actions with owners and dates, and post them where the team will see them.
• Process pages kept current. When a job changes, draft the update to the page that describes it, so the next person does it the new way.
• Compliance evidence. Gather the records an audit or a client questionnaire asks for, and note what is missing.

Sales

• Call preparation. Before each sales call, pull together what the business knows about the prospect: past emails, the CRM record, their website. A one-page brief, not a guess.
• Proposal first drafts. Draft a proposal from the call notes and the standard offer pages, for the salesperson to rewrite in their own words.

None of these is glamorous. That is the point. They are the jobs that eat hours every week and that nobody hired anyone to do.

What a person still owns

An AI employee changes what people spend their time on. It does not remove the need for people.

Decisions. Which campaign to run, which customer to chase, which supplier to drop. The AI can lay out the options and the evidence. A person chooses.

Relationships. Customers, suppliers and staff deal with people. The AI can draft the email; the relationship belongs to whoever sends it.

Accountability. If something goes wrong, a person answers for it. That is why every AI employee needs a named owner, not "the team".

Teaching it. When the AI gets something wrong, the owner corrects it, and the correction is written down so it does not happen again. This is the part people underestimate. An AI employee gets better because someone takes the time to tell it what good looks like.

Where the approval gates sit

The simplest rule: anything that leaves the business, or changes money, records or a customer's experience, waits for a person.

In practice that means:

• Reading is fine. Let it read the systems it needs, within the permissions the owner sets.
• Drafting is fine. Drafts are cheap and a person can throw them away.
• Sending needs approval. Emails to customers, posts on social media and messages to suppliers are left as drafts for a person to approve.
• Writing to core systems comes later. Changing a record in the CRM or the accounting system is the last thing to switch on, and only for jobs that have run reliably for a while, with a person signing off.

Start with read and draft. Move a job to fewer approvals only when the owner has seen it get the work right, week after week.

How to start

Pick one department and one job. The best first job is one somebody does every week, dislikes, and could describe step by step. Time it as it runs today, before anything is built. Build it with the person who does it. Then time it again a week later. If it is faster and still right, keep it and pick the next one. If it is not, fix it or drop it.

That is how AI Departments works: one department at a time, set up in the Claude or ChatGPT your team already uses, wired into your second brain, with drafts left for a person to approve and each job measured against your own week-one baseline. Browse the departments to see the jobs each one covers.