You can hire an AI agent three ways: subscribe to a ready-made one, pay a developer to build one, or pay a service to build one and keep it working. Whichever you pick, hire it like a person. Write a job description, give it a trial period where it drafts and you approve, and review it after 30 days.
An AI agent is software that does a defined job inside your own tools, such as your inbox, calendar and CRM. It works on a schedule instead of waiting for you to ask. What your money buys differs a lot between the three. A subscription gets you the software, and you do the setup. A developer gets you a build, and a service gets you the build plus someone who keeps it running.
VantaSoft runs on 13 AI agents of its own, and each one works from a written job description. This guide is what we learned giving them work, including the permissions we widened and one we took back. We also sell the third option, so we say where the others fit better.
Who can you hire an AI agent from?
There are three kinds of seller. Software companies sell a ready-made agent by subscription. Freelancers and agencies build one to order. Services build one and then keep it running for a monthly fee. You can also skip hiring and set up free, open-source agent software yourself. Here is what each one published on November 5, 2026.
| What you hire | Example | Published price | Who sets it up and keeps it working |
|---|---|---|---|
| A ready-made agent, by subscription | Lindy | $50 of free credits to start, then from $29.99 a month per user for 3,000 credits | You, inside the vendor's app |
| A set of AI helpers for common tasks | Sintra | $48.50 a month on a sale (normally $97), or $15.60 a month on a 12-month plan, with 250 credits a month | You |
| A freelancer or agency that builds one | Many | Usually quoted per project after a call | Whoever you pay, for as long as you keep paying them |
| A service that builds and maintains one | VantaSoft | Build from $1,000 one time, then hosting and maintenance from $200 a month | We keep it running, and your team directs its work |
| Nobody: you set up open-source software | Hermes Agent, OpenClaw | No license fee; you pay for a computer that stays on and for AI usage | You, all of it |
The prices come from Lindy's, Sintra's and our own pricing pages. Credit-based prices change often, so check again before you buy.
Read the last column before the price column. The subscription and do-it-yourself routes look cheapest because you are the one setting them up, connecting them and fixing them. That is real time, and it comes back every week. A subscription suits a standard job that the vendor already does well. A build suits a job that runs across several of your systems and follows your own rules.
If you would rather compare tools by the job they do, our guide to AI agents for business sorts them that way. For what developers charge, see the cost to build an AI agent and nine AI agent development companies and what they charge.
How do you hire an AI agent, step by step?
Hire an AI agent the way you would fill a new role. Pick one job and write it down. Choose who to hire from, and ask who keeps it working. Start it on drafts you approve, then review it after 30 days. Most of these steps are decisions only you can make, and they come before anyone builds anything.
- Pick one job, not a department. "Answer every new web inquiry the same day with a reply I approve" is a job. "Help with sales" is not.
- Write the job description. The next section has a blank to copy.
- Choose who to hire from, using the table above.
- Ask who keeps it working. Logins expire, and the apps it works in keep changing. Get the name of whoever fixes it in month three, and get it in writing. Our guide to choosing an AI agent development company has nine questions for that conversation.
- Start it on draft-and-approve. For the first weeks, it drafts each message, booking or change, and you approve it before it happens.
- Review it after 30 days. Count the drafts you sent unchanged, the ones you edited and the ones you threw away.
- Widen what it may do on its own one channel at a time, and only for work it keeps getting right.
For a closer look at each decision only you can make, such as the approval line and who it asks when it is unsure, see how to implement an AI agent in your business.
What goes in an AI agent's job description?
An AI agent's job description says what the job is and what the agent does alone. It also lists what the agent drafts for your approval, what it never does, which logins it gets, who it answers to, when it works and how you will judge it. Keep it to one page. Here is a blank, with an example for a small office inbox.
| Line | What to write | Example: the main inbox at a small service business |
|---|---|---|
| The job | One sentence with a start and a finish | Read every message to the main inbox and make sure each one gets a reply, or goes to the right person, the same business day |
| Does on its own | Low-risk work that is easy to undo | Labels each message, files receipts and newsletters, adds new inquiries to the CRM |
| Drafts for your approval | Anything a customer will read | Replies to new inquiries and to complaints |
| Never does | Your hard lines | Quotes a price, promises a date, issues a refund, deletes mail |
| Logins it gets | Each app, and whether it can only read or can also make changes | Inbox: read and draft. CRM: read and edit. Calendar: read only |
| Answers to | One named person, and where it asks questions | The office manager, in your team chat, such as Slack or Microsoft Teams |
| When it works | Its schedule | Every 30 minutes from 7 a.m. to 7 p.m. on weekdays |
| How you judge it | What you will count at 30 days | Messages that waited past the same day, and drafts sent without changes |
Each of our agents works from a page like this, and two lines matter most. The first is the never-does line. When we added an editor agent in October, her job description named what she edits, which is proposal drafts. It also named three things she never does: choose which bids to chase, set prices or submit anything.
The second is where corrections go. When one of us corrects an agent, the fix goes into its job description as a dated rule. A correction said once in chat can get lost, and the written rule is there every time the agent starts work. Our chief of staff agent's job description holds about a dozen dated rules, and more than half were added right after one of us corrected him, as we describe in what our AI chief of staff does.
Keep the logins line as narrow as the job allows. Our guide to what access an AI agent should have shows how to decide app by app.
How much should a new AI agent do on its own?
Very little at first. Give a new agent the same trial period you would give a new hire. It starts by reading and reporting, then drafts work for you to approve. Only after that does it act alone on one channel under written rules. Money, prices, hiring decisions, legal commitments and anything sent in your own name stay with a person.
| Stage | What the agent may do | When to move it up |
|---|---|---|
| Read and report | Reads your systems and tells you what it found | Its reports stop needing corrections |
| Draft and approve | Drafts each message or change, and you approve it first | You keep sending its drafts without changes |
| Act within written rules | Acts alone on one channel and tells you afterward | Widen it one channel at a time |
| Never alone | Payments, prices, hiring and firing, legal commitments, anything in your own name | It doesn't move up |
Our own agents sit at different stages for different work, and one agent can sit at two. Our content agent publishes to our company blog on its own every morning. An article that would go out under our CEO's own name waits until he approves that piece. Our sales agent started on drafts only. She now answers replies on one owner's sales messaging account without asking, after three replies sat for a day waiting for him, while her email still waits for approval. The rules she works by are in how our sales agents handle follow-up.
Authority can go back down, too. In late August an owner took back our content agent's permission to make small changes to our website. For about six weeks, every change outside the blog went to an engineer as a request. In early October he gave it the whole site, with a short list still kept for an owner, such as domains, hosting and the legal pages.
The next day, the job that publishes our articles pushed five unrelated website changes live along with the article. The limit had been written into its instructions but not built into the publishing tool, so we built it into the tool. When you widen what an agent may do, have the software enforce the limit as well as the job description. The full story is in what went wrong with our AI agents.
Frequently asked questions
Can I hire an AI agent instead of an employee?
For a narrow job that repeats, an agent can take the work. In our experience it works best taking that work off a person who is already stretched, rather than replacing a role. Judgment calls, customer relationships and anything with legal or financial weight stay with your people. Our comparison of a virtual assistant and an AI agent sorts common admin jobs between the two.
Is hiring an AI agent the same as hiring an AI agent developer?
No. The agent is the worker, and the developer builds it. If you hire a developer, you also need someone to keep the agent working after the build, because logins expire and the apps it uses keep changing. Ask any developer who does that in month three, and get the answer in the contract.
How much does it cost to hire someone to build an AI agent?
It depends on the job and on how many of your systems it touches. Most development companies don't publish a price. When we read nine of them on October 21, 2026, six published none. Our build starts at $1,000, with hosting and maintenance from $200 a month. The cost to build an AI agent collects the published ranges.
Do I need technical staff to hire an AI agent?
Not for a subscription or a done-for-you service, though someone on your team still has to write the job down and approve its work. You do for open-source software you set up yourself. Somebody has to install it, connect it to your apps and fix it when an update breaks something. What Hermes Agent costs a small business to run puts numbers on that route.
Can an AI agent help me hire people?
It can take the admin around hiring, such as sorting applications against the requirements you set, scheduling interviews and sending updates to candidates. Who to interview, who to hire and who to turn down should stay with a person.
Where to start
If you want an agent built and kept running for you, it starts with 30 minutes on your workflows. Start Automating and bring the one job you would hand over first. Not sure which job that is? Our five-question quiz on which AI agent to hire first suggests a starting role, and the agent directory has a job description for each role we build.
