Hiring an AI agent works the same way hiring a person does. You define the role, give it the tools, set the voice, manage the work, and hold it accountable for output. That is the five-part checklist we run before any agent we build goes live, and every failed agent I have been asked to look at skipped at least one of the five.
My opinion up front: most first AI hires fail before the technology gets a chance to. The owner signs up for a tool, points it at a vague goal, and waits. Asking ChatGPT for help is asking a smart friend for advice. Hiring an agent is hiring an employee. Nobody would bring on an employee with no job description, no logins, and no manager, but that is exactly how most small businesses adopt their first agent.
This guide walks through each part of the checklist the way we actually run it: what good looks like, what skipping it costs, and the questions to ask any vendor, including us, before you sign.
1. The role: write the job description first
Roles are narrow on purpose. Narrow scope is what makes the work reliable.
Compare two owners. The first hires “a marketing agent.” No defined output, no cadence, no place the work lands. Within three weeks the novelty wears off, the outputs drift, and the tool joins the graveyard of subscriptions nobody remembers to cancel. The second hires an agent that drafts her weekly LinkedIn posts in her voice, pulls from her call notes, and queues every draft for approval on Thursday morning. Six months later it is still running, because “still running” was defined on day one.
A real job description for an agent fits on a page:
- A title a human would recognize. Content assistant. Invoice processor. Report builder.
- One outcome, stated as a deliverable. Not “help with marketing” but “five drafted posts per week.”
- The inputs it works from and where they live.
- The cadence and the handoff. When does work appear, and who signs off?
- What it escalates. When something is ambiguous, it asks a human instead of guessing.
Which role first? The best first hire is usually a recurring weekly chore with clear inputs and a checkable output. Drafting the posts you already know you should publish. Processing the invoices that pile up every Friday. Building the report you assemble by hand each Monday. Pick something you would happily hand a competent junior employee in their first week, because that is the trust level you are starting at.
One more rule we hold clients to: one role first. The second agent is easier to hire once the first one is boringly reliable. If you want to see what narrow roles look like in practice, we published the five jobs small businesses hand an agent first.
2. The tools: give it access to where the work already lives
An employee with no logins produces nothing. Same for an agent.
The good news: this rarely means new software. The agents we build live in the tools a business already runs. Slack, Teams, email, the spreadsheet everyone actually uses. One 60-person company we talked to runs entirely on SharePoint and email, and their first worry was that agents meant adopting a whole new stack. It is the opposite. The agent meets your systems where they are.
Two rules govern access.
Read-only first. An agent earns write access the way a new hire earns it, by proving judgment on lower-stakes work. Start where it can look but not touch, and expand deliberately. We wrote up how we stage that on How It Works.
Tier the data. Marketing data, like your published posts and website analytics, is low-stakes. Financials and customer records are not. We set agents up zero-storage and no-training, meaning your data does not persist in the model or train anyone else's, and the high-stakes systems stay read-only until you decide otherwise.
Done right, tool access is where agents get surprisingly capable. A two-person wine retailer we work with can photograph a bottle on the shelf, and their agent finds the matching invoice in the records. No new interface. A camera, a chat thread, and access to where the answer lives.
3. The voice: it should sound like you
Before any agent writes a word on a client's behalf, we build a one-page voice guide from their real writing. Sent emails, old proposals, the LinkedIn post that got replies. It captures three things: stance (the opinions you state plainly that competitors hedge on), rhythm (whether you write in short declaratives or long careful qualifiers), and a banned-word list (the hype words you would never say across a table from a customer). The agent drafts against that guide, and drafts that miss it get rewritten, not published. We hold our own content to the same gate, including this page.
Skip this step and you get the fate of most AI content: grammatically perfect copy that could belong to any company in any industry. Your customers can tell. Generic output quietly spends the trust you built with the people who know how you actually talk.
Voice matters beyond marketing copy, too. An invoice agent writes payment reminders. A customer-questions agent answers a frustrated email. Every one of those messages is your business talking. Decide what it sounds like on purpose.
4. The management: every agent needs a manager
The part demos never show you: an unmanaged agent decays. Models get updated. A tool changes its export format. Your offer changes and the agent keeps describing the old one. Set-and-forget agents do not fail loudly on day one. They rot quietly over weeks, until the person who notices the outputs went bad is a customer.
This is why we treat management as part of the hire, not an add-on. Someone reviews output quality every week (we call it watering), catches drift early, retunes when the business changes, and checks the edge cases. Our clients get a dedicated account lead, a concierge who manages the agent team day to day and is the human you talk to in Slack.
It is also the honest answer to a question prospects ask on almost every call: why is there a monthly fee if the agent runs itself? Because it does not run itself. Nothing that does real work in a changing business does. You are not buying software seats; you are paying for a role that stays managed. An unmanaged agent is a liability with good grammar.
5. The accountability: deliverables and approval gates
An employee is accountable for output, and so is an agent. Accountable for output means deliverables you can inspect: the drafted post, the processed invoice batch, the finished report. Activity is not a deliverable.
Two mechanisms make that real.
Approval gates. In month one, a human approves everything before it goes anywhere. You approve; it sends. A mistake gets caught at the gate, in a review queue, instead of in front of a customer. As the work proves out, you loosen the thresholds deliberately, the same way you would stop shadowing a new hire.
A measurable definition of done. One team we set up processed 450 products through their catalog in a single day, three people plus an agent, and every batch passed a human check before it touched the live system. That number made the case better than any pitch could: work you can count, verified by someone whose name is on it.
One boundary we hold everywhere, including in our own sales work: agents assist humans. They do not replace human judgment on outcomes that matter. An agent can prep the lead research and draft the follow-up. A person decides what gets sent and owns the relationship.
The checklist, on one page
Before your first AI hire goes live, you should be able to answer yes to all five:
- Role. Is there a one-page job description with a single, countable deliverable?
- Tools. Does it have access to the systems where the work lives, read-only where the stakes are high?
- Voice. Has it been trained on how your business actually sounds?
- Management. Is a named human reviewing output weekly and retuning as things change?
- Accountability. Are there approval gates, and does someone own the definition of done?
A no on any line is not a reason to give up. It is the work to do before go-live, and it is exactly what the build phase of our process covers.
Questions we get on nearly every call
How is hiring an AI agent different from just using ChatGPT?
ChatGPT is a smart friend you ask for advice: useful, general, and it forgets you between conversations. An agent is an employee. It has a defined role, standing access to your tools, your voice, and a manager checking its work. The difference is the checklist above. We wrote a fuller comparison in software you operate vs labor you manage.
How long does it take to get a first agent live?
Two to three weeks from kickoff for the businesses we work with. Most of that time goes to the checklist, not the code: defining the role, wiring tool access safely, and building the voice guide.
Can an AI agent replace a hire?
Sometimes it replaces the output of a role you were about to hire for, at a fraction of the cost of a salary. It still needs management, which either you do or we do. Anyone who tells you it replaces the whole job, judgment included, is selling something.
Can an agent touch our financial or customer systems safely?
Read-only first, always. An agent can pull the numbers for a report without the ability to change a single record. Write access to high-stakes systems comes later, behind approval boundaries, after the agent has proven judgment on work where a mistake costs minutes instead of money.
What happens when the agent makes a mistake?
In month one, nothing reaches the outside world without human approval, so mistakes die in the review queue. We review every output in the first month, then loosen thresholds together as the error rate earns it.
Run the checklist before you sign anything
The checklist is five lines. Any vendor who wants to build you an agent should be able to answer, specifically, how they handle each one: the role they will define, the access they need, how they capture your voice, who manages the thing after launch, and where the approval gates sit. If the answers are vague, the agent will be too.
If you want to hear how we answer them, book an intro call. Bring the messiest job on your plate. It is usually a better first role than the one you had in mind.
Where to start
New to the category? Start with what an AI employee actually is, or see the five jobs owners hand over first in AI agent examples. For the job-by-job breakdown, the small-business guide to AI agents goes deep on each role.
Matt writes Field Notes, a newsletter on running AI agents inside a real business. Sign up here.