In short
- Hire for one job, not for "AI". The best first job repeats weekly and follows rules you can write down.
- Write the brief like a job post: goal, schedule, tools, what it reads, what it must ask before doing.
- Start with everything on approval. Loosen the rules only after a trial week with real work.
- Judge it like a hire: time you got back, how much you edited, what it got wrong.
Most small companies don't fail at AI agents because the technology is weak. They fail because they never decide what the agent is for. "Let's use AI for marketing" is not a job. "Every Monday at 09:00, chase invoices that are more than seven days late, politely, and ask me before each email goes out" is.
Hiring an agent works like hiring a person. You pick a role, write a brief, give access to the right tools, agree on what needs sign-off and check the first week closely. The difference is that the agent starts on Monday and never needs the onboarding talk twice.
Start with a job, not with AI.
Look at the work your team does every week and nobody wants to do. That list is your hiring pool. A good first job for an agent has five traits. Tick the ones that fit the job you have in mind:
Tick what applies. Four or five ticks: hire an agent for it this week.
If a job ticks fewer than three, it's probably a project, not a role. Projects need judgment that changes every time. Keep those with your people and give the agent the repeatable work around them.
Good first hires for a small team.
These are the roles small companies tend to hand over first, because they repeat, they follow rules and they can wait for a yes before anything goes out.
Sends invoices on the 1st and chases the late ones on a polite schedule.
Sorts the inbox every morning, drafts replies and proposes meeting times.
Answers the questions your docs already answer and hands over the rest.
Notices deals that went quiet and drafts the follow-up for you to send.
Pulls the numbers into one weekly report per client, in plain words.
Takes each new client from signed to running without a dropped step.
Pick one. Not three. The point of the first hire is to learn how your company works with an agent, and that's easier to see when only one thing changed.
Write the brief like a job post.
A vague brief gets vague work, from people and from agents. Write it the way you'd write a job post for a careful new assistant. Seven lines are enough:
- Goal
- Get paid on time without the team writing reminders by hand.
- Schedule
- Weekdays at 09:00. Invoices go out on the 1st of each month.
- Tools
- Xero for invoices, Gmail for reminders.
- Reads from
- Each client's living file: agreed terms, who receives invoices, any payment plan.
- Rules
- First reminder at 7 days late, second at 14, then flag it to me. Friendly tone, never threats.
- Asks before
- Every email to a client. Any invoice above $3,000.
- Done looks like
- A Friday summary: who paid, who's late, what I need to do.
Write it in your words.You don't need prompt tricks. If a new colleague would understand the brief on their first morning, the agent will too.
Brainstack agents come with the brief, schedule and approval rules already written. You edit, you don't start from zero.
Give it your company's context.
A new hire is only as good as what they know about your clients, prices and past decisions. The same goes for an agent. If it has to ask you "what did we agree with Northwind?" every morning, you've hired a very fast intern.
That's why Brainstack agents work from a company brain: it reads the tools you connect, keeps a living file for every client with a source for each fact, and only shows each person and each agent what their role allows. The billing agent knows that Northwind pays the finance inbox, not Sara. It knows because someone wrote it in Slack in September, and the brain kept the link.
If you're wondering how that differs from uploading files to ChatGPT, we wrote a whole piece on it: company brain vs ChatGPT with files.
Set the approval rules before day one.
Decide what the agent may do alone and what waits for a person. Start strict. You can always loosen a rule; it's much harder to unsend an email.
| When the agent wants to… | Week one | Once it has earned trust |
|---|---|---|
| Read and summarise | On its own | On its own |
| Draft an email or a post | On its own | On its own |
| Send to a client | Asks you | Routine messages alone, the rest asks |
| Publish | Asks you | Asks you |
| Spend money | Asks you | Alone below a limit you set |
| Delete anything | Asks you | Asks you |
Run a trial week.
Don't judge an agent on a demo. Judge it on five real working days, with every outgoing action on approval. Here's what a good trial week looks like:
Connect the two tools, paste the brief, run it once by hand.
You edit most draftsEvery edit you made on Monday becomes a rule in the brief.
Fewer editsIt runs on schedule. You approve from your phone.
Mostly approvalsGive it the awkward client. See what it asks you.
It asks the right questionsKeep, adjust or let it go. Then loosen one rule, not all.
You trust the routineMeasure it like a hire.
At the end of the week, answer three questions. You don't need a dashboard for this, although every Brainstack agent has one.
- Did it give you time back? Count the hours the job used to take and the minutes you now spend approving.
- How much did you edit? If you rewrote most drafts on Friday too, the brief is still missing something.
- What did it get wrong, and did it ask first? Mistakes caught at approval are the system working. Mistakes that went out are a rule you need to add.
The goal of week one isn't a perfect agent. It's a brief good enough that next week you only approve.
When to hire the second one.
Hire the next agent when the first one runs for two weeks with only approvals and no rewrites. Pick the job next to it: billing leads to financial reporting, support leads to onboarding, follow-ups lead to proposals. Each new agent reads from the same company brain, so the second hire starts with everything the first one taught it.
That's how a small team ends up with a full back office: one agent at a time, each one earning its place.