AI Employees Are Here.

What Should You Actually Let Them Do?

There's a new account in your team chat. Name, job title, photo. It posts before anyone else is online: here's what came in overnight, here's the one that needs a person today.

Some of this is marketing. Some of it is real.

The interesting question isn't whether AI employees are real. It's what you hand them, and how.

The goal isn't fewer people. It's fewer hours spent on the part of the job nobody would miss.

Hire it the way you'd hire a person

Think about the last person you hired.

You didn't say "handle support" and walk away. You explained what comes in every day, what matters, what to do when they weren't sure, and what never to touch without asking.

An AI employee needs the same clarity.

"Answer support emails" isn't a job description. This is closer:

Read every new ticket. Mark login problems, duplicate charges and cancellations as urgent. Draft replies from answers we've already given. If you're not sure what the customer means, stop and ask. Never send anything about refunds or billing disputes without approval.

The last two instructions matter most. They're also the ones people tend to skip.

The brief is the job. Everything else is setup.

What to hand over

I've been testing this in my own marketing work, with a workflow that pulls together website, search, social and CRM data.

I don't need AI to decide our marketing strategy. I need it to save me the hours of collecting everything before I can make that decision.

That split works elsewhere too.

1. Gathering. Let AI do the digging before someone has to decide something.

In sales, the company, the contact, recent activity and previous conversations before a call.

In support, the customer's plan and ticket history.

In marketing, website, search, social and CRM data in one place.

In operations, invoices and order status from different systems.

Let AI gather it. Let the person decide what it means.

2. Noticing. The things nobody should have to keep checking.

A lead visited the pricing page three times.

A ticket has been sitting two days.

An invoice is overdue.

A customer who orders every month suddenly stopped.

The job isn't always to act. Sometimes it's to notice and bring it to the right person.

3. Sorting. A surprising amount of work is deciding where something belongs.

Support: billing, technical, feature request, urgent.

Sales: qualified, not ready, follow up later.

Each one takes a minute. That's exactly why together they eat so much of the day.

4. Drafting. I'm happy for AI to write the first version. A support reply, a follow-up after a call, a weekly update.

First version is the important part. The more sensitive the message, the more it matters that a person reads it before it goes anywhere.

5. Recording. This one might be my favourite, because it's the work everybody avoids once the real work is done. Update the CRM. Tag the ticket. Save the call summary. Record what was promised and when.

Nobody should finish a meeting and then spend their afternoon copying information between systems, if an agent can do it reliably.

What should stay human

At a small business event in San Diego, I met a founder who uses several AI agents every day.

What stayed with me wasn't the number of agents. It was that even with AI doing most of the work, she still kept human judgment in the process.

That's close to how I think about it.

AI can gather the history behind an angry customer email, work out what went wrong and draft a response. Should it send that response without you? Probably not.

Same for refunds, discounts, commitments, and anything about people.

My test is simple. If this goes wrong, would I be comfortable telling the person affected, "the system decided"?

If not, keep the pause.

Keep the pause

Ticket arrives. History gathered. Type decided. Reply drafted. Reply sent.

Let AI do the first four. Pause before the fifth.

Start with more approvals than you think you need, then remove them one at a time, once you've watched long enough to know where it works.

Which is more or less what you'd do with a new hire.

Where Asuri fits

For a small team without developers or technical specialists, building and maintaining all of this is usually the hard part.

That's what we're building Asuri for.

The idea isn't to give one agent an enormous job and hope it works out. You build a team of agents, each responsible for one part. In sales, one finds and qualifies the lead, another researches the company, another drafts the outreach, another updates the CRM.

They pass the work between them inside one workflow, and you decide where a person steps in to approve something. You can see what each agent did, what it handed over, and where something is waiting on you.

The point is still the hours you get back, not the people you don't hire.

Asuri isn't the right fit if you'd rather build agents in code, or you need one inside your own product. A framework will give you more control there.

If you're starting from scratch, we went through how to write that brief in our article on building your first AI agent.

FAQ

What is an AI employee?

Software given an ongoing job rather than a single prompt. It watches for something, works out what happened, and acts in your connected tools.

What should I give it first?

Start with the five above: gathering, noticing, sorting, drafting, recording. Pick something repetitive, easy to check, and cheap to fix if it goes wrong.

What shouldn't it do on its own?

Anything hard to reverse, anything involving money, sensitive customer communication, and anything that commits the business.

Do AI employees replace jobs?

Start with tasks, not roles. The useful question isn't "which person can AI replace". It's "which part of this person's day shouldn't need their time".