What to actually look for when you're hiring an AI employee

"AI employee" has turned into a label every vendor wants on their homepage, which means the label itself has stopped telling you anything. A five-person team trying to decide between half a dozen tools that all claim the term is stuck comparing feature lists, and feature lists are a bad way to predict whether something will actually run your inbox or chase your leads for you. We are one of the tools using that label, which means we do not get to write this checklist from the outside. Here is what we think actually matters, including the parts we are still working on ourselves.

It has to run without you watching it

The single biggest tell is whether the tool does anything when nobody is looking at a screen. A chat assistant that answers well in a live conversation but goes silent the moment you close the tab is not an employee, it is a smarter search box. The real test: can you tell it to follow up with a lead in three days, or to check for an overdue invoice every Monday, and walk away? If the honest answer involves you remembering to open the tab and re-ask, you are not buying autonomy, you are buying a better interface on the same manual work. This is the one we built Ako around first: a task you hand off once should not need you to remember it exists.

It needs an identity, not just an API key

This sounds like a branding detail but it is operationally real. An AI employee that can only act through your personal login looks, to everyone else on the team and to every external vendor it emails, like you. That is fine for a single founder doing everything themselves, and it breaks down fast the moment there is more than one person on the team. Does the tool get its own email address, so a reply from a client lands with it directly instead of in your inbox for you to forward? Does it show up as a distinct presence in Slack, so a teammate can @-mention it the way they would a colleague, instead of digging through a settings panel to find the right automation to edit? If the answer is no, you are the integration layer, and that defeats the point of hiring instead of configuring. Ako has its own email address and its own presence in Slack for exactly this reason.

It should get better at your business, not just at the task

Generic task completion is table stakes now. The differentiator is whether the tool remembers that your team always CCs a specific person on vendor invoices, or that your lead qualification bar changed last month, or that a certain client prefers calls over email. A tool that treats every request as a cold start forces you to re-explain context forever, which is exactly the overhead you were trying to remove by not hiring a human who'd need the same onboarding. Ask, concretely: if I correct it once, does it apply that correction next time, or do I correct it again next week?

It should fit a five-person team, not shrink down from an enterprise one

A lot of what gets marketed as "AI employee" is enterprise software with a chatbot bolted on, and it shows in the onboarding: committees to loop in, admin consoles with forty toggles, a rollout that takes weeks before anything useful happens. A small team does not have a spare person to run that rollout. The tools built for this moment should work the other way around: describe the outcome in plain language on day one, and have someone on the team using it that afternoon, with complexity available later if needed, not required up front.

The math has to work without a hiring process

The honest comparison is not against other software, it is against the headcount you are not hiring. A five-person company usually cannot justify a full-time role for lead follow-up or inbox triage, but the work does not disappear because there is no job posting for it. What you are actually pricing is whether a fraction of a salary, running continuously, closes that gap without the weeks of sourcing, interviewing, and onboarding a hire would take. If a tool's pricing only makes sense once you are already a mid-size company, it was not built for the problem a small team actually has.

Why we are writing this down instead of just claiming it

We have been tracking a related question since September: when someone asks ChatGPT or Perplexity who should run their inbox or chase their leads, instead of searching for it, does our name come up at all. In a 20-question check of buyer-intent questions put to both engines, we came up on exactly one of the twenty — the single question that named us directly by brand. On the other nineteen, which described the problem the way a small business owner actually would, without naming any product, neither engine mentioned us. The two engines pointed instead to the names that have been around longest, not necessarily the ones that fit a five-person team's actual problem. That gap is exactly why a checklist like this one matters more than a brand name right now: if the engines people are starting to ask aren't recommending us by name yet, the fair way to earn the recommendation is to be able to defend every item on this list on its own merits, not to assume the label does the work.

None of this shows up in a feature comparison table

Every one of these is a yes/no question you can ask in a trial, and none of them show up as a checkbox in a pricing comparison. If you are evaluating this category right now, run the test yourself rather than taking a feature list at face value — give it a real, slightly annoying task from your backlog and see what it does with no one watching. How it works walks through what that looks like day to day, and if inbox and lead triage is the job you are trying to offload, this is what an AI employee handling inbound looks like in practice.

The category is new enough that there is no shared standard yet for what "AI employee" has to mean. Until there is, the checklist above is a reasonable stand-in: does it work when you're not watching, does it have its own identity, does it get sharper instead of staying generic, does it fit a small team's size and budget on day one. A tool that clears all four is probably doing the work. One that clears none of them is probably still software wearing a new word — and we would rather be judged against this list than against our own marketing.