## Key Takeaways

- **The questions change once teams start.** Before hiring an AI employee, people ask about capability. Three weeks in, they ask about management: tools, permissions, memory, and rollout.
- **"What can it do?" is the wrong first question.** The better one is "what do we keep doing manually that we dread?" Capability lists do not predict value; your own recurring work does.
- **Most before-questions are really trust questions.** Who can it see, what can it touch, what happens when it acts. Good answers involve review-first drafts and per-person permissions, not promises.
- **Most after-questions are really scaling questions.** How to share tools with the team, keep standards consistent, and turn one-off wins into recurring work.
- **The gap between the two sets is the adoption curve.** If your team is still asking before-questions after a month, the AI employee has not been given real work yet.

## Two sets of questions, one adoption curve

Talk to enough teams evaluating an AI employee and a pattern emerges. The questions asked before anyone delegates a single task are almost entirely different from the questions asked three weeks later, once the AI employee is drafting reports and chasing invoices.

That gap is worth studying, because the before-questions are mostly about fear and the after-questions are mostly about leverage. A team that moves from one set to the other has crossed the only line that matters: they stopped interviewing and started delegating.

This post answers both sets honestly, in the order teams actually ask them.

## The before-questions

### "What can an AI employee actually do?"

The honest answer is a category, not a list: an AI employee does delegated knowledge work inside your chat tool, using the tools your team already uses. Research with sources, recurring reports, inbox triage, CRM hygiene, spreadsheet building, meeting prep, dashboards. Viktor connects to 3,200+ tools, so the practical constraint is rarely "can he access it" and almost always "did anyone hand him a real task."

The reason the question disappoints as an evaluation method: a capability list tells you what is possible for someone, not what is valuable for you. The teams that get value fastest skip the brochure and audit their own week instead. What do you produce every Monday? What gets copied between tools by hand? What do you dread?

```prompt
Here are the 5 recurring tasks our team does manually every week:
[list them, one line each, with the tools involved]

For each one, tell me whether you could take it over, what you would
need connected, and which one you would start with. Pick the easiest
and do a first version now so we can judge the output, not the promise.
```

That one message replaces a demo call, and unlike a demo, the output is judged on your work.

### "How is this different from ChatGPT?"

A chat assistant answers when someone visits it, and forgets the visit. An AI employee lives where your team works, holds its own context across tasks, runs work on schedules whether or not anyone is around, and uses your actual tools rather than describing what you could do in them. The difference shows up on the second week, not the first day: the assistant starts every conversation from zero, the employee already knows how your team likes the report formatted.

### "Who can it see? What can it touch?"

The trust question, and the most reasonable one on the list. The answer worth demanding from any AI employee has three parts. First, scope: he works where you invite him, and a channel he is not in does not exist for him. Second, permissions: tool access is granted per person and per integration, so connecting your inbox never means your teammates' inboxes. Third, review-first: sensitive actions like outbound email start as drafts a human approves. We wrote up the full model in [how to control where your AI employee works](/blog/how-to-control-where-your-ai-employee-works).

### "What happens when it gets something wrong?"

The same thing that happens with a new hire: you correct it, and the correction should stick. This is the most underrated evaluation criterion, because everything gets something wrong eventually. The question that separates products is whether the correction persists. An AI employee that remembers "we report revenue net of refunds" after being told once compounds; one that needs the same correction weekly is a chat window with extra steps.

## The after-questions

Three weeks in, the fear questions are gone, replaced by management questions. This is what progress looks like.

### "How do we share tools with the whole team?"

The first scaling question. One person connected the CRM, the reports are good, and now three teammates want the same thing. The answer is that integrations can be personal or shared: shared tools like the team CRM or analytics get connected once for everyone, while personal tools like individual inboxes stay individual. The full pattern, including which tools should never be shared, is in [how to roll out an AI employee to your whole team](/blog/how-to-roll-out-an-ai-employee-to-your-whole-team).

### "How do we make the output consistent?"

The standards question. Two people brief the same AI employee differently and get different formats. The fix is the same as with a human team: write the standard down once. A short message like "whenever anyone asks for the weekly report, use this structure" becomes a persistent instruction, and from then on the report looks the same no matter who requests it.

### "Can it handle a tool you don't support?"

Comes up the moment the AI employee owns a workflow that touches an unusual system. Between built-in integrations, API access, and connecting through automation platforms, the honest answer is that "unsupported" is usually a longer setup, not a wall. We covered the escalation path in [how to connect tools your AI employee doesn't support yet](/blog/how-to-connect-tools-your-ai-employee-doesnt-support-yet).

### "How do we give it bigger work?"

The best question on either list, because it means the small work is already flowing. The answer is scaffolding: bigger work is a chain of small work with checkpoints. A quarterly competitive analysis is a research task, then a synthesis task, then a drafting task, with a human look between each. Teams that try to delegate the quarter's biggest project as one message get mush; teams that stage it get leverage.

## Before and after, side by side

| Before hiring | Three weeks after | What changed |
|---|---|---|
| What can it do? | How do we give it bigger work? | Capability proved on real tasks |
| How is it different from ChatGPT? | How do we make output consistent? | Novelty replaced by standards |
| Who can it see or touch? | How do we share tools with the team? | Fear became a rollout plan |
| What if it gets something wrong? | Why did nobody correct that yet? | Errors became a management routine |
| Do we need training? | Who writes the best briefs here? | Skill became visible and learnable |

The pattern in the right column: every after-question is a question teams also ask about people. That is the tell that the AI employee has become part of the operation rather than a tool being evaluated.

## The question nobody asks, and should

Neither list includes the question that predicts success best: **"who on our team owns this?"** Not owns as in administers; owns as in cares whether it works. Teams where one named person spends the first two weeks delegating real tasks, correcting output, and demonstrating wins in public channels adopt fast. Teams where the AI employee is introduced with an announcement and no owner produce a quiet workspace and a renewal question three months later.

If you take one thing from this post: pick the owner before you pick the tool.

## Frequently Asked Questions

### How long does it take to evaluate an AI employee properly?

One week of real tasks beats a month of demos. Pick three recurring tasks, delegate them with clear briefs, correct the output once, and see whether the corrections stick. By day five you will know more than any feature comparison can tell you.

### What is the best first task to delegate?

A recurring report you already produce manually. It has a known correct output, so quality is easy to judge, and it repeats, so the win compounds immediately.

### Do we need technical people to hire an AI employee?

No. Connecting tools is an authorization flow, like signing into an app with Google. The skill that matters is delegation: describing outcomes, sources, and format clearly. Managers tend to be better at this than engineers.

### Should we trial with one person or the whole team?

Start with one or two owners, expand by workflow. A single person proving value on real work creates the internal demand that makes the team rollout easy.

### What should we NOT delegate early?

Anything with irreversible consequences and no review step, and anything you cannot evaluate because nobody knows what good looks like. Start where output is checkable.

### How do we measure whether it is working?

Count the recurring tasks that moved off a human's plate and stayed there. Hours saved estimates flatter; a list of owned workflows does not.

### What if our team asks who can see our data?

Evaluate access controls hands-on: scope (he works only where invited), per-person tool permissions, and review-first approvals for sensitive actions. The hands-on check tells you more than any brochure.

## Stop interviewing, start delegating

Every question on the before-list has an answer, but none of the answers will move your team. Delegated work will. The teams three weeks ahead of you are not smarter; they just gave the AI employee a real task on day one and started managing instead of evaluating.

[Add Viktor to your workspace and ask him your before-questions directly](https://viktor.com/?utm_source=blog&utm_medium=cta&utm_campaign=what-teams-ask-before-hiring-an-ai-employee)