Key Takeaways
- The honest answer to "what can it do" is a calendar, not a feature list. The work is concrete: Monday's pipeline report, Tuesday's research brief, Wednesday's invoice chasing, Thursday's ad review, Friday's retro board.
- Most of the week is recurring. The highest-value delegations are the tasks that repeat on a schedule, because they compound: set up once, delivered every week without anyone remembering to ask.
- Every outbound action pauses for a human. Emails to customers, changes to ad campaigns, and code pushes arrive as drafts and proposals. The AI does the work; a person stays the decision.
- The week spans departments because tools do. One employee moving across HubSpot, Google Ads, QuickBooks, Greenhouse, and GitHub replaces a stack of single-purpose bots that never talk to each other.
- Off-hours count. Scheduled tasks run overnight and before standup, so mornings start with finished work instead of a to-do list.
The question every team asks in week one
A founder installs an AI employee, stares at the empty message box, and asks the same thing everyone asks: "okay, what do I actually give him?" The feature list does not help; "connects to your tools and does real work" describes everything and nothing. What helps is watching a week happen.
So here is one: a composite Monday to Friday for a 25-person B2B company, drawn from the ways teams actually use Viktor. Names and numbers are illustrative; the shape of the work is the point. If you are still setting up, start with your first 7 days and come back when you want to see the destination.
Monday: the reporting that nobody wrote this morning
7:30 AM, before anyone logs on, the weekly pipeline report posts to #revenue. It was a scheduled task set up a month ago: deals from HubSpot by stage, closed revenue from Stripe, deltas versus last week, and two flagged risks, deals past 30 days in Proposal with no activity. The sales lead reads it with coffee and answers the only question that matters in the thread: "what happened to the Fairmont deal?" Viktor replies with the deal history and the last email exchange, pulled on the spot.
At 9:15, the ops manager asks for something one-off: last month's churn list cross-referenced with support ticket volume, as a table. It lands in the thread twenty minutes later with sources noted per column. Nothing about Monday required opening a dashboard.
Tuesday: research that ends in a document
Sales calls on Tuesday morning surface a prospect objection about a competitor's new feature. Instead of the account executive spending her afternoon in tabs, the request goes to the AI employee: what changed, what it means, what our honest counter is.
@Viktor a prospect says CompetitorX now does automated onboarding flows.
Check their own site, docs, and changelog: what exactly shipped, what
are its stated limits, and where do we genuinely win or lose against it?
One-page brief with links to every claim, draft here before you file
it in the sales Drive folder.The brief comes back with linked claims, a limits section taken from the competitor's own documentation, and a short "where we lose" paragraph, and that honesty is what makes the rest believable. The AE reviews it, softens one line, and the file goes to the shared folder. Total human time: reading it.
Wednesday: the unglamorous middle of the week
Wednesday is collections and hiring, the two categories of work that everyone agrees matter and nobody wants to do.
At 10:00, the finance channel gets the overdue list: invoices 14+ days past due from QuickBooks, sorted by amount, with a drafted reminder email per customer. The drafts wait as approvals; the finance lead approves four, edits the wording on one long-relationship account, and skips one that is in an active dispute. The emails go out only after those clicks. That pause is the entire safety model in one scene: drafting is delegated, sending is not, and the human stays in the loop exactly where it counts.
In the afternoon, the hiring manager preps interviews for a product-designer role. Viktor assembles a packet per candidate from Greenhouse: resume summary, portfolio links, the take-home exercise, and three tailored questions based on gaps the screening call left open. The manager walks into Thursday's interviews prepared, without having spent the evening prepping.
Thursday: the ad account gets an opinion
Thursday morning, the weekly paid review posts to #growth: Google Ads and Meta Ads spend versus plan, cost per lead by campaign, and one flag: a retargeting campaign whose frequency has crept high while conversions sag. Attached is a proposal, not an action: pause the two worst ad sets and shift their budget to the campaign beating its target.
The growth lead looks at the numbers, asks one follow-up in the thread ("what does the last 14 days look like without branded search?"), gets the cut, and approves the pause. The campaign change executes after the approval, and a note lands in the channel confirming exactly what changed. Proposal, question, decision, action, record. It reads like a competent employee's Slack thread, because that is what it is.
Friday: the week closes itself
Friday afternoon, two things happen without prompting. The retro dashboard, a small internal app with a live URL that Viktor built and now updates, refreshes with the week's numbers: shipped items from Linear, revenue movement, support volume. And the week-in-review summary posts to #general: what shipped, what slipped, what is at risk next week, assembled from the channels and tools the team already uses.
The COO skims both in ten minutes. Nobody assembled slides. The recurring machinery behind this kind of Friday, and how to set it up so it keeps running, is covered in how to set up a recurring task for your AI employee.
The week at a glance
| Day | The work | Tools touched | What the human did |
| Monday | Pipeline report + ad-hoc churn table | HubSpot, Stripe, support inbox | Read it, asked one follow-up |
| Tuesday | Competitive brief with linked claims | Web, competitor docs, Google Drive | Reviewed, edited one line |
| Wednesday | Invoice reminders + interview packets | QuickBooks, Gmail, Greenhouse | Approved sends, ran interviews |
| Thursday | Paid review + campaign change proposal | Google Ads, Meta Ads | Questioned, then approved the pause |
| Friday | Retro dashboard + week-in-review | Linear, Stripe, Slack history | Skimmed for ten minutes |
Notice what the human column never says: "wrote the report", "built the table", "drafted the emails". It says read, reviewed, questioned, approved. The judgment stayed human. The production moved.
What did not happen this week
An honest tour includes the boundaries. Nothing was sent to a customer without a person clicking approve. No campaign changed without a proposal being accepted first. The strategy questions, which market to enter, whether to raise prices, who to hire, were never delegated, because they should not be; the AI's job was to make sure those conversations happened with correct numbers in the room.
And at least once, the human said "no". The Wednesday invoice draft for the long-relationship account was rewritten by hand, because the finance lead knew context no tool holds. That is not a failure of the AI employee. That is the division of labor working.
How to get from here to that week
Do not try to install the whole calendar at once. Pick the Monday report, the single most annoying recurring deliverable your team produces by hand, and delegate that first. When it arrives correctly two weeks running, add Wednesday. The path from first task to full week is shorter than most teams expect, and the write-up on which workflows to automate first is a good place to choose your opening move.
Frequently Asked Questions
What tasks can an AI employee actually do?
Concretely: recurring reports from tools like HubSpot, Stripe, and Google Ads; research briefs with linked sources; drafting customer emails and invoice reminders for human approval; interview prep from your ATS; proposing and executing approved ad-campaign changes; and building and updating internal dashboards. The pattern is production of real work products, with humans keeping the decisions.
Does an AI employee work without being asked?
Yes, for scheduled work. Recurring tasks such as a Monday pipeline report or a Friday week-in-review run on their own schedule, often overnight or before standup, without anyone prompting them. One-off requests still happen conversationally in the channel whenever they come up.
Can it send emails or change ad campaigns on its own?
Sensitive outbound actions arrive as drafts and proposals that wait for explicit human approval: an email is drafted but not sent, a campaign pause is proposed but not executed. Admins choose which action types require approval, and the defaults are conservative.
How is this different from a chatbot like ChatGPT?
A chatbot answers when you visit it and forgets your company between sessions. An AI employee lives in your Slack or Teams, holds connections to your actual tools, runs scheduled work unprompted, produces files and dashboards rather than only text, and routes real actions through approvals. The difference is between asking questions and delegating work.
How many people does it take to manage an AI employee?
No dedicated operator is needed. In the week above, five different people delegated directly in their own channels: sales, ops, finance, hiring, and growth. Each request took the form of a normal Slack message, and the corrections they made became persistent behavior rather than repeated instructions.
Where should a team start?
With one recurring deliverable, not a grand rollout. Pick the report your team most resents producing by hand, delegate it with clear definitions, and spot-check it for two weeks. Once it arrives reliably, expand task by task; the full-week picture above is an accumulation of single delegations, not a day-one configuration.
