Part of our guide to How to delegate work to AI agents: tasks, autonomy levels and review loops
On this page · 7 sections
The first week with an AI operations assistant should be boring on purpose. You are not trying to prove it can run the business by Friday. You are trying to learn, cheaply, where its judgement matches yours, where your own rules are fuzzier than you thought, and which single task is ready to run with less supervision next week.
That means starting small: one research task, one drafting task, narrow access, and reading everything it produces. The plan below is the one we suggest to owner-operators. It assumes you work with the assistant from your phone in a chat app, and it applies the autonomy levels from our guide to delegating work to AI agents.
Before day one: pick a lane and a no-go list
Choose one area of the business — enquiries, receivables, bookings, prospecting — rather than “admin” in general. Then write two short lists. The first is five recurring tasks in that lane that you do at least weekly. The second is your no-go list for the first fortnight: things the assistant must not touch at all.
A sensible default no-go list: no money out, no messages sent without your approval, no deleting, no contact with anyone tagged as a complaint or dispute, and no changes to pricing. You can relax it later; it is much harder to rebuild confidence after a surprise.
Finally, connect only what the lane needs. If the first tasks are about enquiries, the assistant needs the enquiries inbox and the CRM, not your personal email and the accounting file.
First tasks for an AI assistant, day by day
Day 1: a read-only research task
Give it something useful that changes nothing. Two good options: research (“build a list of 20 commercial cleaners within 15 km of Parramatta with their website, phone and whether they mention strata work”) or study (“read my last 30 sent emails about quotes and summarise how I usually write: greeting, length, sign-off, phrases I use”). Both produce something you can check quickly, and the second teaches the assistant your voice for day two.
I asked Evie, my Fulcrum Cortex agent, for Australian independent hotels targeting Japanese travellers. That same day, I had 29 companies in Excel.
That is a well-shaped first task: clear criteria, a defined output format and easy to spot-check. Individual results vary, and a list like that still needs a human read before anyone contacts the companies on it.
Day 2: draft-only writing
Ask for drafts of real work you would otherwise write tonight: replies to this morning’s enquiries, a follow-up after yesterday’s site visit, a reminder to a client. You copy, edit and send them yourself, or approve them for sending. Every edit you make is information. Tell the assistant why you changed it (“too formal”, “never promise a date before we’ve seen the job”) so the rule gets written down.
Day 3: organising and a first recurring job
Hand over something repetitive and internal: tagging new enquiries in the CRM, filing supplier invoices into the right folder, or turning your voice notes from site visits into tidy job notes. Make one of them recurring — every weekday morning, say — and ask for a short end-of-day message that lists only what needs you.
Day 4: the first gated action
Choose one outward-facing task where the rules are crisp, such as appointment confirmations or polite payment reminders, and run it in act-with-approval mode. The assistant prepares each message, shows you the exact text and the reason, and sends only after your OK. Our article on approval gates shows what a good approval request looks like.
Day 5: review and rewrite the briefs
Spend half an hour looking back. What did you edit most? Which instructions were ambiguous? Did the assistant stop and ask when it should have, or guess? Update each task’s written instructions, using the structure in how to write instructions for AI agents, and decide which one task gets more room next week.
Days 6 and 7: nothing new
Let the recurring jobs run. A quiet weekend tells you whether the assistant behaves sensibly without you prodding it, and whether its end-of-day message is something you actually want to read.
What not to hand over in week one
- A big project. “Redo the website” or “clean up the CRM” has too many hidden decisions. Break it down later.
- Your hardest task. It is usually hard because of judgement, which is what the assistant knows least about yet.
- Anything with money leaving the account. Preparing a payment list is fine; releasing it is not.
- Customer messages without approval. Even if the drafts look perfect on day two.
- Broad access “to save time later”. Add systems as tasks need them.
How to tell trust has been earned
Trust should follow evidence, task by task. The signals below are the ones we look for before moving a task up a level. The thresholds are illustrative; set your own and write them down so you are not deciding on mood.
| Signal | Good sign | Not yet |
|---|---|---|
| Edits to drafts | Most drafts approved with no changes by day five | Still rewriting most of each draft |
| Stopping to ask | Asks when data is missing or rules conflict | Fills gaps with confident guesses |
| Following written rules | Applies a rule you gave on Tuesday every time after | Forgets corrections |
| Reporting | End-of-day message is short and lists only exceptions | Long recaps you skim or ignore |
| Evidence | Shows where facts came from | Asserts things you can’t trace |
What week two usually looks like
If week one went well, week two is about depth rather than breadth. Promote one task (often the recurring internal job) to act-and-report, add a second lane only if the first is calm, and start reading the daily exception report instead of every individual output. If week one went badly, the fix is almost always in the instructions or the access, not the assistant. Go back to the brief.
For larger requests that span several tasks — month-end, a tender, onboarding a new client — the orchestrator splits work across specialists and people. We explain that in multi-agent orchestration explained.
Working from your phone from the first day
Most owner-operators we talk to do this work in the gaps of the day, so it matters that the assistant lives where you already are. Cortex works through Telegram, WhatsApp, Slack or Discord, and accepts voice notes, photos, documents and short videos as instructions. If you are deciding which app to use, our guide to running your business from chat covers the set-up, and the channels page lists what each supports.
Frequently asked questions
What should I ask an AI assistant to do first?
A research or summarising task that changes nothing, followed by drafts of real messages you would otherwise write yourself. Both are quick to check and teach the assistant your rules and voice before it touches anything outward-facing.
How long does it take to trust an AI assistant?
Trust should be earned per task, not per assistant. Many owners move one internal task to act-and-report after a week of clean results, while customer messages stay on approval much longer. Set written thresholds so the decision is based on evidence.
What access should an AI assistant have at the start?
Only what the first tasks need. If you are starting with enquiries, connect the enquiries inbox and the CRM, not your personal email or accounting file. Add systems as you add tasks, and review access when a task is retired.
What if the AI assistant gets things wrong in the first week?
Expect some misses and treat them as information. Most problems trace back to a vague instruction, missing context or access to the wrong data. Rewrite the brief, add a stop condition, and keep that task at draft-only until it settles.
