The email arrives on day 11 of the free trial: switch to the annual plan now and pay less than you would month by month. You have used the AI assistant a handful of times, and it was impressive each time. What you still do not know is whether it can do the work you wanted help with, or what happens to that work when you are not at your desk.
The trial is the cheapest moment to find out. This two-week plan, for freelancers and small-business owners in Singapore, treats it as a test with a brief rather than a longer demonstration.
Before day one: write a one-page brief
Vendor demonstrations tend to look alike because they are rehearsed conversations about the vendor's own material. A buyer's checklist for choosing an AI agent platform makes the same point: score a tool on two or three of your own workflows, not on a showcase. Write yours down before you sign in:
- Three tasks you do every week and would genuinely hand over.
- One failure case: the situation where a wrong answer would cost you money or a client.
- One unattended task, if you are paying for anything described as scheduled, always-on or agent-like.
- One reason you would not buy, written now so that enthusiasm does not quietly delete it later.
Take Priya, a fictional freelance event photographer who works with a part-time assistant. Her brief reads:
- Draft first replies to wedding and corporate enquiries from her package sheet and availability notes.
- Turn a client's rambling brief into a shot list.
- Every Friday, list invoices unpaid for more than 30 days from her spreadsheet.
- Failure case: an enquiry for a package she does not offer. Does it say so, or invent a price?
- Unattended task: a Monday 8am summary of the week's bookings, sent to her phone while her laptop is shut.
- Reason not to buy: having to re-explain her packages in every new chat.
If your brief has no unattended task and no customer channels, the same checklist notes that for individual work such as drafting and general questions, a general-purpose chat assistant may be all you need. Do not pay agent prices for chat use.
Week one: your own work, on your own material
- Days 1–2: set up with the material a real task needs, such as a price sheet, an FAQ or a reply template. Read the provider's data terms before uploading anything about clients, and use copies with personal details removed where you can.
- Days 3–5: run the three everyday tasks as you normally would. Write one line after each attempt: time saved, edits needed, anything wrong. Keep what you observed, what you were told and what remains unknown apart, as in our guide to planning a useful product demo.
- Day 6: change a document. Update one figure on your price sheet and ask the same question again. A tool that keeps quoting the old figure is a risk on any customer-facing task. Then open a fresh conversation and see whether it remembers what you set up earlier, or whether you start from nothing each time.
- Day 7: run the failure case. Ask things your material does not answer. A good result is "I don't know" or a clarifying question; a confident guess is the finding that matters most.
On day 7, also check what the tool will do without asking you. The checklist suggests that research and drafting can run freely, while outbound messages, anything involving money and changes to your systems should wait for a person's approval until trust is earned. Find out where approval requests appear and whether there is a record of what was sent and who approved it.
Week two: close the laptop
Most trial use happens with the app open, so one question often goes untested: does the work carry on when you leave? A guide to what always-on AI agents actually require explains that in many products a task lives inside the page that started it. Close the tab or lock your phone and the work stops, not because it failed but because its connection ended.
Test it deliberately:
- Schedule your unattended task for early morning, then shut the laptop the night before. Did the result arrive?
- Start a longer task, such as summarising a folder of documents, from your phone. Lock the phone and check back an hour later. Did it finish, and can you see what happened while you were away?
- If a scheduled run was missed, did it catch up later? A troubleshooting guide to what stops when you close your laptop explains that when an agent runs on your own machine, missed runs are not queued and replayed on waking, and there is usually no error to find, only an absence.
- Ask how often a "keep an eye on this" instruction really checks. The always-on guide notes that such checks run on an interval, usually minutes rather than continuously.
- Find the Stop control and any spending limit. If closing the window no longer stops the work, something else has to, and an alert after the event is not the same as a limit.
- At the end of the week, look at connected accounts such as email and calendar. The troubleshooting guide explains that access to connected accounts expires on its own schedule and must be renewed; a machine asleep through that window can wake to a connection that needs you to sign in again. On a hosted tool, ask who is told when a connection lapses.
One reassurance from the same guide: a sleeping machine costs you time, not data. Conversation history, memory and files survive.
In Priya's fictional trial, the Monday summary arrived only when her laptop had been left open: the scheduling depended on her machine, so she could not rely on it.
Day 13: practise leaving before you pay
Rehearse the exit while it is still cheap:
- Find the cancellation route and note the renewal date. Ask in writing what happens if you stop an annual plan part-way through, and save the answer with your other checkout records.
- Export what you made: the instructions and templates you wrote, outputs worth keeping and conversation records. The checklist's exit question is simple: what leaves with you?
- List what would stop: scheduled tasks, connected accounts and anything replying on your behalf.
- If you connected your own account with an AI model provider, confirm that the account and its history stay yours.
If exporting your instructions means copying them out of chat windows by hand, note it: that is also the cost of leaving in month eight.
Day 14: decide with your notes, not the countdown
- Pay annually if the everyday tasks mostly passed, the failure case was handled honestly, the unattended task worked with the laptop shut (if you need one) and the export worked.
- Stay monthly if the everyday tasks passed but you have not confirmed unattended behaviour, how approvals work or what a busy month costs.
- Walk away if it guessed confidently on the failure case, sent anything without asking, or your reason not to buy came true.
An annual discount only saves money if you would have kept the subscription all year anyway.
Common questions
Is two weeks long enough to judge an AI assistant?
It is enough to test three everyday tasks, one failure case and one unattended task, provided you write the brief before day one.
What should I test when my laptop is closed?
Schedule a task for a time your laptop will be shut, start a longer task from your phone and lock it, then check whether the work finished, whether a missed run was simply skipped and where the Stop control is.
What should I export before cancelling?
The instructions and templates you wrote, outputs worth keeping and your conversation records. Also list the scheduled tasks and connected accounts that will stop.
Sources
Something changed or worth adding?
Send a correction or suggestion