You will be able to estimate whether AI help on a task saves time once you count checking and fixing.
Choosing which work tasks to try with an AI assistant is usually done on gut feel. Most people reach for the job that eats the most hours, but the largest job on your list is often not the one where an assistant returns the most time. A rough estimate takes about five minutes per task. It tells you where to begin, and it can save you from spending a month on a task that was never going to pay.
People who say AI saved them time usually mean that a draft showed up fast. A fast draft covers only part of the job, because using an assistant brings three kinds of work you did not have before. Prompting time is the time spent writing the request, pasting in the material it needs and replying to any questions the assistant asks you back. Checking is reading what comes out carefully enough that you would sign your name to it. Fixing is correcting mistakes, filling in anything left out and putting the result into the format your manager expects.
Your time saved on a task is the time it used to take by hand, minus prompting, checking and fixing time. When that figure is positive, the assistant helps with that task. When it is near zero or below zero, it does not help, no matter how good the first draft looked.
Priya (an invented example, as are all of Priya's figures) runs operations for a mid-sized logistics firm in Tai Seng. Following lesson 1.1, Priya listed their ten biggest tasks. The weekly report sat at the top at two hours, followed by email, meeting notes and a monthly reconciliation of supplier invoices against delivery records. Priya's first thought was to aim an assistant at the largest number on the list, the report.
Working through the report, Priya put the hand time at 120 minutes. With an assistant they expected 10 minutes prompting, 40 minutes checking the figures against the source spreadsheets and 20 minutes fixing, a total of 70 minutes. 120 - 70 leaves a saving of 50 minutes a week. The report is worth testing, though the saving is smaller than its two-hour hand time had suggested.
How often a task comes up matters as much as how long it takes, because frequency is where savings come from. A minute saved on a task you do many times a week is multiplied many times over. A task you do twice a year will hardly change your week even if it becomes much faster. For each task, write down how often it occurs each week or month next to how long it takes.
Priya's routine email replies take about 6 minutes each by hand. With an assistant, a reply needs a 1 minute prompt, 1 minute reading and 1 minute fixing, which is 3 minutes and a saving of 3 minutes per email. Priya writes around 40 of these a week, so 3 x 40 = 120 minutes a week, more than the report saves.
Some tasks need a good prompt written and tested before you can rely on it, and that setup is a cost too. Frequency decides whether it pays back. Priya considered giving 30 minutes to writing and testing an email prompt, with a short note describing their tone and the points they always cover. Dividing 30 minutes by the 3 minutes saved per email shows the prompt pays back after 10 emails, about two or three days of Priya's normal work. The same 30 minutes put into a prompt for an annual budget memo might never pay back, since by next year the format, the numbers and possibly the tool may all have changed.
Watch for tasks where checking the assistant's output would mean doing the whole task again. On those, AI help rarely pays back. Detailed financial work often sits in this group, particularly where a mistake is expensive and redoing the work is the only way to find one.
Priya was most keen to hand over the reconciliation, because they find it tedious. Thinking it through, Priya saw that to trust a reconciliation from an assistant they would have to compare every line against the invoices and delivery records, and that comparison is the reconciliation itself. By hand it takes 90 minutes a month. With an assistant it would be 10 minutes prompting plus 85 minutes checking, 95 minutes in all, which is a loss of 5 minutes before any fixing is counted.
You can still get help with this kind of task another way. A formula that matches records automatically (module 5) might help a great deal, whereas asking an assistant to match the records and then going through its work line by line usually does not.
A quick test picks these tasks out early: ask "how would I know if this output were wrong?" If a minute's reading would show you, the task is a good candidate. If you would have to do the whole thing again to find out, it usually is not.
The 2023 field experiment described in lesson 1.1 points the same way. A research team that included Harvard Business School academics worked with consultants at Boston Consulting Group. On tasks that suited the model, people using it did better and worked faster. On a task chosen to sit outside what the model did well, people using it were more likely to get the answer wrong than people working without it.
Priya ended with a short list: email replies first, the weekly report second, and the reconciliation moved to module 5 so it can be handled with a formula instead of an assistant.
Take the tasks you wrote down in lesson 1.1 and, for each one, record:
the time it takes by hand your estimate of prompting, checking and fixing time with an assistant the time saved per use, found by subtracting those three from the hand time that saving multiplied by how often the task comes up each week or month for a task that needs a prompt written and tested first, the setup time compared with the saving per use, to see when it pays back your answer to "how would I know if this output were wrong?"
Then rank what is left into a short list you can defend, and move any task where checking means redoing it to another approach, such as a formula.
Every figure you produce is a guess, a hypothesis, until you have tested it. The baseline in lesson 1.3 and exercises later in the course will confirm or overturn your estimates. A task that looks ideal may turn out to need heavy fixing, and a task that looks risky may turn out to be easy to check.
Your next step is to run the same sum on your own tasks, starting with five of the ten you wrote down in lesson 1.1.
Take five tasks from your list. For each, estimate how long it takes now, how long prompting would take and how long checking would take, then mark the ones that should save time.
Junxiong-WFG Organisation is an authorised representative of AIA Financial Advisers Private Limited (Reg. No. 201715016G).