Start from the work, not the technology
The usual mistake is to begin with a tool and then look for something to use it on. That produces a subscription and no change. Begin instead with a list of the jobs that reliably slip: the quotes written at 9pm, the invoices nobody chased, the enquiries answered two days late.
Write ten of them down as they actually happen, in your own words. You are not writing a specification. You are collecting the raw material for the first sentence you will hand over.
Pick the job you would be relieved to lose
From that list, choose one job that is repetitive, low-risk if it is wrong on the first attempt, and already annoying you. Repetitive means you will see the benefit within a week. Low-risk means a bad draft costs you two minutes, not a client. Annoying means you will actually use it.
Resist starting with the most valuable job in the business. Valuable work usually carries judgement, and judgement is the part you should keep.
Describe the outcome, not the steps
When you hand the job over, say what you want to end up with rather than how to produce it. "Draft replies to last night's enquiries with the right prices and availability" is a better instruction than a list of eleven steps, because it leaves room for the work to be broken up sensibly and for you to correct the output rather than the method.
Check the first ten, then check the tenth
For the first handful of runs, read everything. You are calibrating: seeing which parts come back right and which parts need your context added. Once a job is consistently good, move to spot-checking — but keep reading anything that leaves the business.
Only then widen it
When one job is dependable, add a second that touches the same information. Growth by adjacency is much easier to control than starting five unrelated experiments at once.
There is no reward for adopting AI quickly. There is a real reward for having one job permanently off your desk by the end of the month.
A first hour, if you want one
If you have precisely one hour, spend it like this.
- Fifteen minutes listing the jobs that slip.
- Five minutes choosing the dullest, safest one.
- Five minutes writing the request in one plain sentence.
- Twenty minutes reading and correcting what comes back.
- Fifteen minutes deciding whether to run it again tomorrow.
Written by the 8ORA team. We publish no customer names, figures or certifications we cannot evidence.