Automation & AI

Draft, do not decide

The useful line is not between tasks a machine can do and tasks it cannot. It is between drafting and judgement.

Most arguments about what to automate get stuck on capability — what the tool is or is not able to do. That is the wrong axis, because capability keeps moving and the argument has to be had again every six months.

A more durable line: does this step produce a draft, or does it commit the business to something?

Drafting is safe. Committing is not.

Summarising a call, restructuring a supplier quote into your format, pulling next actions out of messy notes, writing the first version of a reply — these produce something a person then reads. If it is wrong, it is caught before it costs anything.

Deciding whether to take a job, what to charge, whether to give a discount, what to promise on a deadline — these commit you. A wrong answer here reaches a customer, and a human at the end of the chain is the only cheap way to catch it.

Keep the person where the risk is

In practice this means the shape of the workflow matters more than the tool in it. The model drafts, a person approves, and the approval step is real rather than ceremonial. The moment approving becomes a reflex click, you have removed the check without noticing — which is the failure mode to watch for once the drafts get good.

One routine at a time

Adoption fails when it is framed as adopting AI. It works when it is framed as automating one repeated routine completely before touching the next. Pick the routine you repeat most often and finish it. Ten shallow automations are worth less than one that is trusted.