If you're not all in on AI

In our holding company, one person now produces drafts that used to need three people. The bottleneck moved to reviewing them.

That is the most useful way I can explain something I posted on X: If you're not all-in on AI today, AI will not be all-in on you tomorrow.

It drew agreement and arguments. Most people read "all in" as buying the subscriptions. A company can pay for every tool on the market and leave the way it works almost untouched.

The drafting example made the distinction practical for us. We had more output, but also more to compare, check and reject. Measuring only how quickly a draft appeared would have hidden the work that followed it. The question became: how much sooner do we get to something we can use?

Count the review

When an agent writes code, we run it. When it writes copy, someone who knows the customer reads it. A plausible answer can reduce the urge to question it. The review needs to survive that feeling of reassurance.

Cheaper drafting has changed distribution too. We can try more variations, publish faster and stop experiments that aren't working. More attempts give us more chances to learn, provided we look at what happened. Otherwise, we've reduced the cost of repeating ourselves.

Change the process

This is where I find the conversation about adoption too narrow. Using a tool once is easy to count. Changing a process raises less comfortable questions. Can the person doing a task change how it's done? Does everything still wait for the owner? Did one disappointing attempt become a reason to stop trying?

A small company has an advantage here. We can change how we work on a Tuesday morning. We can also make a mess quickly if nobody checks whether the change helped.

For us, being all in means learning something about the tools each week, putting useful discoveries into the process and measuring the resulting work. Did it need less correction? Did it get to a user sooner? Did the time saved in production survive the time spent reviewing?

Those answers may be less impressive than a demo. They are what I'd want to know before giving the same task to the tool again.