For the last few years a large part of my job at D&G has been putting AI into how we produce work. Almost all of it is internal. Very little of it is anything a client would ever see.
The approach has been deliberately narrow. Rather than adopting AI broadly and hoping something sticks, I've looked for the specific places where time gets lost, built something for that, and then checked whether it actually improved the work. Most of what I've built is unglamorous. One system automates a complicated production workflow for a major automotive client, and it exists because that workflow was quietly costing us weeks a year.
The hard part has not been the technology. It's getting people to use the thing. Plenty of internal tools work fine and still sit untouched because they were built for a problem nobody actually had, or they arrived without anyone explaining why. Most of what I've learned has been about that, not about models.