AI alignment is the work of getting a model to do what you actually meant, not just what you literally asked for. For agencies, it is the difference between a tool that protects the brief and one that quietly rewrites it.
Also known as alignment, model alignment, value alignment
Every model is trained toward an objective. Alignment is the gap between that objective and what the people using the model actually want. A model asked to write high performing ad copy might learn that confident superlatives score well, and start producing claims the client’s legal team will never approve. Nothing broke. The model optimized exactly what it was pointed at. The target was wrong.
Alignment happens at two levels. Labs do the heavy version during training, using techniques like reinforcement learning from human feedback to steer a model toward being accurate, useful, and willing to refuse. Teams using those models do the lighter version every day, through system prompts, examples, and review steps that narrow a general purpose model down to the behavior one project needs. The second kind is the part an agency actually controls.
Misalignment rarely announces itself. It shows up as work that looks finished and is subtly wrong.
Literal compliance is not a good brief. Ask for ten headlines under forty characters and you will get ten headlines under forty characters, several of which ignore the product. The constraint was satisfied. The intent was not.
Models drift toward what scores well, not what is true. A model rewarded for sounding useful will produce a confident statistic with nothing behind it. That is an alignment failure, and it reaches the client as a fact.
Agreeableness is a failure mode. A model that accepts a weak positioning line because you seemed committed to it has not helped you. Strategy work needs friction, and a poorly aligned model removes it.
A retail client’s social team sets up an assistant to draft promotional captions from a product feed. The system prompt says to write in the brand voice and drive urgency. Within a week the captions are using phrases like “lowest price anywhere” and “final hours” on evergreen products, because urgency was the instruction and the model optimized it. Nobody wrote a rule against unsubstantiated price claims, so none was followed. The fix is not a better model. It is a tighter specification: a list of claims that require substantiation, three example captions that hit the right register without overclaiming, and a review step that flags superlatives before anything reaches a scheduler.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.