Six months of AI inside an agency: what the honest retrospectives agree on

The most useful writing about AI in agencies isn't coming from vendors; it's coming from practitioners posting honest retrospectives. A twelve-person agency's six-month review is typical: batch production down from three-to-five days to one-to-two, eight tool subscriptions consolidated — and a candid list of what "blew up in our faces". Read enough of these and the pattern is consistent.

What reliably works

• Production speed. Drafting, variants, adapting one concept across platforms and formats. The days-to-hours claims hold up.
• Consolidation. Teams that pick one workflow shed the stack of single-purpose subscriptions they'd accumulated.
• Research assistance — summarising, clustering, first-pass analysis — where a human uses the output rather than forwarding it.

What reliably blows up

Every blow-up story in these threads happens at the same boundary: AI output that reached a client or an audience without a person owning it first. Voice drift a client noticed before the agency did. An invented statistic in a client deck. Generated visuals that were nearly right in a way that read as carelessness. The failure is never "we used AI"; it is "nobody stood between the model and the client".

The staff question deserves a straight answer

The other thread we'd point at is the employee asking whether their new agency using AI "for EVERYTHING" is normal. The honest answer: using AI for production is normal now; having no defined gate is common but shouldn't be. A healthy shop can say precisely what AI drafts, what humans decide, and who signs off client-facing work. If nobody can answer that, the discomfort is well placed.

That sentence — what AI drafts, what humans decide, who signs — is the whole operating model. It's the one we built Brandaris around, and it's writable into any agency's process this week, whatever tools are underneath.