Our Year with AI: What We Kept, What We Cut

Every year-end retrospective published this month will tell you 2024 was the year of AI, and for once the cliché is roughly true. This was the year the tools stopped being a topic and became a texture: present in the research phase, the motion pipeline, the codebase, the media plan. It was also the year we learned, sometimes expensively, where they do not belong.

So instead of predictions, here is a ledger. One studio, twelve months, actual usage. What we kept, what we cut, and the line we are holding.

What earned a permanent place

Research synthesis. The unglamorous winner of the year. Transcribing, clustering, and first-pass synthesizing of stakeholder interviews and customer calls now happens in hours instead of days, which means our strategists spend their time on the part that matters: deciding what it means. Priya's team estimates the research week got roughly thirty percent deeper at the same price. Nobody is giving that back.

Previsualization and iteration. Nadia wrote about this in August and the conclusion held all year: generative tools collapsed the distance between an idea and a rough. Motion tests, style exploration, layout variations. The waiting died; the deciding remained.

Engineering assistance. Code scaffolding, test coverage, migration drudgery. Alex's team treats the tools as a fast pair of hands with no judgment, which is exactly the right size for them. Velocity is up; the review bar did not move an inch.

The boring middle of production. Resizing, versioning, transcription, first-draft alt text, localization prep. This is where the hours actually went, and nobody mourns them.

What we quietly retired

The failures were instructive. We stopped using generative tools for final brand copywriting; the output is fluent, plausible, and voiceless, and fixing voiceless costs more than writing well the first time. We stopped generating concept territories, after noticing the tools produce the average of what exists, and the average of what exists is precisely what a brand pays us to escape. And we abandoned two vendor platforms whose entire offering turned out to be a thin interface over a model our team already used directly, with a markup attached. The year taught a reliable sorting rule: AI is transformative where the task is convergent, and a mirage where the task is meant to be divergent.

The line we are holding

Some of what we refuse to automate is practical and some of it is principle, and we think it is worth saying both out loud.

• Judgment stays human. Which of the forty options is right, and why, for this brand, this audience, this moment. This is the product. Everything else is logistics.
• Client trust stays hand-built. No generated meeting summaries sent unread, no synthetic sincerity. People can tell, and they remember.
• Provenance stays documented. Every client knows what was generated, what was made, and what was licensed. This went from a courtesy to a contract clause during 2024, and rightly so.
• Juniors still do the reps. The tools tempt you to skip the apprenticeship stage entirely, and a studio that does so is eating its seed corn. Taste is built by making things badly first. We are protecting that runway on purpose, even when it is slower.

What we expect from ourselves in 2025

Not predictions about the technology; the industry has enough of those. Commitments about the practice. We will keep publishing honestly about what works and what does not, because the gap between AI marketing and AI reality stayed wide all year and somebody should stand in it. We will keep pricing outcomes, not hours, so efficiency lands as client value rather than agency anxiety. And we will keep applying the oldest test in the studio to every new tool: does this make the work better, or just faster? Faster is nice. Better is the job.

The tools changed more this year than in the previous five combined. The job did not change at all: find the true thing, say it with craft, care more than is strictly rational. See you in January.