Last year the question was whether generative tools were a toy. This year the question is which parts of the job they should touch. We ran a deliberate program across the agency. Pick a task, measure the current cost and quality, use the tool for a month, then decide. Here is what stayed.
The nine that stuck
Research synthesis, turning twenty competitor pages into a structured comparison. First draft ad copy variants, where volume matters more than polish because the platform picks the winner anyway. Meta descriptions at scale. Transcription and call summaries. Translating campaign copy between English and Arabic as a first pass. Structured data markup. Turning a rough analytics export into a chart ready summary. Brainstorming hooks for short video. And reformatting one long asset into eight channel specific versions.
The common thread: tasks where you can check the output in seconds and a mistake costs nothing. Our creative team stopped losing afternoons to reformatting. They started spending those hours on ideas, which is the actual return.
The four we kept human
Strategy, because a model will confidently recommend the average answer and the average answer loses. Final published copy in a client voice, because fluent and forgettable is worse than rough and distinctive. Anything with a number attached to a claim, since a plausible fabricated statistic is the single most dangerous output these tools produce. And client conversations, obviously.
Writing is the clearest example. We tested the main writing tools on real client briefs and wrote up where they broke. The failures were rarely typos. They were bland claims, invented specifics and a voice that could belong to any brand.
The rules we wrote down for AI in the marketing workflow
Never paste client data into a consumer tool. A human sources every factual claim before it ships. Someone who could have written the piece themselves edits every published word. And we tell clients where we use these tools, because finding out later is a trust problem nobody needs.
A common worry is search. Will Google punish AI assisted pages? Its guidance asks whether the content helps people, not how someone produced it. Its helpful content guidance rewards original, useful pages. Mass producing thin pages to manipulate rankings is the real risk. Our rules keep a human editor on every page, which also keeps the work useful. If you want the search side of that argument, read what our SEO team does with content.
How to run the same trial in your team
Keep it small. Choose one repeated task, such as meta descriptions or call summaries. Time how long a person takes today and ask a senior colleague to rate the result. Then use the tool for a month on that task alone. At the end, compare the hours saved against the extra editing time, and compare the quality ratings. If the numbers do not clearly improve, drop the tool. Do not widen the trial just because the demo looked impressive. Write down what you decided and why, so the next review starts from facts.
The productivity gain across the agency landed around 20 to 25 percent on production tasks and roughly zero on thinking tasks. That ratio is the whole lesson. Speed on the mechanical parts buys time for the part that actually differentiates the work, which is judgment about what to do and why. If you want to see what that judgment produces on real accounts, start with our success stories, or see how we plan content in social media strategy and management.
