AI helps most on tasks with high volume, clear rules and clean data: classifying, summarizing, preparing first drafts, spotting patterns. It helps less (or gets in the way) where data is thin, business context matters, or mistakes are hard to undo. That’s the rule I use to decide where to put it.
There’s pressure to “put AI in everything”, which is why it pays to have a criterion before optimizing anything. Here’s mine, with examples you can apply this week.
Where AI really helps
When a task meets three conditions at once, AI usually pays off: it repeats a lot, its output is easy to check, and the cost of a small mistake is low. These are the cases where I notice it most:
- Classifying and tagging: labeling leads by reason for contact, sorting survey answers, tagging tickets or mentions.
- Summarizing and transforming: turning a call or a long report into actionable notes, or adapting the same content to several formats.
- First drafts: article outlines, ad variants or email subject lines to test. The draft is a starting point, not the result.
- Exploratory data analysis: asking a model to scan a table and flag anomalies or odd segments, which I then verify myself.
- Automating intermediate steps: enriching records, normalizing fields or writing a summary that feeds another process.
Where it delivers less than promised
There are three situations where I’m skeptical by default:
- Thin data: an optimization model needs signal. With few conversions per month, the algorithm learns slowly and makes decisions on noise.
- Strategy and positioning: what to say, to whom and why people choose you depends on context that is rarely in the prompt.
- Anything published without review: “I’ll check it later” doesn’t work here. A text with an invented figure costs more than the time it saved.
There’s data on that last point: an NP Digital report based on a survey of 565 US marketers says 47.1% run into AI errors several times a week and 36.5% admit that inaccurate or made-up content went live. The riskiest tasks were generating HTML and schema, developing full content and reporting (summary on Adweek). It’s a survey, not an objective measurement, but it matches what you see in practice.
Ad platform AI: it automates, but it doesn’t give you judgment
Performance Max, AI Max or Advantage+ are AI optimization you already use, even if you don’t call it that. They work best when you feed them good signal (real conversions, not clicks) and badly when the data is poor. That’s why what reaches the platform from your CRM matters: if you optimize for a filled-in form, the algorithm will bring you more forms, not more sales.
It’s also worth watching the controls Google is adding. For Performance Max it has announced first-party audience exclusions, a budget report, full audience reporting and network segmentation in placement reports (Google announcement). In short: more ways to steer and to see what happens inside the black box. My advice is to use those controls before increasing budget.
A quick criterion to decide
Before putting AI into a process I ask myself these questions. If two or more come back as no, I keep it manual or keep a human well inside the loop.
| Question | If yes | If no |
|---|---|---|
| Does it repeat at enough volume? | Good candidate | Probably not worth building |
| Can I check the output quickly? | Automate with sampling | Always human review |
| Is the data clean and accessible? | I can start | Fix the data first |
| Is a mistake cheap and reversible? | More autonomy | A human approves before it goes out |
How to start without overcomplicating it
- List your repetitive tasks for the week and run them through the criterion above.
- Pick just one: the highest volume, lowest risk.
- Define up front what a good result is and how you’ll check it.
- Test it for a few weeks with human review and note the errors that appear.
- Only then expand its autonomy, or drop it if the saving is small.
Frequently asked questions
Which marketing tasks should I automate with AI first?
Repetitive, easy-to-verify, low-risk ones: classification, summaries, tagging and first drafts. Start with one and measure.
Can I publish AI-generated content without reviewing it?
I wouldn’t. Models can invent data or sources with complete confidence, so every figure and claim should be checked before publishing.
Does automatic campaign optimization work with few conversions?
It performs worse, because the algorithm has little signal to learn from. Optimize for a higher-volume event that is a good proxy for real value, and reinforce it with CRM data.
How do you decide where to use AI in your team? Tell me in the comments, or write to me from /en/contact/ if you want to share how you do it.
