I Let an LLM Write Our Media Plan — Here’s What Happened

A Cautionary Tale in the Age of Generative Hype

By Dr. Elara Voss, PhD in Artificial Intelligence

📊 The Setup

Our team manages a $2.4M annual media budget across digital, TV, radio, and out-of-home placements. For the Q3 campaign, I decided to do something a little bold: I fed our full brief — audience segments, brand voice, KPIs, historical performance data, channel constraints — into a large language model and asked it to produce a complete media plan. Channel mix, budget allocation, timing, creative briefs, even a risk register. I then sat down with our senior planner and we went line-by-line against the LLM output.

This is not a success story. It is a learning story.

What the LLM Got Right

Let me be fair. The LLM produced a structurally coherent plan in roughly four minutes. The channel mix it suggested — 58% digital, 22% TV, 14% radio, 6% OOH — was within a reasonable band of what our historical data would support. The creative briefs were on-voice. The risk register flagged three genuinely useful risks: creative fatigue on the hero spot, seasonal CPM inflation in week 3, and a dependency on a single DMP vendor.

If you stopped reading here, you might think "AI just replaced the planner." You would be writing a much shallower article.

The Numbers That Did Not Add Up

Here is where it got interesting. The LLM allocated 58% of budget to digital, which sounded right. But when I traced the line items:

Individually, small deltas. But our historical flighting data showed that underweighting TV by 6 points in a brand-lift objective campaign costs roughly 40–55 bps of incremental reach, which translates to about $180K in foregone contribution margin at our average LTV. The LLM knew the right shape of the answer. It just couldn't do the arithmetic that connects channel weight to business outcome.

It had read enough media plans to know what one looks like. It had not run the optimization.

The Timing Error That Could Have Cost Us

The LLM recommended a 4-2-1 flighting pattern: 4 weeks heavy, 2 weeks light, 1 week maintenance. Our data says the optimal for this audience is 3-3-2 — a longer initial flight to build frequency of 4+ before the taper. The LLM's 4-week heavy flight pushed us into a frequency saturation zone where marginal cost per additional exposure rises 34% (I ran the regression on 14 months of panel data). The LLM did not know our specific saturation curve. It had seen the shape of flighting patterns in training data and picked the most common one.

Most common is not most optimal. The LLM optimized for plausibility, not for our P&L.

The Creative Brief: On-Voice, But Generic

The creative briefs were the part I was most impressed with. The LLM wrote three 30-second radio scripts and a 15-second digital pre-roll script that our art director called "surprisingly on-brand." The tone was right. The structure was right.

But every brief followed the same skeleton: problem statement → product feature → social proof → CTA. Our top-performing creative over the last two years uses an inverted structure: scene → emotion → product reveal → CTA. The LLM had learned the average creative structure from thousands of briefs and produced the median brief. It was competent. It was not creative.

For a media plan, that distinction matters less. For a creative strategy, it matters a lot.

The Risk Register: Three Good, Two Missing

The three risks it flagged were legitimate. But it missed two that our planner would have caught in the first pass:

1. Regulatory timing: Our product category has a seasonal FDA review cycle that could affect claims in the 2–3 week window. The LLM had no access to our regulatory calendar.

2. Vendor concentration: We have two agencies running 80% of our digital. The LLM had no knowledge of our agency structure. It optimized in a vacuum.

The LLM could only risk-model what was in the prompt. It could not risk-model what was in our internal systems.

What I Would Change Next Time

The LLM saved us about 2 days of drafting time. It did not save us the 2 days of validation, because we could not validate in good faith without doing the validation. The net time savings is modest. The quality risk was real.

The Deeper Lesson: Plausibility vs. Optimality

This is the core insight, and it generalizes well beyond media planning.

Large language models are stochastic parrots of the training distribution. They produce the most plausible next token, the most common structure, the median answer. They are not running your optimization problem. They are not reading your P&L. They are not sitting in your regulatory office.

In media planning, plausibility and optimality are usually close enough that a good planner can spot the gaps. In a high-stakes, high-dimensional problem — a national campaign, a M&A model, a clinical trial design — plausibility and optimality can diverge by millions.

The LLM is a very fast, very well-read junior analyst. It is not your senior planner. It is not your CFO. It is not your creative director.

A Practical Framework for Using LLMs in Planning

1. Use the LLM for structure, not for numbers. Let it draft the skeleton, the channel logic, the creative briefs. Then run your optimization model on top of it.

2. Use the LLM for edge cases, not for core allocation. Ask it to stress-test your plan, to generate alternative scenarios, to write the risk register. Don't ask it to pick the budget split.

3. Use the LLM for documentation, not for decision. The narrative, the board deck, the client-facing summary — the LLM is excellent at this. The actual numbers should come from your models.

4. Always validate the arithmetic. The LLM will produce internally consistent numbers that are externally wrong. Check every allocation against your optimization output.

5. Feed it your constraints explicitly. The LLM will not infer your vendor structure, your regulatory calendar, your agency relationships. Write them into the prompt or expect them to be missing.

The Bottom Line

The LLM did not replace my planner. It did not replace my model. It did not replace my judgment. It did something more useful: it gave me a fast, coherent, structurally sound first draft that I could then interrogate. The value was in the iteration speed, not in the answer quality.

If you are considering letting an LLM write your media plan, your marketing strategy, your financial model, or your clinical protocol: do it. Then spend the time you saved doing the validation that the LLM cannot do. The LLM is a very fast, very well-read junior analyst. Treat it accordingly.

And if your junior analyst has a PhD in artificial intelligence and has published three papers on LLM reliability, you might want to check their work a little more carefully than you would the first-year analyst.

Dr. Elara Voss is a senior planner and AI researcher. She has managed over $40M in annual media spend and has spent the last four years studying where LLMs are reliable and where they are not. She remains cautiously optimistic.