How to Use AI to Plan a Seasonal Promotion

Most people who plan a seasonal promotion with AI start at the wrong end. They ask for campaign ideas, get twelve, pick one, and only later work out whether the discount leaves any money on the table. Reverse it. Use AI to do the arithmetic first, produce a discount ladder your margin can survive, and only then let it write anything a customer will read. The planning half is where a model earns its keep, because the maths is tedious and you have all the inputs already.

The three numbers you need before you prompt

Gather these first. They take ten minutes and they determine everything else.

• Gross margin per unit or per job, at full price. Not revenue, margin.

• Stock cover or capacity, meaning how many units you can move or how many slots you can fill in the promotion window.

• Last year's baseline for the same window, if you have it. If not, use the four weeks before the promotion as a flat baseline and say so.

Without these, any model will happily suggest 30% off, which for a business running a 35% gross margin means selling five times the volume to stand still.

Step 1: make the model find your margin floor

My gross margin at full price is 42%. Fixed costs for the promotion
window are £600 (staff overtime and print). My baseline is 180 units
over four weeks.

For discount levels of 10%, 15%, 20% and 25%, calculate:
- the new gross margin percentage and cash margin per unit
- the additional unit volume needed to hold total cash margin flat
  versus baseline
- the additional volume needed to cover the £600 fixed cost as well

Show the working as a table. Do not recommend a discount level yet.

Ask for the working. It lets you check one row by hand, which you should, and it makes the model's arithmetic auditable rather than asserted. The instruction not to recommend yet is doing real work: models that jump to a recommendation tend to reverse-engineer the numbers to support it.

The output usually surprises people. At a 42% margin, a 25% discount needs roughly a 145% volume increase just to hold cash margin flat. Seeing that in a table before you commit is the entire point of the exercise.

Step 2: choose the mechanic, not the number

Once you know the floor, the interesting question is what shape the offer takes. These behave very differently even at identical headline cost:

Give the model your three numbers plus this constraint set and ask it to argue for two mechanics and against two, with the volume implication for each. Asking it to argue against is the part that surfaces the objections, in the same spirit as prompting AI to play devil's advocate on a decision you have already half made.

Step 3: build the calendar backwards from the end date

Work backwards from the last day the offer runs, not forwards from today. A model is good at this if you give it the fixed points:

Promotion runs 14 to 27 October. Lead times: printed material needs
7 working days, email list needs 2 days notice for the segment build.
I post to social myself, no lead time.

Produce a working-day calendar from today to 27 October with:
- the last safe date for each asset
- a three-touch email sequence (announce, midpoint, last chance)
  with send dates
- what has to be true on 13 October for this to launch on time

Flag anything that is already late.

The "flag anything already late" line converts a plan into a decision. If print is already impossible, you find out in the first reply rather than the week before.

Once you have the shape, the content calendar itself is a solved problem and planning a content calendar with AI covers the ongoing version of this.

Step 4: only now, the copy

With the mechanic and dates fixed, the writing is the easy part, and it is the part where a model needs the tightest brief: who the offer is for, what it is not, the exact terms, and the one thing you want a reader to do. Feed it your existing best-performing message rather than a description of your tone, because matching brand voice from examples works considerably better than adjectives.

Two guardrails on the copy. Never let a model invent terms, expiry dates or stock levels: give it the exact wording and tell it those strings are fixed. And run the final version past the question of whether every claim in it is one you can evidence, because a promotion is the moment a regulator or a customer is most likely to hold you to it. In the UK, the CAP Code section on promotional marketing sets out what a promotion has to state up front, and availability claims are the usual failure point.

Step 5: measure it in a way that survives next year

Record four numbers on the day the promotion ends: units sold in window, units sold in the equivalent baseline window, actual average discount given, and total cash margin. Write them somewhere durable.

That single row is worth more than the whole campaign next time you run one, because it turns the following year's plan from a guess into a comparison. It also feeds directly into forecasting demand for a small business, which gets sharper every time you give it a real promotional period to learn from.

For where promotions sit among everything else worth automating, AI for small business is the wider map.

FAQ

What discount level should I actually run?

Whatever your margin floor tolerates at a volume increase you believe. The table in step one gives you the numbers, and your own judgement about achievable volume gives you the answer.

Can AI predict how well a promotion will do?

Not reliably from nothing. It can extrapolate from your own past promotions, which is why recording the four closing numbers matters more than any forecasting technique.

Should I discount at all?

Threshold offers and bundles protect headline price better than straight percentage cuts, and for service businesses filling quiet slots, an early-bird window often beats both.

How far ahead should I plan a seasonal promotion?

Work backwards from your longest lead time. If anything is printed, that is usually seven to ten working days, which sets the real start date regardless of when you had the idea.