AI in Marketing: What Really Holds Up Today and What Is Overrated

Short and direct: AI carries real weight in marketing today in three places. In the bid and budget automation of the ad platforms, in the production of drafts for copy, images and video edits, and in the analysis of large data volumes. It is overrated wherever it is meant to exercise judgement: positioning, offer, price, brand, and the question of which number in your dashboard you are allowed to believe. The gain is speed, not replacement.

Every figure here is labelled industry experience, a platform specification, or a calculation you can redo with your own numbers. All amounts given are excluding VAT. Platform rules change; check them before you act on anything. As of October 2026.

Where does AI hold up in marketing and where does it not?

Why does platform AI come before any tool list?

Most articles start with tools. Wrong order: the AI that moves money in your marketing today sits inside the ad platforms. And there is a distinction worth making that often gets blurred. Smart Bidding decides the bid in each auction, nothing more. Campaign types such as Performance Max, Demand Gen or Advantage+ also decide placement and the combination of image and text; with Performance Max, final URL expansion can vary the landing page as well, unless you switch it off.

So the job is not "use AI", it is: feed the automation correctly and fence it in properly. First, the target event. Optimise for "add to cart" and you will get carts. Without an order value, the system cannot tell a 30-euro order from a 300-euro one. And if you optimise for revenue rather than contribution margin, you will reliably scale your low-margin products too.

Second, data quality: server-side tracking, conversions API, deduplication, value transfer, consent. That is a project, not a checkbox. Third, material and limits: enough creatives, a maintained product data feed, plus exclusions for brand terms, placements and search terms. We have written up what that looks like per channel separately for Meta Ads and Google Ads.

The most common mistake comes on top of that: switching the automation on and then touching budgets and targets every couple of days. Meta cites around 50 optimisation events per ad set per week to leave the learning phase; for Google's Target CPA and Target ROAS, the lower bound is 15 to 50 conversions in 30 days. After a change, reckon with 7 to 14 days. Write the rule down once and stick to it: change budgets by at most 20 to 30 per cent and no more often than every seven days, change targets one at a time, and change nothing structural while a test is running.

What does platform AI need from you?

• A conversion event that matches the business goal, rather than the one that is easiest to measure
• Contribution margin passed through as the conversion value; order value is the minimum version
• Server-side measurement plus the platform API, deduplicated via a shared event ID
• Conversions in the order of magnitude above: around 50 per ad set per week on Meta, 15 to 50 per campaign per month for Google's target-based strategies
• Creatives to the campaign type's specification: for Performance Max, at least three short headlines, one long headline, two descriptions, a logo plus one landscape and one square image per asset group; in practice three to five distinct concepts per ad group
• A product data feed with every required field: ID, title, description, link, image link, availability, price, condition and identifiers such as GTIN
• Exclusions, as far as the campaign type allows them: with Performance Max, placement and content exclusions largely apply at account level rather than per campaign

How do you get contribution margin into the platform?

• Record the margin per product or product group
• Pass contribution margin rather than order value as the conversion value, server-side or via the conversions API
• Deduct returns, discounts and cancellations through retroactive value corrections
• Run the target as POAS, that is contribution margin per euro of ad spend

How many creative variants can you actually evaluate?

A generator producing 200 image variants a week solves no problem you actually have. What matters is how many you can pay for until a difference becomes visible. For that you have to keep two quantities apart.

The learning-phase requirement is the roughly 50 conversions per ad set per week from above: the threshold at which the automation starts working, and no evidence at all that one variant is better. Statistical power is something else. For a relative difference of 20 to 30 per cent at 80 per cent power and 95 per cent confidence, you need roughly 175 to 400 conversions per arm. The approximation behind that: conversions per arm ≈ 16 divided by the square of the relative difference, so 16 / 0.3² ≈ 178 and 16 / 0.2² = 400. Check it with any significance calculator.

At 30 euros excluding VAT per conversion, one cleanly evaluated arm therefore costs roughly 5,250 to 12,000 euros excluding VAT. The second assumption is the share of budget you put into testing, and that is a decision: the usual 10 to 20 per cent in steady state, 30 to 40 per cent in a build-up phase.

The conclusion is uncomfortable: below roughly 10,000 euros excluding VAT in monthly budget, a cleanly significant creative test cannot be funded within a month. At that level you decide over longer periods on the basis of directional indications, and you keep the number of concepts running in parallel small. Launch 30 concepts on a 10,000-euro budget excluding VAT and you get 30 times too little data.

AI still helps with creative, though: variants of a confirmed winner, because the underlying idea already has data behind it. More formats out of one shoot. Cheaper localisation.

Where is the time saving on copy and research real?

In our experience, assistance mainly shortens the first draft, while checking and revising stay the same or increase. How big the gain is for you is something only your own before-and-after measurement will show. It is not the factor of ten that gets promised. The example below is made up to show the logic of the calculation, and is not a measurement.

That is about 44 per cent less time. Push the values towards the edges, so a laborious draft and an involved check, and the gain disappears. It arises at the drafting stage and is partly spent again on checking.

What works well: briefings, before anyone writes. Variants for headlines and primary text that you were going to test anyway. Rough drafts for product descriptions and email automations. Enriching product attributes for the data feed, an underrated lever in shopping campaigns.

What works badly is anything that asserts facts about your product: materials, ingredients, deadlines, certificates, competitor comparisons, figures and studies. Models produce plausible sentences, not true ones.

What do large data volumes really help with?

With working through the things nobody enjoys reading: clustering search terms to see where spend is being wasted. Sorting product data, so which products carry contribution margin and which only carry revenue. Searching reviews and support tickets for objections. Two limits come with that. Language models do not calculate reliably: totals, rates and margins belong in a spreadsheet. And every output stays a hypothesis. The check is a test, for example in the conversion rate optimisation of your website, not an assertion in a slide deck.

What is AI good for in customer service?

Triage and suggested replies are solid, autonomous answering is not: enquiries get categorised, standard cases get a draft, a person signs it off. It gets risky where commitments are involved. A model that invents delivery times, withdrawal periods or warranties creates promises your business cannot keep. Two rules help:

• Replies may only draw on internal documents you maintain, not on the model's free-floating knowledge
• There is a route to a human that stays visible at all times, with no detours

When does personalisation start to pay off?

With a few hundred contacts, the lever sits in the subject line, the timing and the offer, not in a model. In our experience, automatic segments in the common email systems start to work from a few thousand active contacts. The more useful criterion, though, is how narrowly you cut: after all filters, a segment should still have a three-digit number of recipients, otherwise you are measuring noise. In e-commerce work, that makes email stubbornly underrated.

What does AI not replace?

• Positioning. Why someone buys from you is a decision, not a calculation.
• Offer and price. No language model repairs an offer nobody wants, it only phrases it more fluently.
• Brand. Recognisability comes from consistency over months.
• Judgement. Which number you believe, which test gets stopped, which hypothesis is next in line.
• Accountability. If a claim turns out to be wrong or an account gets suspended, no model is liable.

What should you check on data protection?

This section contains general notes from practice and is not legal advice. For a specific tool or contract, bring in a lawyer or your data protection officer.

• Legal basis under Article 6 GDPR (DSGVO). Customer lists, email addresses, support histories and order data are personal data, even when they look unremarkable in a spreadsheet.
• A data processing agreement under Article 28 GDPR with the mandatory content set out there: purpose, processing on instruction only, sub-processors, deletion periods, audit rights. Also check whether your inputs are used for training; that should be possible to switch off.
• Place of processing. If data goes to a third country, Articles 44 et seq. GDPR apply: an adequacy decision, standard contractual clauses or another appropriate safeguard.
• Data minimisation. For work on copy, a model needs no real names and no order numbers.
• Transparency and labelling. AI-generated or AI-modified content falls under the transparency requirements in Article 50 of the EU AI Act, applicable since August 2026 under its implementation timeline, on top of the ad platforms' own policies. Which of these apply to your format is a question for a lawyer.
• Consent comes before loading. Storing and reading information on end-user devices is governed by Section 25 TDDDG, the German implementation of the ePrivacy rules. A consent banner that loads services before consent is given is the actual problem, not the AI behind it.
• Document it. Which tool, which data, which purpose, who signs off. Behind that sit Article 5(2) GDPR and the record of processing activities under Article 30 GDPR.

Why are unchecked AI texts dangerous in advertising?

A plausible-sounding but unsupported claim is two problems at once in advertising: a legal one and a platform one. Typical cases: a product feature that does not exist. An efficacy claim for cosmetics or supplements. A "tested" with no test behind it.

On the platform side it starts with disapproved ads; repeated violations lead to restrictions and ultimately suspension of the ad account. A suspended account costs more time and learning history than text automation saves in a year. The minimum rule: every claim about product, price, deadline or effect needs a source in your own documentation, plus sign-off by a human before it goes live.

How do you start without spreading yourself too thin?

Not with a collection of tools. With two or three use cases, each with one person responsible for it and each with a metric.

Four rules separate an introduction from a purchase:

• Record the baseline before you start. "Feels faster" is not a measurement.
• Count the checking time. The honest calculation is production time plus review time.
• One person per use case. Without a named owner, nobody will be using it three weeks later.
• Define a stop criterion. Otherwise habit decides.

As of October 2026, the list prices of the common providers sit roughly in these ranges; please check scope and current pricing with the provider:

The licences are the small line item, the working hours are the big one. If it turns out along the way that the real gap is not production speed but tracking or channel strategy, that is a useful result. It is the same question as when comparing agency costs: what exactly is supposed to get better, and how will you recognise that it has.

How do we handle this at AdSynergy?

For context: we are a performance marketing agency based in Berlin, we work with DTC and e-commerce brands, over 50 of them so far, and we have managed more than 25 million euros of ad budget in total. The work tends to make sense from around 5,000 euros excluding VAT in monthly ad budget. Being based in Berlin means, in practice, that meetings in person are possible in the Berlin area; necessary they are not.

Our order of work: first the target event and data quality, then setting up the platforms' automation and fencing it in, and only after that the production tools. We calculate on contribution margin, not on the prettiest number in the dashboard. Sign-off stays human. We deliberately do not promise results or timelines.

The way in is a free, no-obligation 30-minute call. After that you get an audit of your ad accounts with concrete recommendations within 24 hours on working days. The audit is free and comes with no obligation, even if nothing follows from it. For that we need read access to the ad accounts. If you would rather write, the contact page is there for that. How we approach performance marketing is written up separately.

What should you take away from this article?

First: give the automation the right target event and reliable data before you start thinking about tools.

Second: reckon with 175 to 400 conversions per arm before you believe a creative test. Below roughly 10,000 euros excluding VAT in monthly budget, that means directional indications rather than significance.

Third: the time saving arises in the drafting, not in the sign-off. Positioning, offer, brand and accountability for the numbers all stay with you.

Frequently asked questions