AI in ad campaigns: what to automate and which decisions to keep under control

Opening a campaign on Google Ads or Meta today means finding automation already switched on: automated bidding, expanded audiences, placements chosen by the algorithm, copy and images the platform offers to generate. The question is no longer whether to use AI in your campaigns. It’s which levers to hand over and which to keep.

The short answer is this: it makes sense to delegate to AI the decisions that require real-time calculations across thousands of signals, that is, bids in each auction, budget distribution across placements and choosing which combination of ad to show each person. What stays under human control is the goals, the overall budget, the messages and commercial promises, the exclusions and checking lead quality. The result depends above all on what you teach the algorithm: if it optimises for completed forms, it will find people who complete forms; if it optimises for real opportunities from the CRM, it will look for customers.

This article is about campaigns. If you want to understand how AI helps find and fix bottlenecks after the click, on the website, forms and sales follow-up, read Alessandro’s article on how to optimise your funnel with AI.

What do Google, Meta and LinkedIn already automate?

Much more than people think, and in different ways on each platform. This is the situation in October 2026, with the status of each feature as declared by the platforms.

Platform	Feature	What it automates	Status
Google Ads	Smart Bidding	The bid in every auction, based on device, location, time, lists, language, query and other signals	Available
Google Ads	AI Max for Search campaigns	Matching searches beyond your keywords, text customisation, final URL expansion	Available
Google Ads	Performance Max	Distribution across all Google channels with a single campaign	Available, with more reporting than in 2024
Google Ads	Asset Studio	Generating images and videos for ads	Available; the Google Marketing Live 2026 updates are rolling out in English
Meta	Advantage+ campaigns	Budget, audience and placements together	Available; Advantage+ status requires all three levers to be on
Meta	Advantage+ creative	Generated backgrounds, format adaptation, image animation	Available; some enhancements on by default
LinkedIn	Accelerate	Building the campaign from the objective	Available; check availability by country in your account
LinkedIn	Predictive audiences	Audiences built from leads, lists or conversions, from 300 members	Available

On Google, we’ve analysed how Search, AI Max and Performance Max combine for lead generation in PMax, AI Max or Search for B2B leads.

How do you tell available features from announcements and experiments?

Every year the platforms’ events announce dozens of new features, and the gap between announcement and availability in your account can be months. At Google Marketing Live 2026, on 20 May, Google presented new formats for AI answers. Reading the details, however, conversational ads and highlighted answers in AI Mode are being tested, other formats will arrive “in the coming months” and direct offers are a pilot launched in January. The Asset Studio updates are also rolling out globally, starting with English.

For Meta, the idea has been circulating for months that by the end of 2026 campaigns will build themselves from an objective and a budget. The source is a Wall Street Journal article from June 2025 based on anonymous sources, picked up by the trade press, not an official announcement. What Meta has announced is Muse Image coming to Advantage+ creative, with Muse Video “in development”.

Three questions before building a strategy around a new feature:

1. Where is it written? In the platform’s help documentation or only in an event blog post? Help pages describe what exists; event posts also describe what’s coming.
2. What’s its status? Available, beta, pilot, test, “in the coming months”. These are different things.
3. Is it in your account? Language, country and campaign type often limit the rollout. Italian almost always arrives after English.

What should you automate and what shouldn’t you?

The matrix below is how we think about every lever in a campaign. For each activity it shows how much to delegate, what data the automation needs to work well, what a person checks and which indicator to watch.

Activity	Automation	Data required	Human control	KPIs to watch
Auction bids	High: Smart Bidding, Advantage+ bidding	Reliable primary conversions, minimum volumes, values	Choosing the objective and the target (tCPA, tROAS)	Cost per qualified lead, pipeline value
Distribution across placements	High	Enough history, creative for every format	Inventory and sensitive content exclusions	Performance by channel, read with caution
Audience	Medium-high: expanded audiences, suggestions	Customer lists, CRM, conversions	Exclusions (customers, suppliers, competitors), geographic area	Share of leads in the target
Search matching	Medium: AI Max, broad match	Clear keywords and website pages	Negative keywords, brand exclusions, excluded URLs	Search terms report
Creative variants	Medium: generation, adaptation, combinations	Brand kit, approved assets, guidelines	Approval of every piece of copy and image, promises and claims	Performance by concept, not by single asset
Overall budget	Low	Margins, customer value, commercial goals	Monthly cap, split across channels	Cost per customer, MER
Message and offer	Low	Positioning, proof, prices	Entirely human	Qualification rate
Lead quality	Low for the analysis, high for the data flow	CRM statuses, rejection reasons	Regular review with the sales team	Qualified leads out of total leads

The logic is simple. The more repetitive, fast and based on measurable signals an activity is, the more it makes sense to delegate it. The more it touches promises made to customers, the company’s money and the brand’s reputation, the more it should stay with a person.

Which signals improve optimisation?

Automation optimises towards whatever you tell it success looks like. Four inputs make the difference.

1. Conversions that correspond to real value. In Google Ads only primary conversion actions feed into bidding; secondary ones are for reporting only. If your primary conversions include visiting the contact page or clicking the phone number, the algorithm will look for people who do those things. For leads, Google recommends counting only one conversion per click (conversion counting).

2. Sufficient volumes. Automated strategies need data. For target ROAS, Google gives minimum thresholds ranging from 15 conversions in 30 days for Search to 50 in 35 days for Demand Gen, and to evaluate bidding performance it suggests waiting for at least 30 conversions, 50 with target ROAS. Below these thresholds it’s better to optimise for a more frequent but still meaningful event, or to consolidate campaigns.

3. Feedback from the CRM. In B2B the best signal is the one that comes later: qualified lead, opportunity, contract. On Google it’s sent with enhanced conversions for leads; on Meta with the Conversions API for CRM, which for the Conversion leads objective requires at least 200 leads a month from instant forms and uploads at least daily. LinkedIn offers a qualified leads objective once the Conversions API is set up. We’ve explained how to build the connection in from lead to contract, and how to stop Google working in the dark when consent is involved in enhanced conversions and Consent Mode.

4. Audience as a suggestion, not a fence. With Advantage+ leads, Meta asks you to enter your audience as a suggestion: narrowing the settings too much switches the automation off. Human work shifts to exclusions, that is, who you don’t want to reach (existing customers, employees, suppliers), and to the quality of the lists you upload.

How do you build and evaluate creative variants?

Variants matter more than before. Meta explains that Andromeda, the system that selects candidate ads for each person, works better with greater ad variety, and that businesses are uploading more and more creative “to support a diversification strategy”. However, Meta doesn’t publish official figures on how many variants you need: the “rules” doing the rounds are third-party interpretations.

Variety doesn’t mean a hundred copies of the same image with a different headline. It’s worth thinking on two levels:

• concepts: different ideas that respond to different motivations (saving time, a customer’s testimonial, comparison with the alternative, a product demonstration);
• executions: formats, crops, the opening of the video, primary text and headline for the same concept.

Generative AI is most useful for executions: adapting an image to different formats, generating backgrounds, proposing text variants to choose from. For concepts you need people who know the client and the market. How to balance variety with brand recognition is explained by Marika in a recognisable brand in the age of AI, and the practical rules for formats and templates are in our guide to the brand kit for campaigns.

Evaluating them takes two precautions. First: in Performance Max asset reports, a conversion is attributed to every asset in the ad, so the sum across assets exceeds the total and comparisons should be made between assets of the same type. Second: the most useful comparison is between concepts, not individual elements. If a variant wins on Meta but brings in leads the sales team rejects, it has lost.

What checks do you need on budget, messages and lead quality?

Automation works well within clear boundaries. These are the checks to set up before switching on the most open-ended features.

Budget

The monthly cap and the split across channels are business decisions, not platform decisions. A single campaign’s budget can be left to the algorithm; how much to invest in Meta compared with Google or LinkedIn can’t. You’ll find how to calculate it from your company’s figures in how much to invest in Meta Ads.

Searches and inventory

With the expanded matching of AI Max and Performance Max, exclusions are your steering wheel. In Performance Max you can add up to 10,000 negative keywords per campaign, which apply to Search and Shopping, while for Display and Video you use content exclusions. Brand exclusions stop you paying for searches on your own brand or appearing on brands you don’t want. The search terms report should be read every week in the early stages.

Messages and images

Google explicitly asks you to review generated or suggested assets to make sure they’re accurate and not misleading, and Meta recommends checking and approving the preview of every ad before publishing, noting that enhancements on by default can be switched off. In practice:

• switch off enhancements that alter products, people or places in ways that aren’t acceptable for the brand;
• set the brand guidelines in Performance Max (name, logos, colours, fonts), knowing that Google doesn’t guarantee they’ll be applied in every format;
• keep a list of banned claims: unconfirmed prices, guarantees, results figures, comparisons with competitors;
• with final URL expansion, exclude pages that shouldn’t receive paid traffic (careers, members’ area, legal pages).

From 2 August 2026 the transparency obligations of Article 50 of the EU AI Act also apply to content generated or manipulated with AI, starting with deepfakes. Google offers an AI labelling setting and states that it doesn’t guarantee compliance; Meta applies labels to ads created or significantly edited with its generative AI tools (Meta). For images with realistic people or places, the decision to disclose should be made before publishing, not after a complaint.

Lead quality

This is the most important check and the most neglected. Once a week, in the first weeks of an automated campaign, it’s worth opening the CRM and looking at the contacts that have come in: how many are in the target, how many the sales team has qualified, which rejection reasons keep coming up. If the share of off-target leads rises while the cost per lead falls, the algorithm is finding the easiest route, not the right one. We’ve explained why platform ROAS and cost per lead can be misleading in why ROAS can fool you.

Which decisions stay human?

Even with the most advanced automation, some decisions stay with the people accountable for the results:

• the goal: what counts as success for the business, and therefore which conversion to teach the algorithm;
• the value: what a qualified lead, an opportunity or a customer is worth, and therefore the cost and return targets;
• the message: what we promise, with what proof, in what tone;
• the boundaries: who we don’t want to reach, where we don’t want to appear, which pages not to push;
• judging the results: whether an improvement is real or an effect of attribution.

What changes is how time is spent. Fewer hours on manual bidding and audience combinations, more hours on signals, creative, lead analysis and the link with sales.

Mistakes to avoid

Switching everything on at once. If you activate AI Max, URL expansion and creative enhancements on the same day, you won’t know which one caused the change. One lever at a time, with an experiment where possible.

Optimising for the completed form. It’s the easiest signal to hit, and the algorithm will hit it. Without feedback from the CRM you lower your cost per lead and worsen quality.

Narrowing the audience out of fear. With automated campaigns, an audience that’s too narrow blocks learning. Clear exclusions and good conversion signals are better.

Mistaking an announcement for an available feature. Before changing strategy for something presented at an event, check its status and availability in your account.

Letting AI write the promises. Copy variants yes, commercial claims no.

Judging creative on the single asset. In asset reports, conversions overlap. Compare concepts and look at the quality of the leads they bring in.

Checklist before extending automation

• Check which conversions are primary in Google Ads and on Meta, and keep only those that correspond to value.
• Make sure you have the minimum volumes for the bidding strategy you want to use.
• Connect the CRM with enhanced conversions for leads and the Meta and LinkedIn Conversions APIs.
• Upload exclusion lists: customers, employees, suppliers, competitors where possible.
• Set negative keywords, brand exclusions and excluded URLs.
• Configure brand guidelines and switch off unsuitable creative enhancements.
• Write down the list of banned claims and who approves creative.
• Agree with the sales team how to flag off-target leads and check the CRM every week.
• For each new feature, note the source, the status and the date it was switched on.

Want to know how much to delegate in your campaigns?

We’ll audit your Google Ads, Meta and LinkedIn accounts: conversions and signals, bidding strategies, exclusions, brand guidelines, active creative enhancements and the connection with the CRM. We’ll tell you which automations are working for you, which are spending in the wrong direction and what to fix before extending them.

You can book a call with Federico or write to us using the contact form. You’ll find how we manage campaigns on our Google Ads, Meta Ads and LinkedIn Ads pages.

Frequently asked questions

Can AI run ad campaigns on its own?

It can manage bids, budget distribution across placements, audiences and creative combinations. It can’t decide what counts as success for the business, what a customer is worth, what to promise in the messaging or whether the leads are the right ones: these choices stay with the people running the campaigns together with the client.

How many conversions does Smart Bidding need to work?

It depends on the strategy and the campaign type. For target ROAS, Google gives minimums ranging from 15 conversions in 30 days on Search to 50 in 35 days on Demand Gen, and suggests evaluating performance after at least 30 conversions, 50 with target ROAS.

Can Meta’s creative enhancements be switched off?

Yes. Meta states that some Advantage+ creative enhancements may be on by default and that they can be switched off at any time. It’s still worth checking the preview of every ad before publishing.

Do you have to disclose ad images created with AI?

From 2 August 2026, Article 50 of the EU AI Act imposes transparency obligations on content generated or manipulated with AI, particularly deepfakes. Google and Meta offer or apply AI labels, but Google states that its setting doesn’t guarantee compliance. For borderline cases you need legal advice.

How can you tell if automation is bringing in worse leads?

By comparing in the CRM the share of qualified leads before and after activation, over the same period and with the same budget. If the cost per lead falls but the qualification rate falls further, the automation is optimising for the wrong signal: you need to connect the CRM feedback.