How to Use AI for Advertising: A Practical Walkthrough

Why AI Is Now Standard in Advertising — and Where Most Teams Still Stall

Two years ago, using AI in advertising was a competitive edge. Today it is simply how the work gets done — and the numbers make that shift unmistakable.

According to a McKinsey survey cited by Criteo, 78% of companies now use AI in at least one business function, and 71% regularly use generative AI. The AI in advertising market is projected to grow from $11.17 billion in 2025 to $36.34 billion by 2030, a 26.7% compound annual growth rate. As Criteo puts it, AI in advertising has shifted from nice-to-have to standard operating procedure.

Yet adoption is not the same as execution. Most marketing teams stall in the gap between knowing they should use AI and actually running a campaign that performs. The friction points are consistent across organizations:

• Unclear goals. Without specific KPIs, AI models have nothing to optimize — the most common first mistake.
• Tool overload. Dozens of copy generators, image tools, and platforms compete for attention, and most teams pick tools before defining the problem.
• No measurement discipline. Pilots launch without control groups or clean baselines, so nobody can tell whether AI actually helped.
• Creative risk. AI-generated ads can backfire on trust, and the backlash is public and immediate.

That last point deserves attention. Nielsen Norman Group research documents how McDonald's Netherlands pulled an AI holiday ad after backlash over "AI slop," while Coca-Cola's AI remakes of its classic holiday spots were widely criticized as "soulless" and "creepy." Notably, these were not low-effort projects — Coca-Cola's 2025 campaign involved roughly 100 staff, and McDonald's ran a seven-week production with up to ten specialists. Even well-resourced teams can misjudge the line between AI-assisted and AI-offensive.

The trust stakes extend beyond a single campaign. BCG research finds that 69% of consumers feel manipulated when brands use AI in advertising without disclosure. In AI-mediated channels, BCG argues, trust is not a soft metric — it is the cost of entry.

The fix is not another tool. It is a repeatable workflow: define goals, choose tools that match the use case, test small with clean measurement, and integrate with guardrails. The surrounding tasks — data feeds, reporting, variant generation — are exactly the kind of repetitive work agent skills can automate, and SkillzBench's free directory and MCP connector are being built to let you find, vet, and connect those skills with confidence. What follows is that walkthrough, step by step.

The Two AIs You Actually Need: Generative for Creative, Predictive for Targeting

Before you open a single ad platform, you need one mental model: modern AI advertising runs on two engines doing two very different jobs. Confuse them and you'll buy tools that solve the wrong half of your problem.

The first engine is generative AI. Feed it prompts, brand guidelines, and product data, and it produces variations of copy, images, and video — the raw material of your campaign. According to Criteo's taxonomy of AI in advertising, generative models handle creative production at a scale no human team can match, churning out dozens of ad variants for testing rather than the two or three a designer might deliver in the same week.

The second engine is predictive AI. It doesn't make anything; it decides things. Drawing on event logs and behavior data, it scores likely outcomes, then chooses which audience sees your ad, how much to bid, and where the budget goes. Targeted advertising — predictive AI's flagship application — already accounts for 31.1% of the AI advertising market, the single largest share of any use case.

Here's the part most beginners miss: these two engines are far more powerful combined than either is alone. As Criteo puts it, AI in advertising works best as a unified system — predictive models pick the moment and the audience, generative models shape the message, and feedback loops keep improving results over time. The pipeline runs on data signals, model training, in-market decisions, measurement, and continuous refinement.

Mapped onto the campaign lifecycle, the two AIs divide the work cleanly:

• Targeted advertising — predictive AI selects who sees the ad and at what price
• Content creation — generative AI drafts the copy, images, and video variations
• Campaign management — predictive models adjust bids and budgets in real time
• Analytics — both feed the feedback loop that sharpens the next cycle

This division of labor also explains why adoption is accelerating. A McKinsey survey cited by Criteo found 78% of companies now use AI in at least one business function, with marketing and sales among the leaders. And the market itself is projected to grow from $11.17B in 2025 to $36.34B by 2030, per The Business Research Company.

Keep one caveat in mind before you touch any tool: AI is the decision-making and content-generation layer that sits on top of your existing data — not a replacement for it. Clean product feeds and consented first-party data are what both engines actually run on. If you plan to automate the surrounding workflow with agent skills, the same logic applies: verify what you deploy, since published research found roughly 26% of agent skills carry at least one vulnerability. SkillzBench is building exactly that verification layer — free security scanning, benchmarking, and version pinning — so you can equip your agents without inheriting their risks.

Hold onto this two-engine model. Every tool, platform, and tactic in the rest of this walkthrough slots into one of these halves.

Step-by-Step: Build Your First AI-Assisted Campaign in Four Moves

Theory only gets you so far — at some point you have to build the campaign. Criteo's four-step framework turns AI advertising from an abstract ambition into a repeatable process, and it starts with a decision most teams skip.

Move 1: Define specific KPIs — and where they'll be measured. "Clear goals tell the models what to optimize," as Criteo's guidance puts it. Decide upfront whether the campaign optimizes for ROAS, CPA, or conversions, and name the platform where each metric lives. Vague goals produce vague models.

Move 2: Match tools to your stage. Don't buy the shiniest platform — buy the one that fits your bottleneck. Per tool comparison research, the mapping looks like this:

• Jasper — governed, on-brand copy at scale with approval workflows
• Copy.ai — rapid A/B testing variants for fast iteration
• Anyword — predictive scoring that estimates engagement "before you spend a dollar on ads"
• Hypotenuse AI — bulk catalog copy via CSV import for large product lines
• Image and video: Midjourney, DALL-E 3, Synthesia, and Runway ML cover visual production at roughly $10–$18/month entry points, per a University of Michigan tool roundup

Move 3: Run a small, cleanly measured pilot. Split a control group from an exposed group, and test a handful of creatives — enough for the system to learn, few enough that it doesn't "get noisy," in Criteo's phrasing. This is where scale discipline pays off: research attributes roughly a 30% reduction in production time to AI-assisted creative development, but only when measurement is clean enough to know what actually worked.

Move 4: Integrate into marketing automation — with guardrails. Pipe in consented first-party data, maintain a clean daily-updated product feed, and connect conversion APIs. Then set hard limits for ROAS, CPA, and frequency so the system optimizes inside boundaries you define, not ones it invents.

One rule governs all four moves: AI drafts, humans approve. Nielsen Norman Group's analysis of high-profile failures — McDonald's Netherlands pulling an AI holiday ad after "AI slop" backlash, Coca-Cola's AI remakes criticized as "soulless" — concludes that generative AI "should therefore be treated as a support tool rather than a replacement of human work." Even Coca-Cola's "AI-made" 2025 campaign involved roughly 100 staff. A human reviews and approves every asset before it ships.

Much of the surrounding workflow — feed updates, reporting, variant generation — is repetitive enough to automate with agent skills. If you go that route, verify before you connect: published research found ~26% of skills carry at least one vulnerability, which is why SkillzBench security-scans, benchmarks, and version-pins every skill in its free directory before you load it via the free MCP connector.

The New AI Ad Channels — and the Trust Guardrails You Need Before You Test Them

A new set of advertising surfaces is forming inside the AI tools your customers already use — and the rules of engagement are still being written. BCG maps this "AI Attention Stack" across three layers: search-embedded AI (Google AI Overviews, Perplexity, Copilot), assistant-native AI (ChatGPT, Gemini, Claude, Meta AI), and retail/commerce AI (Amazon Rufus, Walmart Sparky).

Each layer supports three emerging ad formats: in-answer ads, in-conversation ads, and agentic ads. Some are already live. Google's AI Overviews now show Shopping and Search ads inside AI-generated responses, Amazon introduced ads within Rufus in 2024, and Walmart followed with ads in Sparky, according to BCG's analysis of AI advertising.

OpenAI is moving fastest. Its ChatGPT Ads pilot includes "Sponsored Agents" — users can converse with a business-sponsored agent after clicking an ad — alongside natural-language campaign creation, AI-suggested copy and imagery, and integrations with HubSpot and Shopify. OpenAI stresses that advertisers can review and edit every suggestion before adding it to a campaign.

The budget signal is real: 53% of organizations are already allocating budget to conversational advertising as a distinct media line item, and nearly three-quarters of those plan significant increases over the next two years.

But this channel has a price of admission, and it isn't money. As BCG puts it, "In AI interfaces, trust is not a soft metric. It's the cost of entry." The numbers back that up:

• 69% of consumers feel manipulated when brands use AI for advertising without disclosure
• Nearly 70% name certain data categories — private messages, health information, precise locations — as off-limits for AI use
• Winning visibility requires "model comprehension, not keyword mechanics": clean product data, structured content, and agent-readiness

BCG's guidance for advertisers is to secure early access to alpha and beta programs, run structured experiments across these surfaces, and build governance for disclosure and consent before scaling spend. Preparation, not prediction, is the priority over the next 18 months.

There's a practical takeaway for the surrounding workflow, too. OpenAI notes that tasks like data feeds, reporting, and variant generation "can take significant time" — work that agent skills can automate. But published research found roughly 26% of agent skills carry at least one vulnerability, so it's worth verifying what you connect. SkillzBench is building a free directory where every agent skill is security-scanned, benchmarked, and pinned to a verified version before you attach it via the free MCP tool.

The opportunity is genuine, but so is the trust deficit. Treat disclosure and consent as launch requirements, not polish.

Automating the Boring Parts — and Verifying the Agent Skills You Connect

Here's the part of AI advertising nobody puts in the demo: the supporting work. Data feeds need refreshing, reports need pulling, ad variants need generating, and campaign performance needs analyzing. OpenAI itself notes these tasks "can take significant time" — and they're often what keeps teams from actually running better campaigns.

This is where agent skills earn their keep. A skill is a packaged set of instructions your AI agent follows to handle one of these repetitive jobs — pulling reporting into a summary, generating variant batches from a product feed, or flagging campaigns that drift from your ROAS and CPA guardrails. Instead of manually stitching those steps together each week, you delegate the plumbing and keep the judgment calls.

There's a catch worth taking seriously. Published research found that roughly 26% of agent skills carry at least one vulnerability. A skill that touches your product feed or ad account isn't just a convenience — it's another piece of software with access to your data. Installing one blindly is a bit like hiring a media buyer without checking references.

That's the problem SkillzBench is being built to solve. It's a free, independent directory of agent skills where every skill is security-scanned, benchmarked in a sandbox against the same model with and without the skill, and pinned to a verified version before you connect anything. The planned free MCP tool then lets you toggle exactly which skills each agent can see — so your reporting agent doesn't need access to everything. You'll be able to browse skills, check each one's trust score and measured Lift, and test before committing. The waitlist is open now if you want in when it launches.

A sensible automation sequence for an advertising workflow looks like this:

• Keep a clean, daily-updated product feed feeding your campaigns, as Criteo's implementation framework recommends.
• Automate recurring reporting so performance reviews start with analysis, not spreadsheet assembly.
• Batch-generate creative variants for testing — remember that research attributes roughly a 30% reduction in production time to AI in creative development, per market analysis.
• Set guardrails for ROAS, CPA, and frequency so automated systems have clear limits.

One final caution before you hand over the keys. AI works best as "the decision-making and content generation layer that sits on top of your existing data," as Criteo puts it — not a replacement for it. Automate the boring parts, verify the skills you connect, and keep a human strategist approving what goes out the door.