How AI Agents Are Changing Digital Marketing (and How to Actually Use Them)

Most of what gets sold as an “AI agent” in marketing today is a chatbot with a new name. A smaller number are the real thing, and those are quietly changing how campaigns get planned, run and measured.

I work at a digital commerce agency, and one conversation keeps coming up across this industry. Someone sees a demo where an AI “runs the campaign by itself.” They want it. Nobody has defined what “it” is, what it can touch, or how they will know it is working. This guide is the version of that conversation I wish every team had before buying anything. I’ll cover what agents are, how they differ from the tools you already use, where they help in digital marketing, how to build and launch your first one, how to measure it, and where it can go wrong.

What is an AI agent?

An AI agent is software that takes a goal, works out the steps, uses tools to carry them out, and checks whether it worked. You hand it an outcome, for example “find the ten product pages losing the most organic traffic and tell me why.” It decides how to get there.

Four things separate an agent from the AI assistant most teams already use:

• It works toward a goal, not a single reply.
• It uses tools. It can call Search Console, an ads platform, your CRM, a spreadsheet or a browser.
• It has memory. It keeps context across steps, and sometimes across days.
• It closes the loop. It looks at what happened and adjusts.

Almost every agent runs some version of the same loop.

Here is what that looks like in practice. Say you ask an agent why organic revenue dropped last week. It pulls Search Console and analytics data, notices that a template change added a noindex tag to a few hundred collection pages, writes a ticket for the dev team with the affected URLs, and checks indexing again three days later. A chatbot would have explained what a noindex tag is. That is the gap between answering and doing.

Two more points about the loop. First, the human checkpoint matters. In marketing, one wrong action can mean a paused campaign, a wrong price in an email or a brand mistake in public, so the best agents are designed to stop and ask at the right moments. Second, the quality of the loop depends on the quality of the data the agent can see. A smart agent with bad data just makes bad decisions faster.

Agents vs chatbots vs automation vs multi-agent systems

These four get mixed up constantly, and vendors are happy to let them. The simple way to tell them apart is to ask two questions: how much can it decide on its own, and how much can it do in other systems?

A multi-agent system works like a small team. One agent researches, another drafts, a third checks the draft against your brand rules, a fourth publishes, and an orchestrator keeps them in order. The benefit is the same as in a human team: each agent has one narrow job and does it better. The cost is that mistakes can travel from one agent to the next, so you need checks between the hand-offs.

Two quick tests help when a vendor pitches you.

• If it cannot take an action in another system, it is a chatbot.
• If its path never changes whatever it finds, it is automation.

Gartner has a name for the gap between pitch and product: “agent washing,” meaning existing chatbots, assistants and automation tools rebranded as agents. It estimates that only about 130 of the thousands of vendors claiming agentic AI offer genuinely agentic features.

The agentic framework: what sits around the model

The model is only one part of an agent. The “agentic framework” is everything around it that turns a language model into something that can do a job reliably: the instructions, memory, tools, coordination and safety checks. When someone says they are “building an agent,” most of the real work is in this scaffolding.

Here is what each part does.

• The model. The reasoning engine. It reads the situation, plans and decides what to do next.
• Instructions and goals. What the agent is for, what good looks like, and your brand voice, pricing rules and no-go areas.
• Memory and state. Short-term memory holds the current task. Long-term memory holds things like your product catalogue, past campaign results and customer segments.
• Tools and integrations. The agent’s hands: Search Console, Google Ads, Meta, your CRM, your email platform, your store’s API. Standards such as the Model Context Protocol (MCP) are making these connections easier to build.
• Orchestration. The logic that sequences steps, retries failures, and routes work between agents when there are several.
• Guardrails and approvals. Limits on what the agent may spend, send or change, and the points where a human has to sign off.
• Monitoring and logs. A record of what the agent did and why, so you can audit it and fix it.

Choosing a framework

You have three broad routes. Code-first frameworks such as LangGraph or CrewAI give developers full control over multi-step and multi-agent flows. Workflow platforms such as n8n, Make or Zapier let a marketing team wire agents to its tools with little or no code. Platform-native agents, built into your CRM, ad platform or commerce platform, are the quickest to start and the least flexible.

For most marketing teams the right first move is the middle route or a platform-native agent. Move to code when you hit a limit you can name. When you compare options, check five things: how well it connects to the tools you already pay for, how much of its reasoning you can see in the logs, how it handles approvals, where your data is stored, and what it costs at ten times today’s volume.

Why this matters now

Three things changed at the same time, which is why agents went from demo to real work so quickly.

• Models got better at multi-step work. They can now plan, call tools and recover from errors well enough to be useful on bounded tasks.
• Connecting tools got easier. Common standards and ready-made integrations mean an agent can reach your analytics, ads and store data without a custom build every time.
• Customers started sending agents. People now ask AI assistants to research, compare and shortlist products for them. That changes how they find you.

On the third point there is hard data. Adobe Analytics, which tracks over a trillion visits to US retail sites, reported that traffic from AI sources grew 393% year on year over a single quarter. In the last month of that quarter, that traffic converted 42% better than non-AI traffic and earned 37% more revenue per visit. A year before, it had converted 38% worse. The base is still small compared to search and social, so do not read this as AI replacing your main channels. Read it as a channel that went from poor to better than average in about twelve months.

The warning label comes from Gartner. It expects more than 40% of agentic AI projects to be cancelled, because of rising costs, unclear business value and inadequate risk controls. The technology is not the main reason. The usual cause is a team that automated a vague goal, never measured it, and ran out of patience. Both things are true at once. The upside is real, and so is the failure rate. The rest of this guide is about staying on the right side of that.

Where AI agents work in digital marketing

The best use cases share three traits: the task repeats, the inputs are digital, and a wrong answer is cheap to catch. Here is where I would look first.

A few notes from working across these areas.

• Start with reporting and audits. They are read-only. An agent that can look but not touch has almost no way to cause damage, and it saves hours every week. It is also the quickest way to find out whether your data is clean enough for anything more ambitious — the same discipline behind a measurement stack you can trust.
• Be careful with anything customer-facing. Content, email and social can go wrong in public. Keep a person in the approval step until the agent has a long clean record. If you want to see what a governed, agent-assisted pipeline looks like, I wrote up the content automation system I built for eCommerce.
• Paid media needs hard limits. If an agent can move budget, give it a daily cap and a rule such as “never raise a campaign’s spend by more than 20% in a day without approval.” Limits written in the setup beat good intentions.
• Pick one metric per agent. An agent that tries to improve everything improves nothing. One agent, one job, one number.

On the first row of that table, “whether the brand shows up in AI answers” is becoming its own discipline. It sits close to SEO and generative-engine optimisation for AI search, and it is where a lot of the next few years of visibility will be won or lost.

Agentic commerce: when the shopper sends an agent

Everything so far is about agents you run. The other half of the story is agents your customers run. A shopper tells an assistant “find me a waterproof jacket under 400 dirhams that can arrive by Friday,” and the assistant searches, compares, shortlists and in some cases completes the purchase. This is what people mean by agentic commerce.

The journey changes shape. In the old journey the customer does the searching and comparing, and you compete for their attention at every step. In the new one, the agent does the comparing, and you compete to be one of the options it trusts enough to put in front of the customer.

The agent is picky in ways a human is not. It cannot be charmed by a banner. It reads your data. Adobe warned that retailers are not ready for this: about a quarter of homepage and category-page content is not optimised for AI models, and around 34% of product pages cannot be properly accessed by AI. It is the same shift in behaviour I unpacked in this guide to how shoppers actually behave — only now part of the shopping is being delegated to software.

Standards are forming around this. Two that keep coming up are the Agentic Commerce Protocol from OpenAI and Stripe, and the Universal Commerce Protocol from Google and Shopify, which aim to give AI agents a common way to read catalogues, check out and handle orders. They are young and may change, so I would not rebuild a store around one. Do the groundwork that helps with all of them.

A practical checklist for your store

• Clean product data. Accurate titles, attributes, sizes, colours, materials, GTINs and variants. Agents compare on specifics.
• Structured data. Product, Offer, Review and shipping and returns markup on every product page.
• Live price and stock. An agent that recommends an item that is out of stock will stop recommending you.
• Clear policies. Delivery times, returns and warranty in plain text, easy to find and easy to read by a machine.
• Crawler access. Check that your robots.txt and security tools are not blocking legitimate AI crawlers by accident. Block the bad ones, not all of them.
• Content that answers comparison questions. Size guides, “X vs Y” pages and honest buying guides give agents something to quote.
• Test it yourself. Ask the major AI assistants the questions your customers would ask. See whether you appear, what they say about you, and what they get wrong.

If you run a large catalogue, especially one with many variants and several markets, this is mostly a data quality job. It is not glamorous, but it is the part that decides whether an agent can use your store at all — and it is exactly the kind of work I take on as part of eCommerce strategy and growth.

A 9-step process to build your first agent

This is the order I would follow. Most failed projects skip steps 1 to 4 and jump straight to picking a tool.

1. Pick one job and one metric. Choose a task that repeats weekly, has digital inputs, and where a mistake is cheap. Weekly SEO reporting is a good one. “Improve marketing” is not. Write down the metric that will tell you it worked, such as hours saved or issues caught.
2. Map the manual process. Write out every step a person takes today, including the small judgement calls. If you cannot describe the job to a new hire, you cannot describe it to an agent.
3. Check the data. List where each input lives and whether it is accurate, complete and accessible by API. Fix the biggest gaps first. This step quietly kills more projects than any model limit.
4. Do it by hand first. Run the whole job manually with an AI assistant, step by step, and check each output. You learn where it goes wrong while the cost of a mistake is zero.
5. Chain the steps. Join the tested steps into a sequence, and decide which ones branch. For example, if the anomaly is larger than a set size, escalate to a person; otherwise log it.
6. Choose the framework and tools. Now, and not before, pick the platform that fits the job, using the five checks from the framework section above.
7. Add context and rules. Give the agent your brand guide, product data, past examples of good work, and the limits it must never cross. Add memory so it does not start from zero each run.
8. Add approvals and logging. Decide which actions need a human sign-off. Record every action the agent takes and the reason it gave, so you can review and correct it.
9. Pilot, measure, then widen. Run it on a small slice for a few weeks against your baseline. Only expand its scope or autonomy when the numbers justify it.

Steps 1 to 4 are mostly thinking and spreadsheets, not engineering. They feel slow, and they are what keep you out of Gartner’s 40%.

How to launch agents inside your organisation

Building the agent is the easy half. Getting a team to trust it, use it and keep it running is the harder half. Two ideas make it manageable: give the agent more freedom in stages, and run the rollout on a fixed 90-day clock.

Earn autonomy in four stages

No agent should start with full control. The ladder below is the one I would use. The agent moves up a rung only after it has shown it deserves to.

• Stage 1, Suggest. The agent recommends, and a person does the work. Use it to see whether its judgement matches yours.
• Stage 2, Draft. The agent prepares the work, such as a report, a brief or an ad set. A person reviews and publishes.
• Stage 3, Act with approval. The agent carries out the action once a person approves it, inside set limits.
• Stage 4, Act alone. The agent acts within hard limits and a person reviews the logs. Keep this for low-risk, reversible work.

Decide the promotion rule before you start. For example: “Move up when a reviewer accepts at least nine in ten outputs without changes, over four weeks.” The numbers are yours to set. What matters is that you set them in advance, so enthusiasm does not set them for you.

A 90-day rollout

Who needs to be involved: a business owner who cares about the metric and can say whether the output is good; a technical owner who manages the integrations, access and logs; a reviewer who approves the agent’s work during the early stages; and legal or compliance, brought in early if the agent touches customer data, pricing claims or regulated content.

Tell the team what the agent is for. The message that works is “this takes the repetitive part so you can spend time on the work that needs your judgement.” If people think it is there to replace them, they will quietly route around it, and you will be left with a project that looks good in a slide and does nothing in practice.

How to measure whether an agent is worth it

Measure before you build. If you do not know how long the job takes today, or how often it goes wrong, you cannot prove the agent improved anything. Track six things.

Three habits make the numbers believable.

• Keep a baseline and a holdout. Compare the agent’s results with the old way over the same period, or run the agent on half the campaigns and leave the other half alone.
• Count the review time. If an agent saves ten hours and creates six hours of checking, the gain is four hours.
• Track AI-referred traffic separately. Visits that come through AI assistants are often hard to attribute, so set up their own reporting rather than letting them disappear into “direct” or “other.”

When you review at the end of a pilot, ask one plain question: if this agent were switched off tomorrow, would anyone notice and complain? If yes, it earned its place. If nobody would notice, stop the project and use the budget elsewhere.

Caveats: where agents go wrong

I would rather you hear this from me than from your own post-mortem. These are the problems that come up again and again.

A word on infrastructure. Agents work best on connected, well-organised data with clean APIs. If your customer data is split across five tools that do not talk to each other, an agent will inherit that mess. You can still start, because a narrow agent on one clean data source works fine. But if you plan to run several agents across the business, budget for the data consolidation and transformation work as part of the project. It is not an extra.

None of this is a reason to avoid agents. It is the reason to start with a narrow job, keep the limits tight, and widen them as the evidence comes in.

Where this is heading

The direction is clear even if the timing is not. Routine marketing work, such as pulling reports, resizing creative, checking tracking, and monitoring spend, will increasingly be done by agents. The people who do well will be the ones who know what to ask for, can tell good output from bad, and spend their saved hours on strategy, creative thinking and customer relationships.

Here is my own read on the commerce side, and it is an opinion, not a forecast. If shoppers’ agents do more of the comparing, then price, stock, delivery speed, reviews and clean product data count for more at the comparison stage, and banner-style persuasion counts for less. Brands will still need to be known and trusted, because an agent leans on reputation too. But the days when a well-designed page could carry a weak offer are numbered.

What I would do this month:

• Pick one repeatable job, write down its metric, and record today’s baseline.
• Audit the data that job depends on.
• Run the job by hand with an AI assistant and note where it fails.
• Check how your top products show up when you ask the main AI assistants about them.
• Decide who owns the result and who approves the agent’s work.

Agents are not magic, and they are not just hype. Think of one as a new hire who is fast and tireless but needs clear instructions and supervision. Treat it that way and it will earn its place.

Want a second pair of eyes?

If you want a second pair of eyes on your first use case, or want to check whether your store is ready for shoppers’ agents, get in touch. It is the kind of thing I work on every week across AI and automation and eCommerce growth.