AI Agents in Marketing: What They Can (and Can't) Do Yet

Every second pitch deck I see now has "AI agent" somewhere in the first three slides, and most of them are describing a chatbot with a system prompt. That's not a criticism of the underlying tools - some of it is genuinely useful - but the term has been stretched so far that clients ask me to "build an AI agent" without being able to tell me what decision they want it to make on its own. Before you hand anything over to an agent, you need a clear picture of what the word actually means, what these systems are reliably good at right now, and where they still produce confidently wrong output that looks correct until someone checks it.

What "Agent" Actually Means, Technically

A chatbot answers a question. A single API call to an LLM takes an input and returns an output - one step, no memory of what happens next. An agent is different: it's an LLM given a goal, a set of tools it can call (a CRM lookup, a search query, an email send, a database write), and the ability to decide which tool to use, evaluate the result, and decide the next step - in a loop, without a human choosing each action.

That loop is the whole distinction. A workflow automation tool like Zapier or Make executes a fixed sequence you designed - if this, then that, every time, the same way. An agent decides its own sequence based on what it observes at each step. Ask a workflow to "follow up with leads who haven't responded in 5 days" and it does exactly what you configured. Ask an agent the same thing and it might check the CRM, notice the lead opened three emails but never clicked, decide that's a different signal than a lead who never opened anything, and choose a different follow-up angle for each. That reasoning-and-tool-use loop is what makes agents more capable than automation - and also what makes them harder to predict and audit.

Multi-agent systems take this further: several agents with different roles (a research agent, a drafting agent, a review agent) pass work between each other. This is where most of the current hype lives, and it's also where most of the current failure modes live, because errors compound across handoffs instead of staying contained to one step.

Where Agents Genuinely Save Time Today

I'm not an AI skeptic - I run agents in production for client marketing ops, and in a handful of specific, bounded tasks they save real hours every week. The pattern across all of them: high volume, low-to-medium stakes per instance, and a clear success signal the agent can check itself against.

Lead Qualification and Routing

An agent that reads inbound form submissions, cross-references company size and industry against your ICP, checks CRM history for prior contact, and assigns a lead score with routing to the right sales rep is a strong fit. The task is repetitive, the inputs are structured, and a wrong call on any single lead is low-cost - it gets caught downstream when a rep looks at the account.

First-Draft Content

Agents that pull from a content brief, existing brand guidelines, and past-performing posts to produce a first draft of a blog outline, an ad variant set, or a social caption batch are genuinely useful. The output is never publish-ready - it's a starting point that removes the blank-page problem. Treat it the way you'd treat a junior writer's first pass: useful structure, needs an editor.

Campaign Monitoring and Alerts

This is one of the highest-value, lowest-risk uses I've deployed. An agent that checks Google Ads and Meta Ads spend pacing daily, flags a campaign that's burned 80% of its monthly budget by day 12, checks whether conversion rate dropped alongside a CPC spike, and posts a summary to Slack with the likely cause - that's an agent doing something a human would otherwise do manually every morning, and doing it consistently. The agent isn't making the budget decision; it's surfacing the signal fast enough that a human can.

Customer Support Triage

Reading an inbound support ticket, classifying it (billing, technical, sales inquiry), checking whether it matches a known issue, and either drafting a response for review or routing to the right queue is a well-bounded task agents handle well, because misclassification is cheap to catch and correct.

Where They Still Need a Human in the Loop

The common thread in every failure I've seen or heard about from other agencies is the same: the task required judgment that depends on context the agent doesn't reliably have, or the cost of a wrong call was high and not easily reversible.

Strategic Judgment Calls

Deciding to shift 30% of quarterly budget from Google Search to LinkedIn because of a market shift, or deciding to pause a campaign during a PR situation, requires weighing context an agent doesn't have full visibility into - competitor moves, internal politics, brand risk tolerance that isn't written down anywhere the agent can read. Agents can prepare the analysis. They shouldn't make the call.

Brand Voice at the Final Mile

An agent can draft in a brand's general tone reasonably well if it's been given examples. What it consistently misses is the judgment about when to deviate - when a serious client update needs a different register than the usual playful brand voice, or when a joke that works in one market reads badly in another. That's a nuance layer that still needs a human sign-off before anything goes out under the brand's name.

Anything With Real Financial or Legal Risk

Don't let an agent autonomously change live bid strategies beyond a pre-approved range, send contractual terms, or make claims in ad copy that touch regulated categories (health, finance, legal services) without review. The failure mode here isn't hypothetical - I've seen an automated bidding tool run unchecked overnight and burn a week's budget chasing a tracking anomaly that looked like a conversion spike. The fix isn't "don't automate bidding" - it's "cap the blast radius and require review above a threshold."

Common Failure Modes to Watch For

• Hallucinated data in reports. Ask an agent to summarise "why did conversions drop last week" without giving it a strict instruction to only cite numbers it actually pulled from the data source, and it will sometimes fill gaps with plausible-sounding but invented figures. Always require the agent to cite the exact query or data pull behind any number in a report, and spot-check against the source before it reaches a client or a decision-maker.
• Over-trusting automated decisions because they're fast. Speed creates false confidence. An agent that reroutes ad spend in seconds feels more authoritative than a human who takes a day to think it through, but speed isn't accuracy. Build in a review step for any decision above a defined dollar or risk threshold, regardless of how fast the agent produced it.
• Tool access that's broader than the task needs. Giving a content-drafting agent write access to your CMS "for convenience" means a bad output can go live without review. Scope tool permissions tightly to what the specific task requires.
• No fallback when the agent gets stuck in a loop. Multi-step agents can loop on a task they can't resolve - retrying the same failed API call, or looping between two contradictory sub-goals. Set a hard step limit and a timeout that hands the task back to a human rather than letting it run indefinitely.
• Treating agent output as ground truth for client reporting. An agent-generated summary should be labelled as a draft internally until a human has verified the underlying numbers, especially for anything client-facing.

A Practical Framework for What to Hand Over First

When a client asks where to start, I use two questions to score candidate tasks: how repetitive and high-volume is it, and how reversible or low-cost is a mistake? Tasks that score high on both - repetitive and cheap to get wrong - are the right starting point. Campaign spend alerts, lead scoring, first-draft content, and support ticket triage all sit in that quadrant.

Tasks that are repetitive but expensive to get wrong (bid adjustments beyond a small range, autoresponder sequences that touch sensitive topics) need an agent with a human approval gate, not full autonomy. Tasks that are infrequent and high-stakes - budget reallocation, brand positioning, crisis response - shouldn't be handed to an agent at all right now. Use the agent to prepare the analysis; keep the decision with a person.

Start with one bounded task, measure the time saved and the error rate over a month, and only then expand scope. The agencies and marketing teams getting real value from this are the ones treating it as a gradual capability rollout, not a wholesale replacement of judgment.

For the tracking foundation any of this depends on, see our GA4 migration guide - an agent is only as reliable as the data it's reading. And if you're weighing where automation fits into a broader martech stack, the Marketing Automation ROI post covers how to measure whether any of this is actually paying for itself.