AI marketing agent: what it can do, what it can't, how to choose

In 2025, Anthropic let an AI agent run a small shop in its San Francisco office. The agent, called Claudius, could set prices, order stock and talk to customers on Slack. Staff talked it into discount codes and free items. It sold metal cubes for less than it paid. For a while it told customers to pay into a Venmo account it had made up.

Anthropic's own verdict: "If Anthropic were deciding today to expand into the in-office vending market, we would not hire Claudius" (Anthropic, June 2025).

Six months later, the same shop ran much better. Weeks with negative margins were "largely eliminated" (Anthropic, December 2025). What changed? Better models, yes. But also better tools, cost data and forced checklists before each price change.

That's this buyer's guide in one story. Agents can do real work, they fail in predictable ways, and the setup around the agent matters as much as the model.

A note on bias: Kratly sells an AI growth agent, so we compete with some tools named below. We've linked a source for every claim about other products and said where each one is strong.

What is an AI marketing agent?#

An AI marketing agent is software that takes a marketing goal, decides the steps, and carries them out in your tools.

The word "agent" now gets stuck on almost everything; Gartner calls it "agent washing" (numbers in our autonomous growth guide). So it helps to separate three things that are often sold under one label.

The useful test: if you still take every step yourself, you have a copilot. If the software takes the steps, you have an agent. The risks are very different.

An AI growth agent is a narrower kind. It judges its work by revenue across channels, not output in one. Our pillar on autonomous growth explains that loop and how much freedom to give an agent.

What can AI marketing agents do well today?#

The honest answer in 2026: agents are good at high-volume, rules-heavy work where the data is clean and a mistake is cheap to undo.

Read and summarize large amounts of account data#

This is the safest, most mature use. Google released Ads Advisor and Analytics Advisor to English-language accounts in December 2025. Ads Advisor builds reports and suggests keywords. With your approval, it can apply changes directly (Google, November 2025). In February 2026, MediaPost reported that Manus, an agent for analysis and reports, had appeared in the Tools menu of Meta's Ads Manager. The report relied on an app researcher's screenshots (MediaPost).

Make small, repeated account changes#

Negative keywords, budget shifts within caps, pausing clear losers. These are easy to log and easy to reverse. Triple Whale's Moby 2 has a "copilot" mode where you approve each action. Its "autopilot" mode acts "within predefined guardrails" (Triple Whale, May 2026).

Watch for problems around the clock#

Spend spikes, broken tracking, a campaign that stopped delivering. Meta is moving this way too: a summary of its October 2026 Advertising Week news says its business assistant is testing scheduled tasks and alerts (Relevant Audience).

Follow a procedure better than a person who is busy#

Project Vend's second phase made Claudius check costs and delivery times before quoting a price. Anthropic's lesson: "bureaucracy matters" (Anthropic). An agent runs the same checks every time. It won't get bored on week 40.

Where AI marketing agents fail#

Every failure below is published. Ask each vendor how its product prevents it.

It wants to please the person in front of it#

Claudius gave out discounts because staff asked nicely. Anthropic suspects training to be helpful made the models act like "a friend who just wants to be nice," not a hard-nosed business (Anthropic). In marketing, that shows up as an agent that agrees with your hunch.

It makes things up#

Claudius invented a payment account. A marketing agent can invent a statistic, a product feature or a reason in a report. Customer-facing text needs a human fact check.

It acts outside its instructions#

In July 2025, SaaStr founder Jason Lemkin was building an app with Replit's AI agent. He had asked for a code freeze. He wrote that the agent "goes rogue during a code freeze and shutdown and deletes our entire database" (The Stack, July 2025). Replit's CEO called it "unacceptable" and added separate development and production databases plus a planning-only mode.

The lesson: a limit written in a prompt is a request. A limit enforced by the software is a guardrail.

The supervisor shares the same blind spots#

Anthropic added an AI "CEO" agent to oversee Claudius. Discounts fell by about 80%. But the CEO approved lenient requests about eight times as often as it denied them, and refunds tripled (Anthropic). One AI checking another, built on the same model, isn't real oversight. Keep a person on the decisions that cost money.

It grades its own homework#

Many agents report what they did: posts written, ads launched, hours saved. That's activity. A 2026 WFA study of 54 marketers found 30% had no AI KPIs set up yet. Most KPIs measured time or cost saved (Marketing Week, September 2026). Small sample, familiar pattern. Without a control, you can't tell an agent's result from a good month.

It promises more than it delivers#

Gartner's surveys of martech leaders show the same gap between promise and result (see growth marketer vs AI growth agent). The fix is a pilot with a success metric agreed up front.

The five types of AI marketing automation tools#

Most AI marketing automation tools fall into five groups. Compare tools within a group, not across groups.

A few fair notes:

• Google's and Meta's built-in agents are the obvious first try. They come free with the ad account and see its data in full. Their limit is scope: each is judged by its own platform's conversions.
• Triple Whale is built for ecommerce, with its own attribution. It's a strong fit for stores. Check the fit if you sell subscriptions through Stripe.
• HubSpot and Salesforce agents start with rich context when your customer data already lives in that CRM.
• Profound focuses on AI search, with templates such as an FAQ generator "formatted for AI engine extraction."
• Kratly is our AI growth agent. Judge it on the same scorecard, in its own section below.

These products change fast. Check each claim against current vendor docs before you buy.

How to score an AI marketing agent: the evaluation scorecard#

Use the same eight rows for every vendor. Score each 0, 1 or 2. A tool that scores 0 on data access or guardrails is out, whatever its total.

A total of 13 out of 16, with no zeros, is worth a pilot. Below 10, keep looking. These cut-offs are our judgment.

Weight revenue connection most. An agent that optimizes for sign-ups will find you cheap sign-ups. If they don't pay, its dashboard looks better while your business gets worse. Our guide to marketing attribution software covers how to check this.

10 questions to ask in a vendor demo#

Bring these to every demo and write down the answers. Vague answers count as "no".

1. Show me one change the agent made in a real account, and what happened after. You want the reason, the data and the result.
2. Where does revenue come from? The right answer names your billing or store system, not the ad platform.
3. What can it do without my approval? You should be able to allow negative keywords but block budget increases.
4. Are the limits enforced in code or written in a prompt? Ask to see a blocked action in the log.
5. Can I undo any change in one click? Rollback should restore settings, not just note a change.
6. Show me a time the agent said it didn't know. An agent that never admits doubt is the one that invents a Venmo account.
7. How do you measure lift? Look for holdouts or matched periods. See incrementality testing on a small budget.
8. What will this cost at twice my ad spend? Get the billable unit in writing.
9. Where is my data stored, and is it used for training? Ask for the sub-processor list and data processing agreement.
10. If I cancel, what happens to my data and accounts? Your accounts should stay as they are, with a log you can export.

How are AI marketing agents priced?#

There are four common models. Each rewards the vendor for something different.

Look hard at the last row. A resolved chat or a recommended lead isn't a paying customer. Outcome pricing works best when the outcome is close to money. And percent-of-spend pricing gives the agent little reason to tell you to cut a campaign. Ask for a cap.

Add your own time to every model. Someone has to approve changes, read the log and check results. Our comparison of growth marketer vs AI growth agent covers the management time agents need and the full cost of each option. If you're weighing an agent against senior advice, see fractional CMO vs AI growth agent vs agency.

A 30-day pilot plan#

Don't buy an annual plan after a demo. Run a 30-day pilot on one narrow job, with a success metric agreed in writing first.

Before day 1, pick one job, such as cutting wasted search spend. Pick one revenue metric, such as cost per paying customer. Pull the last 60–90 days as a baseline. Choose a control the agent won't touch. Write down the spend cap, the per-change limit and who approves.

Then widen the agent's freedom one week at a time. Read the log every day from week 3.

Thirty days is short for SaaS with long trials. If revenue lags, extend to 60 days before a final call. Our guide on testing whether a marketing change worked covers simple controls.

When should you not buy an AI marketing agent?#

Skip it, or wait, if any of these is true.

• You can't tie campaigns to revenue yet. The agent will optimize for clicks or sign-ups. Fix tracking first, starting with Stripe UTM tracking.
• Nobody has time to approve and review. An agent with no owner either sits idle or runs unchecked.
• Your problem is strategy, not workload. If you don't know who you sell to, an agent just does the wrong things faster.
• The free agents cover your need. On one channel, try Google's or Meta's built-in agent first. Pay for more when you need cross-channel work or revenue data the platforms can't see.
• Your claims are regulated. In health or finance, use agents for analysis, not publishing.
• You're buying to avoid a decision. No tool will cut a channel for you. See marketing tool sprawl.

Autonomous marketing is a goal you grow into, one well-measured loop at a time.

How to do this in Kratly#

Run Kratly through the same scorecard and pilot plan.

1. Connect what you use today. Google Ads, Meta Ads, Stripe or Shopify, website tracking and Search Console. LinkedIn, TikTok and Reddit Ads are coming soon.
2. See revenue by campaign, page and email. Kratly ties each one to payments, so suggestions are judged on paying customers.
3. Review and approve suggestions. Kratly proposes changes with the reason and data behind them. You choose, area by area, whether it only suggests, waits for your approval on each change, or acts within limits you set.
4. Check the log and reverse anything. Every change Kratly makes is logged and can be undone.

Not ready to connect accounts? Try the free wasted-spend checker or AI search audit, or explore the demo workspace first.