AI agents for marketing: which to build first, and how.

01What are AI agents for marketing?

AI agents for marketing are pieces of software that each own one marketing job, such as search or outbound, and run it on a schedule: they read the business, produce the work and hand it to a person to review before it ships. A small business runs them as a team, all briefed from the same positioning.

The definition of an AI marketing agent covers what makes software an agent. This guide covers which agents to build, in what order, and how.

Few businesses have got this far. Salesforce's tenth State of Marketing report, surveying 4,450 marketers late 2025 including 890 at small and medium businesses, found only 13% using agentic AI¹, while 80% of the SMB marketers expected AI agents to improve their marketing return. Among smaller organisations the picture is flat: McKinsey's State of AI 2026, surveying 1,719 people in mid 2026, found the share of smaller organisations scaling AI agents unchanged at 22%², against 40% at companies with revenue above US$1 billion. The gap is knowing where to start.

02The ten marketing agents, by job

A complete marketing function runs on ten agents: search, content, social, outbound, conversion, launch, paid, community, measurement and reporting. Each one needs specific inputs, has one review step a human owns, and earns its place at a different stage of the business. The table below is the reference; the next section turns it into an order.

Every agent's inputs include something only the business can supply, which is why a generic tool produces generic work. Every review step is one named decision, small enough to make weekly: the human in the loop has a clear question to answer.

03Which AI agent should you build first?

Build the positioning brief first, then the agent for the gap that costs the most. The brief is a one-page record of who you sell to, what you claim and what proves it, and every agent reads it before producing anything. Without it, ten agents produce ten slightly different businesses, and review time balloons correcting the drift.

After the brief, the right first agent depends on where the business is losing the most money today. Three questions usually settle it:

1. Do buyers search for what you sell? If yes, and you rarely appear in Google or in AI answers, the search agent comes first. It also produces the topic list the content agent needs later. The answer engine optimisation work sits here.
2. Do you sell to a list of businesses you could name? If yes, and the pipeline depends on referrals, the outbound agent is the faster return. The guide to outbound sales automation covers the mechanics.
3. Is traffic arriving without turning into enquiries? Then the conversion agent beats anything that adds more traffic.

From there, a default order suits most businesses of 5 to 50 staff. Measurement goes in before the second month of anything, so the first agent can be judged. Content follows search, which names the topics, and social follows content, which it repurposes. Paid comes after measurement, because spend against broken tracking teaches nothing. Launch and community switch on when there is a reason. Reporting starts in month one as a single page.

04How to build an AI agent for marketing

Building a marketing agent takes five parts in sequence: a written brief, scoped access to tools and data, guardrails, a review gate and a measure of success. Each part is a short document or setting, and skipping any one is where most agent projects stall. The steps below apply whether the agent is coded, no-code or an installed skill.

1. Write the brief. One page: the agent's single job, the voice rules, examples of good and bad output, and what it hands back. A brief longer than a page describes two agents.
2. Give it scoped tools and data access. List every system the agent touches and give the narrowest permission that works: read-only analytics, draft-only email, a CRM view limited to the fields the job needs. Salesforce found only 27% of SMB marketers completely satisfied with their ability to unify customer data¹, so expect to tidy the data before the agent can use it.
3. Set the guardrails. Write them as rules the agent cannot override. Daily sending limits per inbox. A human takeover the moment a prospect replies, with automation paused for that contact. Voice rules and banned claims. A spending ceiling for anything paid. A list of topics it never writes about.
4. Put in a review gate. Decide exactly which outputs a person approves, who that person is, and how long they have. Approve everything at first, and loosen the gate only for output that has passed unchanged for several weeks.
5. Measure one number. Pick the number the agent exists to move, such as booked meetings from outbound. Track cost per enquiry where money is involved, and review it monthly alongside the time review took.

An illustration: a content agent for an Auckland accounting firm reads the positioning brief and the search agent's topic list, has draft-only access to the blog, is barred from anything that reads as personal tax advice, and a partner approves each article on Tuesday. Its number is enquiries that start on an article page.

05Build, buy, no-code or installed skills: which route fits?

There are four ways to get a marketing agent running: buy a product, build on a no-code workflow tool, install written skills into an AI assistant, or code it. The right route depends on how specific the job is and who will maintain it. Many businesses mix routes, buying one agent and building another.

A useful rule for a small team: buy where the job is generic, use installed skills where judgement matters, connect them with no-code, and code only what is left. The guide to AI marketing automation covers the delivery layer most agents sit on.

06What a marketing AI agent workflow looks like in a week

A marketing agent workflow runs on a weekly loop: the agent reads what changed, produces the week's work, a person reviews it in one sitting, approved work ships, and the results feed the next cycle. Once the brief is right, the owner's part should settle at under an hour a week per agent.

• Monday: the agent reads last week's numbers and review notes, then drafts the plan.
• Tuesday to Wednesday: it produces the work, such as two articles or a batch of first messages.
• Thursday: a named person reviews it in one sitting and approves or rejects with a reason.
• Friday: approved work ships, and rejection reasons go into the brief.

Salesforce reports that AI agents free marketers 6 to 7 hours a week for creative and strategic work¹ (surveyed late 2025). Those hours appear only if review stays a fixed weekly slot.

07What can go wrong with marketing agents, and how to contain it

The main risks are unreviewed output reaching customers, messages that break consent law, personal data handled loosely, and agents that cost more than they return. Each has a containment that is cheap to set up before launch and expensive to add after a problem, and most sit at the point where work meets a customer.

People trust AI output more than they check it. KPMG and the University of Melbourne found 51% of New Zealand employees had relied on AI output at work without evaluating its accuracy³ (surveyed over summer 2024-25). A review gate with a named owner is the containment.

• Consent and unsubscribe. New Zealand's Consumer Protection guidance on the Unsolicited Electronic Messages Act requires consent, accurate business contact details and a free unsubscribe that works within five business days, with fines up to $500,000⁴. Outbound and email agents need those rules built in, and unsubscribes synced to every tool.
• Enriched lead lists. The Privacy Commissioner confirms that from 1 May 2026 the new IPP3A requires reasonable steps to tell people when their information was collected indirectly⁵. Check it before an outbound agent buys contact data.
• Personal data in public tools. In Australia, the OAIC's October 2024 guidance recommends organisations do not enter personal information into publicly available generative AI tools⁶. Keep customer data in systems you control and pass agents only what the job needs.
• Claims the business cannot back. Guarantees, prices and credentials stay on the banned list unless they appear word for word in the positioning brief.
• Cost without return. Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls⁷. One agent, one number, reviewed monthly, is the cheapest defence.

08How to judge claims about the best AI agents for marketing

Judge any "best AI agents for marketing" list by whether it was tested on businesses your size, names the human review step and shows real output from a live business. Many lists rank products by feature count, and some by paid placement. The questions below work on any vendor page or ranking, and a good one answers them plainly.

The caution is earned. In the same June 2025 release, Gartner estimates only about 130 of the thousands of agentic AI vendors are real⁷, in a passage on what it calls agent washing.

• Who wrote the ranking? A vendor's list that includes its own product is marketing. Look for the method and any affiliate disclosure.
• Was it tested at your size? Enterprise rankings assume a data team a 20 person firm does not have.
• Does it plan its own work? If a person starts every task, it is a tool with an agent label.
• Where does the human sit? The approval points should be named and adjustable.
• What does it read first? An agent that never reads your site and analytics writes for an average business.
• What do you keep if you leave? Pages, content, workflows and logs should stay on your accounts.
• Can you see real output? Ask for a page, a sequence or a report from a live business. Demos do not count.
• What does it report? Enquiries and cost per enquiry. Follower counts and "content produced" are activity.

For the planning role that sits above the agents, the guide to the AI CMO applies the same questions to products that claim to run the whole function.

09Where Flow AI fits

Flow AI is an AI marketing transformation practice: the strategy, the systems and the agents that run a marketing function, engineered end to end. It runs the ten agents in this guide from one positioning brief, installs them one at a time as week-long sprints, and a marketing engineer reviews every send before it reaches a customer.

The scope reads the business and names the first agent to build and why. If you would rather build your own, the free agent templates are a starting point for the brief, the guardrails and the review step.

10The short version

AI agents for marketing work best as a small team built in order. Write the positioning brief, then build the agent for the most expensive gap, usually search or outbound. Give each agent a one-page brief, narrow access, written guardrails, a named reviewer and one number to move.

Add the next agent when the last one runs weekly with under an hour of review. Measure before you spend, and ask every "best agent" list to show real output.

To find out which agent your business should build first, talk to Amber.