fm.

AI agents for marketing: what I hand to agents and what I never do

What an AI marketing agent is, the jobs I hand to agents and what they produced, the calls I keep and how to brief and judge one.

Key takeaways

  • An AI marketing agent takes a goal, plans the steps, acts in your real systems and stops for approval wherever money or customers are involved.
  • I hand agents high-volume work with checkable output: account research, paid operations, creative production, AEO analysis and reporting. At Pipe17 a 12-agent pipeline grew a target list from 441 to 2,795 accounts.
  • I never hand over spend changes without approval, customer sends without review, positioning calls or anything I can't roll back.
  • Judge an agent like a channel, on the cost of a qualified outcome. I don't put much weight on hours saved.

01What an AI marketing agent is

An AI marketing agent is software that takes a marketing goal, plans the steps and carries them out in real systems such as your ad account, CRM and analytics, then reports back with the evidence. The useful ones also know where to stop: they stage anything that spends money or reaches a customer and wait for a person to approve it.

A lot of what gets sold as AI agents for digital marketing is a chatbot or a fixed workflow with a model inside. Both are useful, but the difference decides what you can safely hand over. I run my own agents on Claude Code and Codex. This note covers what I give them, what I keep and how I judge the work.

02Agent vs automation vs chatbot

TestAutomationChatbotAgent
Who plans the stepsYou, in advanceYou, one prompt at a timeThe agent, inside limits you set
What it touchesThe systems you wired, in the order you wired themThe conversation, plus any connectors you addFiles, tools and APIs from an allowed set
When inputs changeBreaks, or quietly does the wrong thingAnswers whatever it's askedAdapts, or stops and asks
How it failsSilently, at scaleConfidently, in textConfidently, in your systems, unless writes need approval
Best marketing useNurture sends, lead routing, threshold alertsDrafts and quick answersResearch, audits, reporting and staged account changes
From my work18 HubSpot workflows and 64 emails at Commas, each measured against a holdout10,560 ChatGPT and AI Overview answers analyzed for AEOA 12-agent research pipeline at Pipe17

Most teams need all of them. The expensive mistakes run both ways: an agent doing a job a fixed workflow does better, like a nurture send, or a fixed workflow doing a job that needs judgment, like scoring accounts for fit.

03The jobs I hand to agents

My test has 3 parts. The output can be checked against a source, a mistake can be undone and the job repeats. These jobs pass.

JobWhat the agents doWhat it produced
Account researchAt Pipe17, 12 parallel agents researched accounts, scored fit and found the champion.A target list grown from 441 to 2,795 accounts and 13,000+ companies classified for tiering
Paid operationsSearch-term mining, negatives, ad rewrites, audits and deal-level attribution through my gads and hs tools.$66.5K in paid influenced $1.3M in pipeline in one quarter
Creative productionAt Commas, an AI creative engine generated, tagged and rotated launch ads against the first-payment signal.100+ launch creatives behind about $1M a month in paid
AEO analysisAt Commas, 10,560 ChatGPT and Google AI Overview answers read for who gets cited.An 8.4% citation share against 21% to 33% for peers, turned into 221 page opportunities and a 35-topic roadmap
Reporting and alerts26 scheduled jobs on an always-on server, and at Commas the data behind a weekly leadership scorecard.A scorecard linking spend, bookings, CRM stages, approvals and processed volume, and alerts that lead with the verdict and fire only when something needs a decision
AI referral trackingA monthly LLM-referral report next to attribution rebuilt on what buyers said they used.32% of 2026 new-business wins at Pipe17 traced to AI assistants, revenue the CRM had credited to paid search

That's the pattern for AI agents in marketing: volume, an answer you can check and a clear place to stop. None of these jobs asks an agent to decide what the company stands for. The full stories are in the Pipe17 and Commas case studies.

04What I never hand over

  • Spend changes without approval. Budgets, bids, campaign status and account settings. Agents stage the change, show a preview and wait for my go.
  • Sends to customers without review. Anything that reaches a customer gets read first, and every sender runs with daily caps, bounce stops and a kill switch.
  • Positioning and judgment calls. Which segment to chase, what we claim and which channel to cut. An agent can pull the deal-level evidence. Cutting a six-figure LinkedIn line that had produced no closed revenue, as I did at Pipe17, stayed a person's call.
  • Anything without a rollback. If I can't undo it, an agent doesn't do it.
  • Facts it would have to guess. Agents will confidently invent an account ID. IDs, budgets and dates live in a canonical facts file, 184 entries, that every session reads first.

05How to write an agent brief

A vague brief gets you 12 confident, wrong answers. The brief in my note on running growth on Claude Code and Codex has 5 parts: the goal, what's ruled out, what to read first, what to return and when to stop. For work that will sit in front of Sales or leadership, add 4 more: the tools allowed, a definition of done, a check and who decides next.

PartWhat to writeExample: account research
GoalOne outcome, with a numberScore 400 accounts for fit and find the likely champion at each.
Ruled outWhat not to spend time onCompanies under 20 people, agencies and current customers.
Read firstThe files that hold your judgmentThe ICP definition, the account list and the last 2 win and loss notes.
ToolsWhat it may use and what it must not touchWeb research and enrichment. No CRM writes.
ReturnThe exact shape of the outputOne row per account: fit score, champion name and title, source URL and confidence.
Done meansThe test for finishedEvery row has a source, or a stated reason it doesn't.
CheckHow the work gets verifiedA second pass re-checks 1 row in 10 against its source.
Stop ifThe condition that ends the run earlyMore than 10% of rows have no source.
Decides nextWho acts on the outputMe. Nothing imports to HubSpot without a go.

When you fan out, give every agent the same brief and a different slice of the input, then run one merge step that dedupes and ranks. Same brief, same output shape. If the agents need to talk to each other mid-task, the job is too big for one brief.

06How to judge whether an agent is working

Judge it like any channel. The number that matters is the cost of a qualified outcome: a sales-qualified opportunity in B2B, a first payment at a payments company, a paying merchant in commerce. At Pipe17 the agent work sat inside a program that cut cost per SQO 74%, from $14,450 to $3,700, in 2 quarters. Hours saved is a weaker signal, because an agent can save hours producing work nobody uses.

Between those reads, watch 4 leading indicators:

  • Spot-check error rate. Re-check a sample against sources on every run. If it climbs, the brief or a source changed.
  • Approval rate. The share of staged changes approved without edits. When it's low, the agent doesn't know the account yet, and the fix is usually the facts file.
  • Time to a sourced answer. When answers take minutes, you ask more questions and catch problems sooner.
  • Alert precision. Alerts should lead with the verdict and fire only when something needs a decision. If nobody acts on an alert, I turn it off.

07Build or buy

There are 2 routes. Agent platforms sell a packaged agent for a job, usually inside one channel or tool. The other route is a general coding agent, such as Claude Code or Codex, connected to your systems and to small tools you write. I took the second route, and the tools I wrote replaced about $24K a year in software. If you're comparing marketing AI agents from platforms, these are the questions I'd ask.

QuestionLean toward a platformLean toward building
Where the job livesInside one tool the vendor controlsAcross ads, CRM, analytics and the site
How often it changesRarely, and the vendor's workflow fitsOften, because your process is part of your edge
Who maintains itNobody on the team wants to touch a scriptSomeone can read a script and rewrite a brief
ApprovalsThe platform's controls are enoughYou need your own limits, previews and rollbacks
DataThe vendor sees everything the job needsThe answer joins data from several systems
ExitYou're happy to stayYou want to own the tools and the memory

Either way, the hard parts are the same: a written brief, a facts file the agent reads first, approvals on spend and one number to judge it by. If you build, start with my Claude Code for marketing setup.

08Where to start

  1. Pick one job that passes the test.Its output can be checked, a mistake can be undone and it comes back every week. A weekly paid audit or a monthly AI referral report is a good first job.
  2. Write the brief before the prompt.Use all 9 parts. If you can't write the “done means” line, the job isn't ready.
  3. Run it read-only for 2 weeks.Compare the agent's answer with your own every time.
  4. Add one write, with a preview and a rollback.Approve each change by hand until the approval rate is boring.
  5. Put the cost of a qualified outcome on the scorecard.That number decides whether the agent gets a second job.

FAQQuestions people ask

What is an AI marketing agent?
An AI marketing agent is software that takes a marketing goal, plans the steps and carries them out in real systems such as an ad account, CRM or analytics, then reports back with evidence. Unlike automation it decides the steps itself, and unlike a chatbot it acts outside the conversation.
What are the best AI agents for marketing?
The best ones are narrow: one job, a written brief, access to only the systems that job needs and a stop before anything spends money or reaches a customer. Judge any of them on the cost of a qualified outcome.
Can AI agents for marketing automation run unattended?
They can run reads, reports, research and staged changes on a schedule without a person watching. Spend changes, customer sends and anything without a rollback should wait for a person to approve.
How do you know if an AI marketing agent is working?
Track the cost of a qualified outcome, such as a sales-qualified opportunity or a first payment, before and after the agent takes the job. Between those reads, watch the spot-check error rate and the share of staged changes approved without edits.
Should you build or buy AI marketing agents?
Buy when the job lives inside one platform and the vendor's agent can see everything it needs. Build on a coding agent with your own tools when the job crosses systems, depends on your rules or is part of your edge.