Writing

• AI in Marketing
• 23 September 2026

Generative vs agentic AI in marketing: uses, risks and permissions

Generative AI in marketing creates content when you ask for it: copy, images, video, summaries, briefs. Agentic AI in marketing takes a goal and works through the steps to reach it using your other software, so it might build an audience, draft a customer journey, shift budget or schedule posts. The first hands you a draft and the second acts on your behalf, which is why it needs permissions, limits and someone checking its work.

Plenty of companies have bought agents, far fewer let them act. In a September 2026 HBR article, Thomas McKinlay (founder of the research advisory firm Science Says), Stefano Puntoni (Wharton) and Serkan Saka (San José State University) cite figures showing that 59% of enterprise organisations say they are using agentic AI, but only 9% have turned that into autonomous workflows, which is the whole point of having an agent in the first place.

What is the difference between generative and agentic AI?

IBM’s explainer on agentic AI vs generative AI draws the line cleanly: generative AI “creates new content such as text, images, video, audio and software code”, while agentic AI “focuses on achieving goals by planning, making decisions and carrying out multi-step workflows with varying levels of autonomy”. Most agents use a generative model underneath, so what matters to a marketing team is what the system is allowed to touch once it has produced something.

Be sceptical of the right-hand column’s labels, though. Gartner’s June 2025 prediction on agentic AI described “agent washing”, where assistants, robotic process automation and chatbots get rebranded as agents, and estimated that only about 130 of the thousands of vendors claiming agentic AI are the real thing. So a fair number of chatbots have simply had a new badge sewn on, and it’s worth asking any vendor what their agent can actually do without a human clicking a button.

How is generative AI used in marketing?

Generative AI in marketing is mostly about the cost and speed of making things. The everyday uses are unglamorous and genuinely useful: twenty subject-line variants before a test, a first draft of a product page, a summary of a 60-page research deck, rough concept images for a mood board, an email translated for a second market. Adobe says its Firefly commercial model was trained on Adobe Stock images, openly licensed content and public domain content, which is why brand teams worried about image rights tend to start there (that claim covers Adobe’s own model, not the partner models Firefly also offers).

The catch is quality you can’t see in review. Adam Peruta of Syracuse University’s Newhouse School co-led a study with Ipsos, reported in HBR as “AI-Generated Ads Perform Worse Than Human-Made Ones”, in which 3,000 US consumers watched 20 video ads for major brands, half made by people and half made entirely with AI from the same briefs. Consumers struggled to tell them apart, yet the AI ads were 14% weaker on short-term sales potential and 17% weaker on long-term brand equity. So the AI version sails through the Tuesday creative review and you find out a year later when the brand tracker dips.

In the UK there’s also the ad code. The ASA’s May 2025 guidance on disclosing AI in advertising says the existing rules “apply regardless of how content is generated, edited, or targeted”, and that you can’t make a misleading claim and then “disclaim” it by saying AI was involved. If your AI-generated cosmetics image shows results the product can’t deliver, a small “made with AI” label won’t save it. For regulated categories, this sits alongside the rules in my piece on UK pharma advertising rules.

How are AI agents used in marketing?

Most UK marketers already run something agent-like without calling it that. Google Ads’ Performance Max takes your conversion goals, optional ROAS or CPA targets, creative assets, audience signals and budget, then sets bids, moves budget towards the inventory that performs and combines your assets into ads across YouTube, Display, Search, Discover, Gmail and Maps from a single campaign. That’s agentic behaviour inside one platform, and most teams have already felt both sides of it: less manual work, and less visibility into why the money went where it went.

The newer marketing agents stretch across more of the campaign. In an October 2024 post on Agentforce Campaigns, Salesforce described agents that write a campaign brief from a plain-language prompt, generate subject lines, body copy and calls to action for emails and landing pages within a branded template, turn a described audience into segment attributes, and configure “a draft journey with all your campaign content that you can further refine, approve, and activate”. HubSpot’s September 2025 Breeze agents update lists a Social Post Agent and a Blog Research Agent in public beta and a Data Agent that answers custom questions about customers by analysing CRM data, call recordings, emails, documents and external web sources.

Realistic use cases for AI agents in marketing today look like this:

• Answering a question like “which lapsed customers bought twice last year and opened our last three emails?” by querying the CRM, rather than waiting a week for the data team.
• Setting up a draft email journey, segment and variants from a brief, for a marketer to check and switch on.
• Reallocating paid media budget between campaigns or placements within caps you set.
• Drafting a week of social posts and slotting them into the calendar at suggested times, pending approval.

Notice how many of those end with a human switching something on. McKinsey’s State of AI in 2026 survey of 1,719 participants found 40% of respondents from large organisations report scaling AI agents, up from 27% a year earlier, while smaller organisations stayed flat at 22%, and agents are most often scaled in IT, knowledge management and software engineering rather than marketing. So if your agency tells you every marketing team is already running on agents, the survey data doesn’t back them up.

How do AI marketing assistants help with scheduling and managing campaigns?

This is where most marketers meet AI assistants first, and today’s tools mostly help with timing, drafting and triage while a person still presses publish. The Optimal Send Times feature in Sprout Social, a social media management platform, analyses how your audience engages with your profiles (comments, likes, messages) and suggests the top seven send times for a given day, falling back on global engagement trends when your own data is thin. In Buffer, another scheduling tool, the AI Assistant drafts, rephrases, shortens and repurposes captions and adapts them to each channel’s limits, and Buffer’s own help page tells users to fact-check everything before posting. Hootsuite pitches Wisdom as an AI agent that drafts posts, flags negative comments, spots trends and summarises performance, while stating that “people remain in control of reviewing, approving, and publishing content”.

For campaign management more broadly, AI marketing assistants earn their keep on admin: turning a brief into a task list, checking every link in a launch email has the right UTM parameters, flagging an ad set that has spent half its weekly budget by Tuesday. A prompt I’d happily give an assistant with read-only access to a campaign calendar looks like this:

You have read-only access to our Q4 campaign calendar and last quarter's
social performance export. For the three product launches in November:
1. List every asset still missing, by owner and deadline.
2. Flag any two posts scheduled within 2 hours of each other on the same channel.
3. Suggest posting slots using our own engagement data, and say where
   you're guessing because we have fewer than 20 past posts on that channel.
Do not change or schedule anything. Output a table I can paste into Slack.

The last line matters most, because an assistant with write access and a vague prompt will cheerfully “tidy up” your launch calendar for you the night before it goes live. If you’re wiring this into a wider process, my piece on how to build an AI marketing workflow goes through the handoffs step by step.

Why do AI agents in marketing stall after the pilot?

Gartner’s June 2025 release predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing rising costs, unclear business value and inadequate risk controls. Three HBR articles from September 2026 explain the mechanics better than any vendor deck I’ve seen.

People don’t trust the agent enough to give it anything important

McKinlay, Puntoni and Saka’s article, “To Adopt AI at Scale, Employees Need to Trust Agents”, summarises their Wharton Blueprint for AI Agent Adoption, built from a review of human-AI research and interviews with leaders at companies including Google, ServiceNow and Zapier. In one study of more than 3,000 people deciding whether to adopt an AI financial adviser, most concerns were about trust rather than performance: privacy carried 31% of the weight in the decision, fear the AI would take unwanted actions 23%, and worry about not understanding its actions 11%.

On what fixes it, they report:

• Telling people where an agent is likely to fail made it seem up to 14.6% more transparent, and people who were told its weaknesses worked with it up to 7.2% more effectively, even though the AI was identical.
• People accepted an AI’s recommendations 54% more often when it showed it understood their broader goal.
• People preferred agents that came across as competent over ones that came across as friendly and warm, which is awkward news for every tool that opens with “Great question!”

The authors’ opening scenario will be familiar to anyone who has set up an agent: it asks to “access, create, and delete all files”, “read and send emails” and “view and make payments”, the manager says no to all three, and the agent then gets stuck constantly. We refuse every permission and then conclude the agent is useless, blaming the vendor, the model, the data and probably the office wifi.

We point agents at the old process

Masha Shunko (University of Washington Foster School of Business) and Serguei Netessine (Wharton) argue in “Stop Automating Old Processes. Design New Ones Instead.” that the workflow, not the task, should be the thing you redesign. Two of their four failure modes, moving the bottleneck downstream and optimising the wrong metric, map straight onto marketing. Put an agent on content production and you get triple the drafts landing on the same two people who do brand and legal sign-off, so the queue simply moves. Put one on email and judge it by sends and open rates, and it will happily optimise for opens while unsubscribes creep up.

Their contrast case is Wayfair, the online furniture and homeware retailer, which worked with OpenAI to redesign its catalogue and supplier-support workflow. Low-confidence product tags go to a human or the supplier to confirm, and the system has corrected 2.5 million product tags and automates 41,000 supplier-support tickets a month. The authors are careful to say those figures show scale, not proven business value.

The agent learns how your team thinks

William Marks, a lecturer in Harvard Business School’s Technology and Operations Management unit, adds a different kind of risk in “Is Your AI Training the Competition?”. Enterprise contracts usually stop providers training on your prompts and outputs, but Marks points out that OpenAI’s Services Agreement and Anthropic’s Commercial Terms treat feedback as a separate category the provider can use, and that OpenAI’s documentation says a thumbs up or down means the whole conversation may be used for training, even if you’ve otherwise opted out. So every time your strategist corrects an agent’s segmentation logic, they may be teaching a tool your competitors also rent. Marks notes Anthropic added feedback controls to Team and Enterprise plans in March 2026 and suggests switching off the Rate Chats function.

He also cites a 2025 MIT study finding that while only about 40% of companies had bought an official AI subscription, employees at more than 90% of surveyed companies regularly used personal AI tools for work. A campaign manager pasting a customer list into a free chatbot is the side door nobody thought to lock.

Where should marketers use agents now, and what should wait?

Shunko and Netessine’s 4A framework is the most useful sorting tool I’ve found for this. Each activity gets one of four modes: Assist (AI suggests, human decides and acts), Approve (AI prepares the action, human approves before it runs), Audit (AI acts, humans spot-check) or Automate (AI acts within fixed guardrails, humans step in only for exceptions). You choose using two questions: who gets hurt if it goes wrong, and can you undo it?

What should shift an activity to a looser mode is a few months of audit logs showing the agent’s approved actions rarely needed correcting. Those same logs are the evidence you’ll need when finance asks what the agent actually returned, which I cover in measuring the ROI of AI in marketing.

How should you scope an AI agent’s permissions?

Adam Seligman, CTO of the automation company Workato, told McKinlay and colleagues that companies “need to be able to say, ‘You can draft this email, but you can’t send it'”, because without clear controls “agents stay stuck on trivial tasks, because nobody trusts them with anything important.” If I were setting up a first marketing agent (I lead AI adoption in marketing at the charity ACTAsia), this is the order I’d work in:

1. Write the pilot sentence first. Shunko and Netessine ask every team proposing a pilot to complete this: “We will improve [outcome] in [end-to-end process] by redesigning [workflow], using AI to [role], with humans retaining authority over [judgment/control], exceptions routed through [escalation path], success measured by [primary KPI], and performance protected by [guardrails].” If you can’t fill in the blanks, you have a technology experiment rather than a project.
2. List every action the agent could take and sort each into read, draft or execute. Start with read and draft only.
3. Default anything irreversible to Approve. Shunko and Netessine cite two coding incidents: in July 2025 SaaStr founder Jason Lemkin reported that Replit’s AI coding tool deleted a production database despite written instructions not to change code without permission, and in April 2026 PocketOS founder Jer Crane reported that an agent in the AI code editor Cursor found an overly broad API token and deleted the company’s production database and backups. The marketing equivalent is an agent holding your email platform’s admin key.
4. Give the agent its own credentials with the narrowest scope, a sandbox or test list for anything it sends, and hard spending caps set in the platform itself rather than in the prompt.
5. Tell the team where it fails. Write a one-page note on the tasks where its output needs double-checking, since that alone raised effective use by up to 7.2% in the trust research.
6. Check whether feedback and interaction logs can be used for training, switch off Rate Chats-style features, and make the enterprise account easy enough that nobody reaches for a personal one.
7. Design the unhappy path. When data is missing, confidence is low or a request clashes with a rule, the agent should stop and escalate to a named person, and every human override gets logged.

A permissions brief for a first social media agent might read:

Agent: social scheduling assistant
Can read: content calendar, last 12 months of post performance, brand guidelines
Can draft: posts, captions, suggested slots, weekly summary
Can execute: nothing. All posts go to the approval queue
Must escalate: anything mentioning pricing, health claims, competitors,
               or replying to a complaint
Cannot access: CRM, ad accounts, DMs, payment details
Review: 10% of drafts checked against brand guidelines each Friday

If your brand sits in a regulated sector, tighten the escalation list further; the rules in my piece on how pharma engages HCPs digitally in the UK don’t relax because an agent wrote the post. For a wider view of where all this fits, start with what AI in marketing actually covers.

The question I’d put to any marketing team this quarter is a plain one: which single campaign task would you let an agent do in Approve mode for the next 90 days, and who owns the log?

FAQs

Is Performance Max an AI agent?

Performance Max behaves like an agent inside Google’s own platform. You provide conversion goals, optional ROAS or CPA targets, creative assets, audience signals and budget, and Google’s AI sets bids, allocates budget and assembles ads across YouTube, Display, Search, Discover, Gmail and Maps. It pursues a goal across several steps without asking you each time, though it can only act within Google Ads, not across your wider marketing stack.

Can AI agents run marketing campaigns on their own?

Not sensibly yet, and the main vendors don’t pitch it that way. Salesforce describes Agentforce Campaigns producing a draft journey that marketers “refine, approve, and activate”, and Hootsuite says people stay in control of approving and publishing. Agents can prepare most of a campaign, but sending to customers, spending beyond set caps and changing customer data should still need a person to approve them.

What is the difference between marketing automation and agentic AI?

Traditional marketing automation follows rules you write in advance: if a contact downloads a guide, send email two after three days. Agentic AI is given a goal and decides the steps itself, such as building the segment, drafting the emails and setting up the journey. Gartner warns that some vendors rebrand chatbots, AI assistants and robotic process automation as agents, a practice it calls agent washing, so ask what the tool decides without you.

Do UK marketers have to disclose AI-generated ads?

There is no blanket UK rule requiring disclosure. The ASA’s May 2025 guidance says the CAP and BCAP Codes apply however an ad was made, and suggests asking whether audiences would be misled without disclosure. Disclosure can’t fix a misleading ad: an AI-generated image exaggerating a cosmetic’s results breaks the rules whether or not it carries a “made with AI” label.

What permissions should a marketing AI agent have at first?

Start with read access to the data it needs and the ability to draft, with nothing that executes. Anything irreversible, such as sending emails, publishing, deleting records or making payments, should need human approval, following Shunko and Netessine’s 4A framework in HBR. Give the agent its own narrowly scoped credentials, hard spending caps set in the platform, and a named person it escalates exceptions to.

Sources

1. Harvard Business Review, Thomas McKinlay, Stefano Puntoni and Serkan Saka, “To Adopt AI at Scale, Employees Need to Trust Agents”, 9 September 2026.
2. Harvard Business Review, Masha Shunko and Serguei Netessine, “Stop Automating Old Processes. Design New Ones Instead.”, 14 September 2026.
3. Harvard Business Review, William Marks, “Is Your AI Training the Competition?”, 8 September 2026.
4. Harvard Business Review, Adam Peruta, “Research: AI-Generated Ads Perform Worse Than Human-Made Ones, Even When Customers Can’t Tell Them Apart”, 3 September 2026.
5. IBM, “Agentic AI vs. generative AI”, 11 February 2025.
6. Gartner, “Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027”, 25 June 2025.
7. McKinsey & Company, “The state of AI in 2026: On the road to ROI”, 25 August 2026.
8. Google Ads Help, “About Performance Max campaigns”, accessed 23 September 2026.
9. Salesforce, “8 Ways Marketing Agents Can Help You Build, Launch, and Track Campaigns Like Never Before”, 21 October 2024.
10. HubSpot, “Spotlight product deep dive: Build your AI team with over 20 Breeze Agents and Assistants, plus the new Breeze Studio”, 3 September 2025.
11. Sprout Social Support, “Optimal Send Times”, accessed 23 September 2026.
12. Buffer Help Center, “Using Buffer’s AI Assistant”, accessed 23 September 2026.
13. Hootsuite, “Wisdom by Hootsuite: AI social media management agent”, accessed 23 September 2026.
14. Adobe, “Adobe Firefly FAQ”, accessed 23 September 2026.
15. ASA and CAP, “Disclosure of AI in advertising: striking the balance between creativity and responsibility”, 29 May 2025.