Why start here
AI assistants such as ChatGPT, Gemini, Claude and Microsoft 365 Copilot can now draft, summarise and analyse in seconds. Like any tool, they are good at some jobs and bad at others. In this level you will learn three things:
- How to sort your marketing tasks into three groups.
- The main ways AI assistants fail.
- How to read surveys about marketers' AI use, including what each one leaves out.
Think of an AI assistant as a fast, well-read junior colleague on their first day. They write quickly and know a lot in general. They do not know your brand, your customers, today's news or your legal duties. They will also give a confident answer even when they are wrong.
Task 1: Sort your work into three groups
Every marketing task fits one of these groups:
- AI-suitable: low risk, easy to check, and never published as it stands. Examples: brainstorming 20 subject-line angles, rewording your own notes into a tidy summary, writing first-draft alt text, producing a list of interview questions.
- AI-assisted with human review: AI produces a draft and a competent person checks it, edits it and signs it off. Examples: blog drafts, ad copy variants, social posts, a first pass at a campaign report, keyword clustering.
- Human-only: tasks that need accountability, real relationships, confidential judgement or a final decision. Examples: approving a claim before it goes live, setting brand positioning, choosing between agencies, handling a complaint from a key client, deciding budget.
DO this: sort one task using these steps.
- Name the task in one sentence, e.g. "Write three Meta ad headlines for our spring sale."
- Check the facts it needs. Does it depend on facts that must be true and up to date, such as prices, dates, statistics or product specifications? If so, it cannot be AI-suitable.
- Check the cost of a mistake. Would an error reach customers, regulators or the budget? If so, it needs human review at the very least.
- Check for judgement and accountability. Does it need a decision someone must answer for, or a relationship? If so, it is human-only.
- Assign the group and write down the check you will run before anything is used.
Worked example: the spring sale headlines include a discount figure and an end date (step 2) and will be seen by the public (step 3). They are AI-assisted with human review. AI can write the variants. You check the offer details against the promotion brief.
Task 2: Learn the main ways AI fails
Look out for these four failure modes in every output:
- Made-up facts ("hallucinations"). OpenAI's own help centre says ChatGPT can produce incorrect or misleading answers, including made-up quotes, studies and citations, and can sound confident even when it is wrong. Marketing example: a draft blog post quotes "a 2025 Nielsen study" that does not exist.
- Out-of-date knowledge. A model learns from data up to a cut-off date. Unless it searches the web, and you check what it found, it can miss recent algorithm changes, competitor launches or new rules. Example: it describes a social platform's ad formats as they were two years ago.
- Generic output. With no context, AI falls back on average, safe-sounding phrasing such as "unlock your potential" or "in today's fast-paced world". The copy reads fine but could belong to any brand.
- Missing context. The tool does not know your audience data, brand guidelines, legal position or past campaigns unless you give it them, so it fills the gaps with guesses.
DO this: take any AI output you have used recently. Underline every fact, number, name and date, then confirm each one against a primary source. Circle any phrase a competitor could equally have written.
Task 3: Read the evidence on how marketers use AI
Use dated, sourced figures, and always ask what a figure does not tell you.
Statistic 1: Salesforce, State of Marketing (10th edition), announced 19 February 2026.
- 75% of marketers have adopted AI. The survey covered nearly 4,500 marketing decision makers, with fieldwork from 8 October to 17 November 2025.
- In the same survey, 84% admitted to running generic campaigns.
- What it leaves out: "adopted" does not tell you how deeply AI is used or how good the results are. The respondents were decision makers, not all marketers. The figures are self-reported. Salesforce also sells AI marketing tools.
Statistic 2: European Marketing Agenda 2026 (European Marketing Confederation, supported by CIM, with McKinsey & Company and Lukkien), published 2 February 2026.
- Over 1,800 marketing and sales professionals took part. Only 5% of European respondents made no use of AI, yet only 12% had professionalised and scaled their AI use.
- 45% named lack of expertise as the main barrier to successful AI projects.
- What it leaves out: the figures are European rather than UK-only. They mix sales and marketing staff. They do not measure business outcomes.
Taken together, the two reports show that almost every marketer now uses AI, but few use it well at scale. The gap is skill and judgement, and that is what this course builds.
DO this: for any AI statistic you meet, write down four things: the source, the date, the sample, and one thing it does not measure.
Try it on your own work
- List ten tasks from your last fortnight.
- Sort each one using the five steps in Task 1.
- For each AI-assisted task, write the check you will run.
Checklist
- Every task has a group and a named check
- All facts, numbers and dates are confirmed against a primary source
- Generic phrases are replaced with brand-specific ones
- Any statistic you quote has a source, date, sample and stated limitation
Key takeaways
- Sort tasks into AI-suitable, AI-assisted with human review, and human-only.
- The more a task depends on current facts, public exposure or accountability, the more human involvement it needs.
- The four main failures are made-up facts, out-of-date knowledge, generic output and missing context.
- Salesforce (Feb 2026) reports 75% adoption, and the European Marketing Agenda 2026 reports only 12% professionalised and scaled use. Adoption is high, but mature use is rare.
- Always ask what a statistic leaves out.