The most common AI enablement programme I have seen inside a marketing team is a Slack message with a link in it. Then a quarter goes by and the CMO asks why AI hasn't shown up in the numbers.
Distributing a tool is not the same as teaching someone to use it, and most marketing teams have only done the first. The structural reason is money: software comes out of a procurement budget and training comes out of a discretionary one. This article looks at what a starter marketing-specific AI training should include: five named workflows, a reusable brand voice asset, and evaluation, which is the skill nobody teaches. It also looks at why untrained leadership makes the gap worse, and gives you an eight-week starter you can run while you make the case for funding it properly. It is deliberately a floor, and the article says exactly where that floor runs out.
The most common AI enablement programme I have seen inside a marketing team is a Slack message with a ChatGPT link in it. Sometimes there’s a deck. Occasionally somebody records a 40-minute demo that four people watch at 1.5x speed.
Then a quarter goes by, the CMO asks why AI hasn’t shown up in the numbers, and everyone looks at the copywriter who has been quietly using it to rewrite subject lines.
Research from NewtonX, surveying 500 marketers for Adweek, put a number on how common that is: 69% said their company mandates or strongly encourages using AI, while 43% said they received no formal training on it[1]. Two thirds told to use it, nearly half handed nothing.
What I’d add from a decade of running marketing teams is that 43% is generous. Formal training in most organisations means a licence, an announcement, and a Notion page nobody updates after the first fortnight.
Access is not adoption. Distributing a tool is not the same as teaching someone to use it.
Before going further, it’s worth knowing where your own team actually sits. Most people assume they’re somewhere in the middle. The audit below usually says otherwise.
Answer yes or no for your team, honestly, before you spend anything. These are not really matters of degree.
| Question | Yes / No |
|---|---|
| Can you name three people who would pass an assessment on the AI tools you pay for? | |
| Is there a written list of what AI must not touch in your team? | |
| Can two people produce on-brand copy from the same brief without rewriting each other? | |
| Does anyone check AI output against a standard, rather than a vibe? | |
| If your best AI user left tomorrow, would anything they built survive? | |
| 4 to 5 yes stage 4, shared capability · 2 to 3 yes stage 3, repeatable workflows · 1 yes stage 2, experimentation · 0 yes stage 1, you have access | |
The count maps to the five stages below. Run it in a team meeting, because the argument it starts is worth more than the total.
The audit gives you a number. This gives you what the number means, because “are we using AI” is the wrong question and it produces a yes from almost every team.
Adoption arrives in stages, and the gap between two of them is where most marketing teams are stuck right now.
Stage two to stage three. That’s the whole game.
| Stage | What it looks like on a Tuesday | What it needs next |
|---|---|---|
| 1. Access | Licences and approved tools. A few people have opened one. | Basic literacy and a written list of what AI must not touch |
| 2. Experimentation | Individuals trying prompts, with private use cases nobody else knows about | Name the workflows out loud, as a group |
| 3. Repeatable workflows | AI used in recurring work, the same job twice with a comparable result | A check step for each workflow, written down |
| 4. Shared capability | Common standards, workflows and checks. Two people produce work you cannot tell apart by author. | A prompt library and a standing review slot |
| 5. AI-powered work | AI embedded into how the team operates. The workflow survives its author going on holiday. | Measurement, and harder problems to point it at |
Match this to your audit count. One yes or fewer is stage 1 or 2, whatever the enthusiasm in the room suggests.
The trap is stage two dressed as stage four.
There are demos. People share prompts in the channel. Somebody built an agent and it works. That looks a lot like a practice, and it can still be five people solving the same problem separately in five private accounts.
The test I’d use is holiday cover. If the person who built your best AI workflow went away for a fortnight, would anyone else run it? If the answer is no, you have capable individuals, which is a genuinely good start and is not the same as a team that has learned something.
Stage three is where the value actually turns up, and it is a smaller step than it looks. It needs one thing that stage two lacks: the workflow written down with its check, so a second person can run it and get a comparable result.
The structural reason has almost nothing to do with anyone being lazy: software is a procurement decision and training is a discretionary one. One has a vendor, a contract and a renewal date. The other has a calendar invite that anybody can move.
The money shows it. Gartner’s 2026 CMO Spend Survey found marketing leaders putting an average of 15.3% of budget into AI initiatives[2]. The CMO Survey out of Duke’s Fuqua School found training budgets sitting at 3.8% of marketing spend[3].
Those come from different surveys measuring slightly different denominators, so it isn’t a clean ratio and I wouldn’t present it as one. Treat it as a direction of travel. The direction is unmistakable: the tool gets funded, the capability to use it does not.
Christine Moorman, who directs The CMO Survey, put the consequence in one line: “Companies will need to ensure that their investments in technology are matched with investments in the capabilities needed to use it effectively”[3].
Budget the training in the same conversation as the licence, or it will not happen.
This is also the moment to notice that adoption isn’t additive. It multiplies, and that changes what a training budget is actually buying.
Tool × Workflows × Behaviour = AI-Powered Professional
| Variable | What it means | The question to ask |
|---|---|---|
| Tool | Approved technology, licences, access | Can people use AI at all? |
| Workflows | The specific recurring jobs AI fits into | Do they know where to use it? |
| Behaviour | Repeated use, evaluation, judgement, safe-use habits | Are they actually working differently? |
Multiplication, not addition. Any variable at zero takes the whole result to zero, and the invoice still arrives.
The licence buys you the first variable. That’s it, and it’s the one every company gets right, because it has a vendor and a renewal date and someone whose job is to sign it.
The other two are what training is for. Workflows are the specific recurring jobs AI belongs in, which is knowledge about your business rather than about the tool. Behaviour is whether people actually work that way when the quarter gets busy, which is a habit, and habits need practice and reinforcement rather than a demo.
Run your own team through it. If workflows are near zero because nobody has named where AI fits, buying a better tool multiplies against a zero and you get the same nothing you had before, at a higher monthly cost. That is the mechanism behind most disappointing AI rollouts, and it isn’t a technology problem at any point.
We’ve written the longer version of that argument in our guide to driving AI adoption on a marketing team.
Whenever training gets cut, it gets cut for one of three reasons. All three sound sensible in the room.
They’re easier to open. They are not easier to get good output from. The gap between an average prompt and a genuinely useful one is wider in marketing than in most functions, because our work is judged on voice, positioning and legal exposure rather than on whether the code compiles.
Some will, and they’ll do it privately with their own accounts and their own workarounds. Microsoft and LinkedIn’s Work Trend Index found most AI users were bringing their own tools to work while only a minority had received any training from their employer[4]. That data is a couple of years old and adoption has moved, but the shape holds: people self-teach, employers stay absent, and nothing that gets learned is ever shared.
There is no after. The next campaign is already late, and it will be late again in six weeks. This is the single most reliable way a training plan dies, and it dies the same way every time: nobody kills it, it just never gets scheduled.
There is no after. The next campaign is always already late.
One question worth sitting with before the next budget conversation. If your team had to sit an assessment tomorrow on the AI tools your company already pays for, who would pass? Not who would say they’d pass. Who actually would.
Generic AI training often misses marketers for a simple reason: it teaches prompting as a skill in isolation, and marketing work is never in isolation. Every output carries a brand voice constraint, a legal constraint and a channel constraint at the same time.
So don’t teach the tool. Teach the five jobs your team actually does every week, and build one repeatable prompt and one check step for each.
| Workflow | What AI does | What the human still owns | Check before it ships |
|---|---|---|---|
| Weekly performance summary | Reads the export and surfaces what changed | Verifies the numbers, adds the business context, decides the action | Does every number match the source file? |
| Creative brief | Builds a first structure from your last five briefs and the audience definition | Owns the strategy, the positioning and what gets cut | Would the person receiving this know what to make? |
| Ad variant set | Generates variants inside your voice asset and channel limits | Picks the idea, tests the claim, signs off the copy | Do the claims survive legal, and do the lengths actually fit? |
| Customer research synthesis | Finds themes across raw feedback and attaches the quotes | Decides which themes deserve action and which are noise | Can you trace each theme back to real verbatims? |
| Landing page QA | Compares the page against the offer, the ad and the brand guide | Judges which mismatches actually matter | Did it catch the thing you already knew was wrong? |
Pick your own five if these are not your five. The point is that there are five, they are named, and each has a check.
Notice the last column. A prompt without a check is a faster way to ship something wrong, and the check is the part that teams skip because it feels like admin rather than skill.
Two of those rows lean on a voice asset, so build that first. It’s one document holding your tone rules, your banned words, three examples of copy that sounds right and three that don’t, and the reason each one fails. Paste it into every prompt that produces public copy. It takes an afternoon, it’s the single highest-leverage thing on this list, and it’s the difference between output you edit and output you rewrite.
This is the other half of the same conversation, and it belongs in writing before anyone gets creative. For most marketing teams the list starts here:
Yours may differ, and it should be argued about rather than handed down. What matters is that it exists somewhere other than in one cautious person’s head, because that person will be on holiday the week it matters.
There’s a compliance dimension too, and it’s now a legal one in places. Article 4 of the EU AI Act has applied since February 2025 and requires providers and deployers to take measures ensuring “a sufficient level of AI literacy” among staff operating AI systems on their behalf[5]. If you sell into the EU, literacy is an obligation rather than a nice-to-have.
Teach the five workflows you already run, not the tool you just bought.
Marketers are good at judging finished work and surprisingly bad at judging AI output. The reason is mechanical: AI output is fluent enough to bypass the part of your brain that notices problems. A weak answer and a strong one arrive with the same clean structure, the same confident tone, the same tidy bullet points.
Which is why evaluation has to be taught on purpose, as its own session, with real examples. It is the one skill on this list that nobody picks up by using the tool more.
Which means evaluation has to be taught explicitly, as its own session, with real examples. Four questions, applied to every output before it goes anywhere:
Run this as a live session rather than a document. Bring three real outputs, one good, one plausible but wrong, one obviously wrong, and have the team score them silently before anyone speaks.
The obviously wrong one is easy and everybody catches it, which is reassuring and teaches nothing. The good one is also easy. The whole lesson sits in the middle example, where half the room will pass something that contains a confidently invented statistic or a claim your legal team would never sign off. When people see colleagues they respect wave it through, the point lands harder than any slide about hallucination ever will.
Do this quarterly rather than once. The failure mode isn’t that people never learn to evaluate, it’s that they get good at it, then get busy, and stop.
Most AI training I’ve watched put the whole department in one room with one set of examples. It’s efficient to schedule and it teaches the CMO and the junior copywriter the same thing, which serves neither of them.
They aren’t doing the same job.
The copywriter needs to get genuinely good at voice, iteration and knowing when a draft is finished. The CMO needs enough fluency to tell whether a workflow is worth the licence. Those are different skills and they take different amounts of time to build.
| Who | What they actually need | How you know it landed |
|---|---|---|
| CMO and senior leaders | Enough hands-on use to judge a workflow, a vendor demo and a disagreement | They can referee whether an output is good enough to ship |
| Marketing managers | Workflow design, the check steps, and coaching their team through weak output | They notice a bad AI habit before it spreads |
| Practitioners | The five workflows, the voice asset, evaluation, and safe-use boundaries | They stop rewriting from scratch and start editing |
| Your one power user | Agents, automation, multi-step chains, and how to hand work over | What they build gets used by someone who did not build it |
Same foundations for everyone on safe use and evaluation. Different examples, different depth, different room.
The foundations should be identical. Everyone learns what must never go into these tools, everyone learns that fluent output can be wrong, everyone learns the check step. Then the examples split by job.
The one I’d protect hardest is the fourth row. Every team has a person who runs ahead, and the usual mistake is to leave them alone because they’re doing fine. They are, and their work is trapped inside their own account.
Give them the harder training and one explicit job: make one thing you built usable by someone who did not build it.
I used to think this was a bottom-up problem. Train the doers and leadership would catch up by osmosis. I don’t believe that anymore.
Gartner surveyed senior marketing leaders across North America and Europe and found 65% expect AI to dramatically change the CMO role within two years, while only 32% said significant changes are needed to the CMO profile and skill set[6]. Gartner calls the space between those two numbers an AI blind spot, and predicts a lack of AI literacy will become one of the top reasons CMOs are replaced at large enterprises[6].
The detail that made me sit up is that Gartner published a nearly identical figure two years earlier. Same expectation of disruption, same reluctance to change the skill set. Two years, no movement on the half that requires anyone to learn something.
The practical consequence is not abstract. A leader who hasn’t used the tools cannot tell a good AI workflow from an expensive one, cannot judge whether a vendor demo is impressive or ordinary, and cannot referee the argument when two people on the team disagree about whether an output is good enough to ship. That referee job doesn’t delegate.
If you lead a team, the fastest fix is to do the five workflows yourself, badly, once. An afternoon is enough to stop being the person who approves things they don’t understand.
The objection to everything above is always cost, so let’s deal with it directly rather than pretending it away.
You should budget for this. That’s the whole argument of the piece and I’m not going to undercut it in the last third. But budget conversations slip, and a quarter spent waiting for one is a quarter where nothing improves and you arrive at the meeting with nothing to show.
So run this in the meantime. You can start some of it internally, and its real output is not a trained team. It’s evidence: a specific list of what your team can’t do yet, which is the strongest thing you can walk into a funding conversation holding.
The audit. Everyone writes down the three AI tasks they already do and the three they avoid. You will learn more from the avoid list.
Pick your five workflows as a group. Argue about them. The argument is the training.
One prompt per week. One workflow, one person demonstrating live, everyone else stealing it.
The evaluation session. Bring three real outputs: one good, one plausible but wrong, one obviously wrong. Score them as a team.
A standing slot. One thing that worked, one thing that did not. This is the part that makes it stick.
Five and a half hours of scheduled sessions, plus the fortnightly slot. Call it six, which is roughly the threshold BCG measured.
The dosage matters more than the format. BCG’s research on AI at work found regular usage climbs sharply once people have had at least five hours of training with access to in-person sessions and coaching, and that only about a third of employees say they’ve been properly trained at all[7].
Five hours. Not a certification, not a six-week programme. Most marketing teams could find that in a quarter if anyone scheduled it, and the reason nobody does is that five hours of eight people’s time has a visible cost while quiet incompetence does not.
Now the ceiling, because starting internally has one and it arrives sooner than people expect. Read what BCG actually measured: five hours with access to in-person sessions and coaching[7]. Running it yourself gets you the hours and not the second half of that sentence.
On the ladder from earlier, this plan reliably moves a team from stage two to stage three. The harder part is turning scattered experimentation into shared capability, and that is where structured training, facilitation and role-specific practice start earning their value.
Five signals that you’re at that point:
That last one is the most common and the least discussed. It’s also the point where the cost of not training properly stops being abstract and starts showing up in someone’s calendar.
What good external training buys you at that point is mostly time. A facilitator who has done this with other teams can spot a weak workflow before you’ve spent a month on it, bring patterns that have already been tested somewhere else, and challenge the assumptions a team doesn’t know it’s holding. That shortens the trial-and-error cycle, which is the expensive part of learning this internally.
If that’s where you are, our corporate AI training is built on exactly these workflows rather than on a tour of the tools.
If this works and you want to make it durable, the natural next step is turning those weekly prompts into something shared rather than something five people each rebuilt. Our guide to building an AI prompt library for your team covers that part.
You can run an excellent session. People leave with workflows that work, prompts that are genuinely better and a lot more confidence than they walked in with.
Then Monday happens.
A month later some of it has survived and some of it hasn’t, and that isn’t a failed workshop. That’s how learning works when it meets a busy quarter. The mistake is treating the session as the whole intervention, when it’s the first of five moves.
Training starts the behaviour change. Steps three to five are what stop it evaporating by week six.
Most training stops after step two, which is why most training doesn’t stick. Apply is where people hit the problems that teach them something, and if there’s nowhere to bring those problems back to, they quietly conclude the tool doesn’t work for their job and go back to doing it by hand.
Steps three to five are also the cheapest part. Fifteen minutes a fortnight covers review and improve. Share is a document somebody owns. None of that needs a budget, and all of it needs someone to actually schedule it, which is the part that fails.
The instinct at the end of something like this is to send round a satisfaction survey.
Resist it.
You were trying to change how the work gets done, so measure the work.
Usage data is the tempting substitute, because it arrives on its own and it looks like evidence. “Forty-one sessions this month” is not a result. “The weekly performance summary takes twenty minutes instead of two hours, and the numbers are checked” is.
Ask the team, not the tool. Run it at 30, 60 and 90 days and watch the direction rather than the total.
| Question | Score |
|---|---|
| I know which parts of my job AI is genuinely useful for | /5 |
| I can run our five workflows without asking anyone for the prompt | /5 |
| I check output against the standard rather than reading it and nodding | /5 |
| I know what must never go into these tools, without looking it up | /5 |
| At least one recurring job is genuinely better than it was in week one | /5 |
| 20 to 25 it stuck · 12 to 19 it is sticking in patches, go back to the weak workflow · below 12 you trained the tool, not the work | |
Then one open question, which is worth more than the scores: what can you do now that you could not do in week one?
Then look past the people to the workflow itself. Four things worth watching, and none of them need a dashboard.
Survival is the one I’d watch closest, because abandonment is nearly always silent. Nobody announces they’ve gone back to doing it by hand.
Then the question your finance director will actually ask, which is what any of this moved. Worth deciding the answer before you spend, not after.
| Learning focus | Change in the work | Business signal |
|---|---|---|
| Context and prompting | Stronger first-pass output | Less rework |
| Evaluation | Fewer weak or unsupported claims | Higher quality, lower risk |
| Workflow training | Recurring jobs finish faster | Shorter cycle time |
| Shared methods | Less dependence on your power user | Consistency across the team |
| Leadership training | Better calls on where AI goes next | Better prioritisation |
Pick the row you are buying before you buy it. Training that is not aimed at one of these is a workshop, not an investment.
Notice that none of the right-hand column is “people used AI more”. Usage is the thing that’s easy to count, and it’s a proxy for a proxy. Rework, cycle time and consistency are the ones a business recognises without translation.
Not a transformation programme. One meeting.
Put 30 minutes in the diary with your team. Run the five-minute audit from earlier and total the scores out loud, which is uncomfortable and exactly the point. Then ask everyone to name the one task they’d most like to stop doing by hand.
That list is your curriculum. Pick the workflow that appears most often, build one prompt for it together, agree the check step, and run it for a fortnight before you build anything else.
If you get to the end of that fortnight and the prompt is still being used by more than the person who wrote it, you have the beginnings of a practice rather than a licence. If it isn’t, you’ve learned something cheap and useful about where the real blocker sits, which is usually the workflow rather than the tool.
If it stalls, resist the instinct to run more training. In almost every case I’ve seen, a prompt that gets abandoned wasn’t a prompting failure: the workflow underneath it was never quite agreed, so two people were solving slightly different problems and neither output looked right to the other. Go back to the workflow, name it properly, then try again.
Either way you’ll know more than you did on Friday, and it cost you half an hour.
Less than most people assume, and more than almost anyone schedules. BCG’s research on AI at work found regular usage climbs sharply once someone has had at least five hours of training with access to in-person sessions and coaching. Five hours is achievable inside a single quarter for most teams. What matters more than the total is the shape: short repeated sessions tied to real work beat one long workshop, because the thing you’re building is a habit rather than a body of knowledge. A team that does thirty minutes a week for six weeks will outperform one that sat through a full day and went back to their inbox.
Start by naming the five workflows your team genuinely runs each week, then build one repeatable prompt and one check step for each. For most teams that’s the weekly performance summary, the creative brief, the ad variant set, the customer research synthesis and the landing page QA. Alongside those, teach two things generic courses skip: how to build a reusable brand voice asset so output is consistent, and how to evaluate AI output against a standard rather than a feeling. Also write down what AI must not touch, which usually includes unreleased pricing, anything customer-identifiable and the first draft of a crisis statement.
Start internally over eight weeks, on about six hours of team time, and treat it as the case-building exercise rather than the training itself. Week one, an audit where everyone lists the three AI tasks they already do and the three they avoid. Week two, pick your five workflows as a group and argue about them, because the argument is the training. Weeks three to six, thirty minutes a week building one prompt, with one person demonstrating live. Week seven, an evaluation session using three real outputs: one good, one plausible but wrong, one obviously wrong. Then a fifteen-minute fortnightly slot to keep it alive. That reliably gets a team from individual experimentation to repeatable workflows, and it produces a specific list of what you still cannot do, which is what a funding conversation actually needs. Turning that into shared capability across roles is the part that usually needs structured training and facilitation.
Because software is a procurement decision and training is a discretionary one. The licence has a vendor, a contract and a renewal date; the training has a calendar invite anyone can move. The spending pattern reflects it: Gartner found marketing leaders putting around 15.3% of budget into AI initiatives while The CMO Survey found training sitting at 3.8% of marketing spend. Those are different surveys with different denominators so it isn’t a clean comparison, but the direction is clear. The practical fix is to budget the training in the same conversation as the licence, because afterwards it competes with everything else and loses.
No, and this is where most programmes waste their budget. Practitioners need the five workflows, the brand voice asset, evaluation and the safe-use boundaries. Managers need workflow design, the check steps, and how to coach someone through weak output. Your one power user needs agents and automation, plus an explicit job: make one thing they built usable by someone who did not build it. Leaders need the least volume and it matters the most, because Gartner found 65% of senior marketing leaders expect AI to dramatically change the CMO role within two years while only 32% thought significant changes were needed to the CMO skill set, and published a nearly identical split two years earlier. A leader who hasn’t used the tools can’t tell a good workflow from an expensive one or referee whether an output is good enough to ship, and that job doesn’t delegate. The foundations stay identical for everyone, which is safe use and evaluation. The examples and the depth should not.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.