Your team has the licences. What most of them are missing is a decision about which recurring jobs Copilot owns.
Five recurring marketing workflows that suit Copilot: the Monday meeting recap, the campaign brief, first-pass copy variants, the performance summary, and the customer-feedback theming pass in Excel. Each one below has the prompt written out and filled in. Two things matter more than the prompts. First, point Copilot at your own files rather than asking it to invent from nothing: Word lets you reference up to 20 files, emails or meetings when drafting. Second, give each workflow a day and an owner, because a prompt that works once and never gets used again has changed nothing.
Say it’s Tuesday. You’ve got a campaign brief due Thursday, three subject lines to write before the send window closes, and a performance recap your director asked for on Friday and will read on her phone. You open Copilot, ask it to tighten up a paragraph in an email, paste the result back, and close the pane.
That’s the whole relationship. It happens again next Tuesday.
I don’t think that’s a skills problem, and I’ve stopped believing it’s a prompting problem either. It’s that nobody on the team ever decided which recurring jobs Copilot owns. Everything else on your week has an owner and a slot. The AI tool has neither, so it gets used in the gaps between real work, which is exactly where the low-value tasks live.
Microsoft’s own usage data points the same direction, and it’s worth reading carefully because it cuts against how most people use the thing. In an analysis of more than 100,000 Copilot chats, Microsoft found 49% of conversations support cognitive work, meaning analysing information, solving problems, evaluating and thinking, while only 17% are producing work[1]. Worth noting that this is Microsoft measuring its own product, so treat it as a signal about what Copilot is good for rather than a neutral finding. Still, the split is the opposite of the drafting-assistant reputation.
So the useful question isn’t “what prompts should I use.” It’s which five jobs come round every week whether you like them or not, and which of those you’d hand to a competent new starter who has read all your files and none of your context.
Here are the five I’d build first for a marketing team. They’re chosen because they recur on a schedule, they start from material you already have, and a human can check the output in under five minutes.
| Workflow | Where it runs | Starts from | What you check before it goes anywhere |
|---|---|---|---|
| Monday meeting recap | Teams | The meeting transcript | That the owner names are right and nothing was invented as an action |
| Campaign brief | Word | Last quarter’s brief, the results deck, the product page | Audience, offer and the success measure, which is where it guesses |
| Copy variants | Word or Copilot Chat | Your approved brief and voice rules | Claims. Every specific number needs a source you can point at |
| Performance summary | Excel | The platform export, unedited | The numbers, against the sheet. Not the sentences |
| Feedback theming | Excel | Reviews, survey free text, support tickets | A sample of rows against the theme it assigned them |
The five recurring workflows described in this article, with the app each runs in and the specific check that matters for each one.
Before you promise any of this to your team, it’s worth knowing what Microsoft currently documents as working, because a lot of what circulates online is a year out of date. I checked these against Microsoft’s own support pages in August 2026.
A few limits worth knowing before you promise anyone anything, because each of these makes people conclude the tool is broken when it isn’t:
One genuinely counter-intuitive thing. Web grounding works on Copilot Chat as well as on the paid Copilot licence. The paid licence’s real advantage is reasoning over your work data, meetings, emails and chats, plus access to the more capable agents[6].
Worth being precise about what Copilot Chat costs, because “free” gets thrown around here and it isn’t quite right. Microsoft’s wording is that it comes “at no additional cost” for users with an eligible Microsoft 365 subscription[6]. So it’s included in something you already pay for rather than free, which matters if anyone on your team is about to build a business case around the word. If your team keeps saying Copilot is useless, check which of the two they’re actually in before you buy anyone anything.
The brief is the job I’d automate first, and not because writing briefs is hard. It’s because the brief is where the whole campaign either gets specific or stays vague, and the reason briefs stay vague is that people write them from memory at 4pm on a Thursday.
Copilot in Word is genuinely good at this for one reason: you can point it at last quarter’s brief, the results deck and the product page all at once, so it starts from your actual campaign history rather than a general idea of what a brief looks like.
| Paste this into a blank Word document with Copilot |
|---|
| Draft a campaign brief for our Q4 retention push, referencing the three files I’ve attached: the Q3 retention brief, the Q3 results deck, and the current pricing page. Use these headings, in this order: Audience, What we want them to do, Offer, Message we’re testing, Channels, Success measure, What we’re not doing. Rules: under 500 words. Use the same audience segment names as the Q3 brief. Where the Q3 results contradict what we assumed in the Q3 brief, say so in one line under the relevant heading. If you don’t have the information for a heading, write MISSING plus the specific question I need to answer, rather than guessing. |
A complete brief prompt with the constraints filled in. Swap the three referenced files and the headings for your own; keep the MISSING instruction, which is the part that does the work.
That last rule is the one I’d fight for. Without it you get a beautifully complete brief in which the success measure has been invented, and invented success measures survive all the way to the retrospective. With it you get a brief that says MISSING: what counted as a retained customer in Q3, 60 days or 90? Which is a question somebody can answer in about a minute.
Tell it what to do when it doesn’t know, or it will fill the gap and you won’t see the seam.
Copy variants work the same way, with one difference. Ask for more than you need. Six subject lines built on the same angle isn’t six options, it’s one option with variations, and you’ll spot that immediately if you make it label them. Ask for four lines, each testing a different angle, with the angle named above each one. Then you’re choosing between ideas instead of choosing between adjectives.
If ChatGPT is what your team already reaches for, our guide to ChatGPT for marketing workflows covers the equivalent moves there. The structure transfers. What doesn’t transfer is the file grounding, which is Copilot’s genuine advantage inside a Microsoft shop.
Everyone’s got the Friday report that keeps arriving even though nobody’s opened it since about March. The reason is almost never the data. It’s that the report describes what happened instead of saying what to do about it, and the person writing it has run out of energy by the time they get to the recommendation.
Copilot in Excel handles the description part well enough that you get your energy back for the recommendation.
| Paste this into Copilot in Excel, with the raw platform export open |
|---|
| Using only the data in this sheet, write a summary for a marketing director who has four minutes. Structure: three bullets on what moved, one paragraph on the most likely explanation, and one decision I need to make next week. Rules: every number you quote must appear in the sheet, and name the column it came from. Flag any week where the change is smaller than the normal week-to-week variation, and say so rather than calling it a trend. If two explanations fit the data equally, give me both. |
A summary prompt for a raw platform export. The column-naming rule is what makes the output checkable in two minutes rather than ten.
The naming rule matters more than it looks. When Copilot tells you cost per acquisition rose 12% in the column labelled CPA_blended, you can check one cell. When it just says cost per acquisition rose 12%, you have to reconstruct where it got that, and in practice nobody does, which is how a wrong number ends up in a board pack.
Check the number, not the sentence. The sentence is always convincing.
The other Excel job worth knowing about is theming free text. If you’ve got a column of survey responses, review text or support tickets, Copilot can summarise it, identify key themes and sentiment, and add the results back as a new column with citations to the source rows[3]. Spot-check ten rows against the theme it assigned them before you use any of it in a deck. Not the ten that look interesting. Ten at random, which is a different and more honest exercise.
Something I’ve watched happen more than once, and it’s a specific kind of disappointing. Somebody on the team finds a prompt that genuinely works. They share it in the channel. Three people react with the fire emoji. Nobody, including the person who wrote it, uses it again.
The prompt wasn’t the problem. It had no day, no owner and no place in anyone’s week, so it competed with everything else on Monday morning and lost.
BCG puts a number on where the value in AI actually sits, based on client work across hundreds of companies: roughly 10% comes from the algorithms, another 20% from the technology needed to implement them, and the remaining 70% from rethinking the people side[7]. To be precise about what that is, it’s BCG’s attribution of where value comes from, not a survey of how anyone spends their budget. But as a way of framing your own Copilot rollout it’s useful: if the licence was the whole plan, you bought the 10%.
We’ve written up the full version of how adoption actually compounds in our piece on taking an organisation from AI awareness to AI fluency, and there’s a marketing-specific version in practical ways to drive AI adoption on a marketing team. I won’t repeat the framework here. The short operational version is the bit that matters this week.
A prompt becomes a workflow the moment it has a day and a named owner. Until then it’s a screenshot.
The complaint I hear most often about AI copy isn’t that it’s wrong. It’s that it’s fine. Grammatically clean, structurally sensible, and it could have come from any company in your category. Everyone’s landing pages have started to rhyme.
Asking for “our brand voice” doesn’t fix this, because the model has no idea what yours is and will substitute a generic professional register that reads like a competent stranger. What does fix it is giving it rules that are specific enough to be broken.
Here’s the block we use. Paste it at the top of any drafting prompt, and edit it until the rules describe things your brand actually does and doesn’t do.
| Paste above any drafting prompt |
|---|
| Write to one person, not an audience. Second person throughout. Contractions always. You’ll, we’ve, doesn’t, isn’t. Never these words: unlock, unleash, supercharge, revolutionise, seamless, robust, leverage as a verb, dive into. Never an em dash. Use a comma or start a new sentence. No sentence fragments for drama. No question as an opening line. Sentences average 16 to 20 words with real variation. Never four short sentences in a row. Every claim about our product names the specific thing it does. If you can’t be specific, cut the claim rather than softening it. If a competitor could put their name on a sentence unchanged, rewrite that sentence. |
A voice block written as testable rules rather than adjectives. The banned-word list is the part to customise first, because it is the part your team will actually notice being violated.
Notice that almost none of these are about tone. They’re about mechanics, because mechanics are checkable and tone isn’t. “Warm but authoritative” gives a model nothing to act on. “Never these eight words” gives it something it can obey and something you can grep for.
The last rule on that list does the heaviest lifting, and it’s worth running as a separate pass. Take the draft, cover your logo, and ask whether your closest competitor could publish it unchanged. If yes, the draft isn’t finished. Our guide to evaluating AI-generated marketing content goes deeper on what to check before anything ships.
None of the above is much use as a list of possibilities. It becomes useful when it’s a timetable, because a timetable survives a busy week and a good idea doesn’t.
Teams recap of the weekly planning meeting turned into owned actions. Owner: whoever ran the meeting. Check: the names.
Campaign brief drafted in Word against last quarter’s brief and results. Owner: campaign lead. Check: the MISSING lines get answered before the brief circulates.
Four copy angles drafted and cut to one. Owner: whoever owns the send. Check: every claim has a source.
Performance summary from the raw export. Owner: the person who currently writes the Friday report. Check: the numbers against the sheet.
Feedback theming pass on the last two weeks of reviews and tickets. Owner: whoever owns customer insight. Check: ten random rows.
A worked example of the five workflows placed in a week with named owners and the specific check each one needs. Total: under two hours across five people.
Under two hours a week, spread across five people, replacing work they were already doing worse and slower. That’s the honest size of the prize here, and I’d rather tell you that than promise you a transformed department.
Pick one row from that table. The Thursday one, if you want the fastest visible win, because the performance summary is the job most people already resent. Run it for two weeks with the check step actually performed, not assumed. Then decide whether it earns its slot before you add a second.
Five recurring jobs suit it well: turning a Teams meeting into owned actions, drafting a campaign brief in Word grounded in last quarter’s brief and results, generating first-pass copy angles, summarising a raw performance export in Excel, and theming customer free text such as reviews and survey responses. What they have in common is that they recur on a schedule, they start from material you already have, and a human can check the output in under five minutes. The tasks that suit Copilot worst are the ones where you can’t quickly tell whether the answer is right.
Start with your voice rules, then the brief, then the constraint that does the real work: ask for four lines, each testing a different angle, with the angle named above each one. Naming the angle forces genuinely different ideas rather than six rewordings of the same one. Add a rule that any specific claim about your product must name the thing it does, and that if it can’t be specific it should cut the claim rather than soften it. The house rules block in the brand voice section of this article is written to be pasted directly above any drafting prompt.
Yes, with conditions worth knowing. Microsoft documents Copilot in Excel returning insights as charts, PivotTables, summaries, trends or outliers, and summarising unstructured text such as survey results or reviews with citations back to the source rows. Editing only works when calculation options are set to automatic, and it fails on files checked out in SharePoint. In PowerPoint you can generate a deck from a prompt or from one Word document at a time, though the controls for adjusting length, tone and style were still rolling out to Insiders rather than generally available when this was written in August 2026.
Give it rules it can obey and you can check, rather than adjectives. A banned-word list, a ban on em dashes, a sentence-length target, a requirement that every product claim names something specific, and a rule that any sentence a competitor could publish unchanged has to be rewritten. Tone instructions like warm but authoritative give a model nothing to act on. Mechanics are testable, which means you can tell whether it followed them and your team can tell when it hasn’t.
The main practical difference is grounding in your own material. Copilot in Word can reference up to 20 files, emails or meetings you already have permission to access in SharePoint or OneDrive, which means a brief can start from last quarter’s actual brief rather than a general idea of what briefs look like. The prompt structures transfer between the two tools almost unchanged. Worth knowing that web grounding works on Copilot Chat as well as on the paid Copilot licence, so if your team says Copilot is useless, check which one they’re actually using before you buy anything. Copilot Chat is not free, despite how often it gets described that way: Microsoft’s wording is that it comes at no additional cost for users who already have an eligible Microsoft 365 subscription.
Every Microsoft capability described here was checked against Microsoft’s own current support documentation in August 2026 rather than against secondary summaries, and anything still in preview or rolling out to Insiders is labelled as such. Two commonly repeated claims were dropped after checking: that Copilot in Excel requires data formatted as an Excel Table, and that Word can only reference three files. Neither appears in Microsoft’s current documentation, and the current file-reference figure is 20. The Copilot usage split is Microsoft’s own first-party telemetry about its own product and is attributed that way in the text. The BCG 10/20/70 figure is BCG’s attribution of where value comes from in its client work, not a survey of budget allocation, and is described that way rather than as a spending statistic.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.