Does AI Save Marketers Time?

01What I looked at

The studies cover writing, consulting, product ideas, office work, coding, and online retail. I chose experiments with a comparison group and enough published detail to check the findings. Some involve marketers; others help explain what to watch for when measuring AI at work.

This is a research review. I haven’t run a new experiment or combined the studies into one average. The source notes explain the selection and versions used, including revisions published through 2026.

A study of someone writing with an older ChatGPT model doesn’t tell us how well ChatGPT Work handles a campaign today. I’m using that research to ask better questions about these tools, without treating it as a current product test.

02Where Claude Cowork ChatGPT Work Copilot and Gemini fit

I’m looking at the work around a campaign: reading the brief, finding information, preparing reports, and getting files into a form the team can use. These four tools approach that work differently.

The features come from Anthropic, OpenAI, Microsoft, and Google. The test ideas are mine. Check which features your plan and connected accounts allow.

Anthropic announced on September 16, 2026 that Cowork and chat were merging into one Claude experience, starting with Pro and Max. I’m using the Cowork name here for those task features, even if your app has already changed. Anthropic’s announcement.

I’d pay attention to how much moving things around each tool saves me. If I still have to copy the answer into a document, rebuild the formatting, and add the sources, the job isn’t finished.

03What I would ask Claude Cowork and ChatGPT Work to do

Both companies describe tools that can take an assignment and produce work for review. Claude can use chosen folders and tools; ChatGPT Work can use files, plugins, and approved tools. Claude Cowork, ChatGPT Work.

I’d give them a whole task: use these customer notes to create a campaign brief, show where the claims came from, and return a document I can edit.

Then I’d ask for a revision. That’s where I’d want to see whether the tool kept the requirements straight or made me explain the brief again. I’d also check the sources myself and open the file in the app where the team will use it.

The studies below don’t tell us which of these current products would win that test. They do give us reasons to look beyond how good the first answer sounds.

04Check the tools already in your work apps

Microsoft offers Copilot inside its office apps. Google offers Gemini in Workspace, with automation through Workspace Studio. The exact app and features matter here; “we have Copilot” doesn’t tell you everything a person can access. Microsoft’s overview, Google’s overview.

Before paying for another tool, I’d try the same assignment in whatever the team already has. A tool that works inside the document you’re editing might save enough handling to be useful, even if it isn’t the app you’d open for everything.

The Copilot study below covers one tested setup. Its results don’t tell us what every Copilot feature can do now or how Gemini would perform.

05What the six studies found

These studies measured different things. Less time writing an email and more sales on a retail platform can both matter, but averaging them into one “AI productivity boost” would be misleading. The study IDs link to the research notes at the end.

Sources: S1 MIT, S2 published paper, S3 published paper, S4 revised paper, S5 paper, S6 revised paper.

06Writing was faster but the test had limits

Noy and Zhang’s writing experiment included marketers among the participants. They completed short assignments related to their jobs, with one group given ChatGPT access. Evaluators scored the work without knowing who had used AI. S1 institutional report.

There’s a catch I’d want marketers to see alongside the result. The assignments didn’t require detailed company knowledge or precise factual accuracy, and the outputs weren’t independently fact-checked. That leaves out work I’d have to do before publishing a sourced article or making claims for a client.

I’d take this as a good reason to try AI on a draft with clear instructions. I’d still count the time spent checking facts and getting approval before claiming the same saving on a real campaign.

07Good campaign ideas and good decisions are separate tests

In the BCG experiment, GPT-4 helped consultants on tasks the model could handle well. On a difficult managerial task it struggled with, people using AI were less likely to get the answer right. S2 published study.

That matters when one brief asks you to come up with campaign ideas and choose an audience. A tool can be useful for the ideas and still miss something in the customer evidence that should change the recommendation.

I’d check the recommendation separately from the writing. A deck can be clear and convincing with the wrong conclusion in it.

08What happened at Procter and Gamble

P&G professionals worked on product ideas alone or in pairs, with or without AI. Individuals using AI matched the performance of pairs without it. Their proposals also combined technical and commercial thinking more evenly. S3 published study.

I’d try that when preparing a proposal: ask what I’m missing from the product or commercial side, then check those suggestions with the people who know the business. Matching a team’s submission in this exercise doesn’t mean replacing someone’s whole job.

09Less email time can be useful on its own

In the revised office-work study, giving workers Copilot access reduced email time by 1.4 hours a week on average. The researchers estimated about two hours for people whose use changed because they were given access. Those figures answer different questions; two hours isn’t the average saving for everyone. The study didn’t find a meaningful change in the amount or mix of tasks completed. S4 revised paper.

I’d be happy to spend less time in my inbox without producing another campaign. But if the goal is more finished work, we need to measure that too. Decide which result you want before you start counting.

10People can feel faster while taking longer

The developers in METR’s early-2025 experiment thought AI had helped them finish faster. The timer showed they’d taken longer. It was a small group of experienced developers working in familiar projects, so I wouldn’t apply the percentage to marketing. S5 paper.

The follow-up needs to be included too. In February 2026, METR said problems with who participated, which tasks they submitted, and how time was recorded made its newer results unreliable as an estimate of current effects. Measuring people working with several agents at once was also difficult. METR follow-up.

For me, this is a reason to keep a timer running. Watching a tool work can feel productive even when I’m spending too much time helping it along.

11More sales and less work are different results

The retail paper tested seven uses of AI, including advertising and customer service. The reported sales gains came mainly from more shoppers converting, rather than buying larger baskets. These were experiments inside one retail platform, so the findings don’t measure someone’s workday in Cowork or ChatGPT Work. S6 revised study.

I’d track campaign performance alongside the time spent making it. Finishing sooner is useful, but I’d want to know if the campaign did worse. Equally, better results might justify the same amount of work.

12The marketing tasks I would test first

I’d start with a task where I can check what’s right and wrong. These are the assignments I’d use for a trial, not a list of proven winners.

For competitor research, I’d check the sources before reading the recommendation. If a basic product claim is wrong, I want to catch that before it becomes the argument for our strategy.

13Keep counting after the first draft

I’d record two times: the minutes I spend working on the task, and the time until the finished file is ready. If a tool takes half an hour while I do something else, that’s different from half an hour of me correcting it.

Write down what the output has to include before starting. If it misses a deadline or invents a claim, it hasn’t passed. If I fix it myself, those minutes go in the total.

For several tasks, divide all the human time spent by the number of outputs that passed review. Include failed attempts in the time total, and show how many passed. Otherwise, a tool could fail most assignments and still look good because you only reported its best one.

Keep setup time separate so you can see whether the second attempt gets easier. Record usage costs too, including any extra tools needed to finish the job.

14What I would use this research for

There’s enough here that I’d try AI on drafting and routine office work. I’d keep the first few assignments small enough to check properly and see whether they took less of my time.

For Claude Cowork, ChatGPT Work, Copilot, or Gemini, I’d start with one recurring job and keep the outputs. After a few attempts, I’d want to be able to point to what was finished, what needed fixing, and how much time I spent on it.

If I’m still doing most of the job after the tool says it’s done, that’s the result I’d report.

15Frequently asked questions

Does AI save marketers time?

They can help, but these studies don’t give us a saving that applies to every marketer or current tool. Time the assignment through final review and see whether the result holds up.

Is generating more content proof of productivity?

It tells you how much content was made. To know whether the process improved, check quality, time, cost, and what happened when the content reached its audience.

Does this research show whether Claude Cowork or ChatGPT Work is better?

No. The studies tested specific tools and assignments, often with older models. They don’t provide a head-to-head comparison of the current products.

16Research record and source notes

The search covered writing productivity, marketing and consulting experiments, AI teamwork, office time use, developer productivity, and retail experiments. It was a selective search completed October 16, 2026, not a systematic review of every study. Vendor adoption surveys and testimonials were excluded as evidence of cause and effect. The figures come from the linked papers and institutional reports; participant-level data wasn’t reanalyzed. Research was first released in 2023–2025, with revisions and journal versions checked through 2026. Publication dates should not be mistaken for tool-testing dates.

S1 — Noy and Zhang, Experimental evidence on the productivity effects of generative artificial intelligence, Science, 2023. DOI 10.1126/science.adh2586. The journal page could not be retrieved; the detailed MIT report and study abstract reproduced by Stanford were consulted. The quality score is an evaluation outcome, not a conversion metric. MIT report, Stanford abstract.

S2 — Dell’Acqua and colleagues, Navigating the Jagged Technological Frontier, Organization Science, published online March 11, 2026. DOI 10.1287/orsc.2025.21838. Full published text and HBS’s 2026 discussion were consulted. Older institutional summaries give different quality-effect figures; this article omits a numerical quality effect rather than mixing versions. Published text, HBS discussion.

S3 — Dell’Acqua and colleagues, The Cybernetic Teammate, Organization Science, published online June 12, 2026. DOI 10.1287/orsc.2025.20702. The published paper reports 791 workshop participants and complete data, including post-task surveys, from 776. The paper also excludes 35 nonrandomly assigned participants from its main analyses. P&G participated, and the paper includes company-affiliated authors and support disclosures. Published text.

S4 — Dillon, Jaffe, Immorlica, and Stanton, Shifting Work Patterns with Generative AI, arXiv 2504.11436v4, revised November 13, 2025. Full revised text consulted. Earlier abstract figures were superseded. Outcomes use application activity, which does not observe every form of work. The study was conducted with Microsoft involvement and participating firms interested in early adoption. Versioned paper.

S5 — Becker, Rush, Barnes, and Rein, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, arXiv 2507.09089v2, July 25, 2025. Paper abstract and METR’s February 24, 2026 methodological update consulted. The original effect’s reported confidence interval is a 2%–39% increase in completion time. Do not treat the later, selection-affected estimates as a clean current comparison. Paper record, follow-up.

S6 — Fang, Yuan, Zhang, Donati, and Sarvary, Generative AI and Sales Productivity, arXiv 2510.12049v6, revised June 29, 2026. Full revised text consulted. Experiments cover 2023–2024 at one cross-border retailer. The upper-end sales effect is not an average across all workflows; the platform setting limits generalization. Versioned paper.

No combined effect was calculated, and confidence intervals aren’t reproduced for every finding. The outcomes are too different to read the table as a ranking of tools or occupations.

Further reading

ChatGPT Dot vs Grok Bot vs Meta Muse in 2026

How to Tell When AI Is Making Things Up