A new AI model comes out and suddenly your feed is full of demos and people telling you how much has changed. For a marketing team, it’s not always clear what you’re supposed to do with that information. Should you switch models? Change how you brief it? Can it do work the old one couldn’t, or will it just do the same work better?
Often, the response is fairly random. Someone tries it on a task and says it’s excellent. Someone else tries something different and can’t see much improvement. Neither gives the rest of the team much guidance on what to change.
Models will keep evolving, so marketing teams need a repeatable way to put a new one through its paces. If your team uses subscriptions rather than paying for AI through an API, comparing the price of each task is less useful. There are two practical things you want to learn:
- What work can you now do with AI? Find out where results have improved and whether a new model can handle something you previously couldn’t get to work.
- How far can you stretch your subscription? Find which models do routine work well enough, so you can save your allowance for a more capable model when it makes a useful difference.
For B2B enterprise marketing teams, the activities below give you a way to answer both. Choose a few that matter to your team, including something you’ve previously struggled to do with AI.
Set up a comparison you can use again
Save the brief, files and instructions for each activity, then give each model the same material in a fresh conversation. Keep any follow-up instructions, too, so you can compare how much help each needed.
Record the application, model, settings and tools you’re using. If one can browse the web and another can’t, remember that you’re assessing that difference as well as the models.
Research and analysis
1. Research a target company
Pick a company you’re interested in and ask for a report on its business, customers, locations, recent news and relevant people. Explain why you’re researching it, and ask for dated sources and a note of anything it couldn’t find.
When you review the report, check that it has the right company and that its sources support the details. Did it find what you needed, finish without repeated help and present something you could use?
2. Qualify contacts
Give the model some contact profiles and explain who you’re trying to reach. Ask it to recommend contacts, explain its choices and flag missing information.
Review the choices with someone who knows those accounts. Does it consider people’s responsibilities, business units and locations, or rely on job titles that can mean different things at different companies?
3. Combine CRM reports into a dashboard
Take a few related CRM reports you’re familiar with and ask the model to combine them into a dashboard. Explain your metrics and the questions you want it to answer, including any filters you’d like to use.
Check that the totals match the original reports and that combining them hasn’t duplicated or lost records. Has the dashboard revealed useful connections between the reports, and can you understand it and use its filters?
Writing and content
4. Develop a content plan and draft
Give the model an audience, an objective and source material, then ask it to develop a content plan, check its claims, outline and draft. Keep each stage so you can see where it did well or went off course.
Ask your content lead to assess whether the idea would interest the audience and how close the draft is to something they’d publish.
5. Rewrite to a specification
You could take a technical article and ask the model to rewrite it for a marketing director. Give it your style guidance and a length to work within, making clear which claims and examples need to stay.
As you read the result, check whether it has made the explanation easier to follow without changing its meaning. Read it aloud, too. Where does it sound awkward or unlike your company?
6. Adapt content for social media
Give the model an article or webinar transcript and ask for a LinkedIn post aimed at a particular buyer audience. Explain the tone and length you’re after.
Does the opening give that reader a reason to keep going? Check whether the post makes sense on its own and stays faithful to the source, without losing the explanation that made the original interesting.
Images and video
For images and video, start with the material your team actually produces. You might want to update an old product photograph, create a campaign image or make a short video showing a product in use. Give the model your existing images and brand guidance, then assess both the quality of the result and how accurately it represents your product. An image can look good while getting the logo or proportions wrong.
7. Create a campaign visual
Give the model your brand assets and a brief for a campaign image. Explain the composition, colours, wording and dimensions, including where you need room for a headline.
Check whether it has followed those instructions, then look at the image at the size you intend to use. Are the words correct and readable? Even if it has included everything you asked for, does the design work?
8. Replace a product in a scene
Take an existing product photograph and provide reference images of a different product. Ask the model to replace it while keeping the surroundings the same.
Look closely at the new product’s proportions, logo and distinctive details. Then check whether it fits the scene: are the scale, perspective, shadows and reflections convincing? You need both a faithful product image and a believable setting.
9. Update a product logo
Give the model a product photograph and your approved new logo, showing where it should go. Ask it to leave everything else unchanged.
Check the lettering, colour and position, including whether the logo follows the surface it’s on. It’s also worth looking over the rest of the image for unrelated changes or traces of the old logo.
10. Update photographic lighting
If your product photography has changed style, try giving the model an older photograph and a newer one as a reference. Ask it to match the lighting while preserving the product and composition.
Check that the product’s colour, finish and shape are still accurate. Whoever normally approves photography can judge how well it matches the new style and how close it comes to the quality they’d accept from a shoot.
11. Make a product video
Provide reference images and a simple storyboard showing a problem, your product in use and the result. Tell the model how long the video should be and what captions to include.
Watch whether the sequence makes sense and the product stays consistent throughout. Alongside the overall quality, look for invented features or actions the real product couldn’t perform.
Build something a buyer could use
12. Create a calculator
You could also try a small interactive calculator that helps a buyer work something out. This draws on the mini-tool suggestion in Tom Orbach’s growth ideas.
Give the model the formula and examples with known answers. Check its calculations, including what happens with blank fields or unusual inputs, and ask whether a buyer could understand the result without help. Keep it as a prototype for this assessment; finding out whether buyers would use it is a separate exercise.
Record why a result is good enough
Keep a shared spreadsheet of each run, with the date, model, settings, inputs and results. Agree how you’ll rate completion, accuracy and presentation, and record why you gave each score. If you wouldn’t publish a draft, explain what would need to change.
Ideally, you’d have a human-produced example to compare against. A different model can also audit a research report by opening sources and checking claims. Review its findings yourself, including whether the report missed anything you needed. Correct citations don’t mean the research is complete.
Try several examples and repeat a test before recommending a model, so you’re judging more than one impressive result. Date online research, since the sources can change.
Turn the findings into model choices
You might find that a cheaper model handles routine rewrites well, while a stronger one gives you a better content plan. You might also discover that a task you’d given up on now works well enough to use.
Those findings can help you get more from your subscription. More expensive models can use up your allowance faster, although the rules vary by plan; Claude’s guidance, for example, explains how model choice affects usage. Keep any usage information your plan provides alongside the results.
Have someone keep the tests current and share what you learn, so it carries into the next project. That’s the discipline behind You Can’t Learn AI the Way You Learn Software.
Rerun the tests when a new model arrives. If all models now do well, add harder examples while keeping earlier results and inputs for comparison. Mike Taylor’s personal-benchmark guide has more on organising these tests.
You don’t need to build the whole collection at once. Start with one routine activity and one thing you’ve struggled to do with AI. Agree what a good result would look like, then see how your current model and the new one compare.