Marketing workflows are changing

Picture a campaign that should have gone out yesterday. The copy is ready, but the images are in another folder and two people are checking different versions. The work is waiting for someone to put it all together.

GPT-6 and comparable models can carry out much of that work when connected to the right tools. For a well-defined workflow, AI can handle production tasks that previously needed several people. GPT-6 Astra, for example, supports research, document creation and tool use.

The workflow still needs someone who understands the customer, sets a clear goal and can judge the result. A polished campaign can make the wrong promise or reach the wrong audience. Experience matters because someone must notice these problems and know how to fix them.

Start with a clear brief

Ask AI to “create some content about our product” and it may return a polished page before you have decided who should read it. That is a good way to produce work you will have to rewrite.

Start with the audience, what they should learn or do, and the facts that support the message. Then specify the format, tone and call to action.

Give the workflow reliable source material: product information, research, customer interviews and approved examples. Keep shared guidance, such as brand rules, separate from the brief for each campaign. Update both when something changes.

Set a clear standard for completion. A draft is only one step. The finished work must be accurate, fit its channel and be ready for review. If a claim has no evidence, the system should flag it.

Connect the steps and tools

A good draft sitting in a chat window is still unfinished work. If you have to copy it into several tools and fix the formatting each time, much of the job is still yours.

The model needs access to source files and the tools used to produce the work. APIs and connectors can provide that access. With function calling, the model requests an action and the connected software carries it out.

Use a clear sequence: read the sources → produce the content → run checks → publish or request review → record the result. Save progress outside the chat so an interrupted run can resume.

Pass information between tools in defined fields, such as source_version, asset_url and review_status. Structured outputs help keep the format consistent. They do not prove that the content is correct.

Examples of fully automated workflows

These are possible setups once the sources, templates and publishing rules are agreed. Routine runs can finish without a person moving files or approving every step. New claims, missing information or failed checks go to a reviewer.

1. A weekly newsletter

Friday arrives, and the newsletter is still on the to-do list. A scheduled workflow can handle the regular edition without another round of chasing files.

Trigger: a weekly schedule reads an approved feed of product updates.

Automatic steps: select new items → write summaries → fill the email template → check links and layout → send to the agreed subscriber group → record delivery results.

Output: a sent newsletter and a short run report. A change to the audience or an unsupported claim stops the send for review.

2. Short videos from approved content

The script is approved, yet the video is waiting for captions, resizing and export. Those finishing steps can become a workflow of their own.

Trigger: an approved script and its images or recordings arrive in a folder.

Automatic steps: assemble scenes in a fixed template → generate captions and voice-over → export the required sizes → check duration and file quality → schedule publication.

Output: captioned videos ready for each channel, with publication records. Separate media tools produce the video and audio; GPT-6 Astra coordinates the work. Missing footage or a failed check stops that run.

3. A case study across several channels

Getting a customer story approved can take weeks. Once it is ready, turning it into different formats should not become another project.

Trigger: a case study with confirmed facts, quotes and customer permission is marked ready.

Automatic steps: create a website page and branded PDF → adapt the approved copy into social posts and an email excerpt → check links and formatting → publish on the agreed schedule.

Output: a published page, a downloadable PDF and scheduled promotional content. The workflow can reformat approved claims, but cannot add new results or quotes.

Creating a new case study from raw notes is a different task. AI can prepare the draft, but someone must confirm the facts and customer permission before it enters this automatic publishing flow.

Build checks into the process

Everything looks fine until you click the main link and land on last month’s offer. Automation can repeat a small mistake across every channel.

Automate checks with clear answers: do links work, are required files present, and does the output meet the requested format? Check the finished email or video, as well as its source files.

Leave decisions that need context to an experienced reviewer. Is the message useful to this audience? Does it imply more than the evidence supports? Is the result worth publishing?

Choose the release rule in advance. Routine work can be published automatically within agreed sources, templates, audiences and limits. New claims or work outside those rules need approval of the exact version. The model cannot grant itself an exception.

Plan for failures too. Before retrying a failed publishing or sending step, check whether it is already completed. Keep a record of sources, outputs and approvals so a person can see what went wrong.

Review samples of successful runs as well as failures. A second AI check can help find errors, but another model agreeing does not make a claim true.

Expertise means knowing what good looks like

The awkward moment comes when everyone likes the draft, but an experienced reviewer spots a promise the product cannot keep. That is the judgement the workflow needs: someone who can explain what is wrong and how to fix it. Prompting alone does not provide it.

One person may supervise a routine workflow, but different decisions need different skills. A marketer checks customer relevance. A designer checks visual communication. An engineer makes the automation reliable. They do not all need to review every output.

Start with one recurring workflow. Track the time spent, the corrections needed and the full cost, including tools and review. Then check whether it achieved its purpose, such as useful replies or qualified enquiries.

Expand when the process works consistently. Let AI take on more production while people remain responsible for the goal, the standards and the result.

Sources:

• OpenAI: GPT-6 Astra model capabilities

• OpenAI: Function calling

• OpenAI: Structured model outputs