Most Teams Use AI to Post, Not to Listen.

Nearly every social team now runs AI somewhere in the workflow, and almost none of them point it at the audience.

HubSpot’s Social Media Marketing Report, a survey of more than 1,100 social media marketers, found 94% using AI in their workflows and only 13.54% using it for social listening and trend monitoring.

That is the whole problem in two numbers. AI has been aimed at the output and hardly ever at the input, which is why so much AI-assisted content reads fluently and says nothing. Generation is the cheapest thing these tools do. Comprehension is the expensive thing, and it is sitting there unused.

Generation is the cheapest thing AI does

Writing a caption is one of the lowest value tasks in a social team’s week, and it is the task almost everyone automated first. It is easy to see why. The output is visible, the time saved is obvious, and a blank document is uncomfortable.

It is also where the tools disappoint most. In the same survey, the jobs marketers said AI fell short on most often were image generation (40.2%), image editing (29.58%) and video generation (28.7%). The three creative tasks people reach for first are the three that come back needing the most fixing.

So the return on generation is thin at both ends. The work it replaces was cheap, and the work it produces needs a human pass anyway.

Comprehension is the expensive thing

Reading is where a language model has a real advantage over a person, and it is an unglamorous advantage: volume.

A social manager can read two hundred comments properly in a morning. A model can read twenty thousand across six months of posts and report which four complaints keep coming back, which product question nobody has ever answered publicly, and which phrase the audience uses that the brand never does.

None of that is creative. All of it is decision-grade. It is the difference between a team guessing at what to make next and a team knowing.

Fluent and empty is an input problem

In the same report, 45% of marketers named consistently producing high-quality content as their top challenge, and 41% said it is harder than ever to stand out organically. Both findings usually get read as a content problem. They are more often an input problem.

A model given a topic and a tone of voice will produce something competent and generic, because competent and generic is the average of everything it has read. Give that model fifty real comments from your own audience alongside the same brief, and the output changes character, because it now has something specific to be about.

The quality ceiling on AI-assisted content is set by what you feed in, not by which model you use.

What listening actually means

Listening is not a dashboard of brand mentions with a sentiment score attached. That is a report, and nobody makes a decision off it.

Useful listening means reading the text your audience actually produces, at a scale a person cannot manage, and turning it into instructions. Four sources are worth the effort, and you already own three of them:

• Comments on your own posts, including the old ones.
• DMs, replies and message requests.
• Comments under competitor and category posts.
• Reviews, support tickets and any other free text customers write.

Public forums belong in the same pile. Reddit is a listening tool before it is a content channel, and the same holds for anywhere your category argues with itself in public.

Start with the comment section you already own

Export the last six months of comments on your own posts. Most platforms allow it, and where they do not, your management tool will.

Then ask the model narrow questions rather than “summarise this”. Narrow questions produce usable answers:

• What questions appear more than five times, and were they answered?
• Which posts drew complaints, and what specifically was the complaint?
• What words does this audience use for the product that we never use ourselves?
• Which comments show buying intent, and what did we do with them?

That last one usually returns something uncomfortable, which is rather the point.

DMs are the highest intent text you have

A comment is a reaction. A DM is somebody choosing to start a conversation, and the expectations attached to it are unforgiving. The Sprout Social Index found 73% of consumers expect a response within 24 hours or sooner, and 73% say they will buy from a competitor if a brand does not respond.

Speed is the obvious use for AI here and the wrong one to start with. Triage first: have the model sort the inbox into questions, complaints, partnership approaches and noise, then route each pile to a person. Drafting replies can come later, once you trust the sort.

The pattern across a month of DMs is a content plan. If forty people ask the same thing privately, that is not a customer service issue, it is a post you have not made.

Read the comments under competitor posts

Most competitive analysis reads competitor content. That tells you what a rival decided to publish, which is the least interesting thing about them.

The comments underneath tell you what the category’s audience wants and is not getting. Unanswered questions on somebody else’s feed are the clearest content brief available, and they cost nothing to collect.

Feed a model a few hundred of them and ask what people keep requesting that nobody in the category provides. The answer is usually one or two specific things, repeated for years.

The weekly listening rhythm

Listening fails when it is a quarterly project. It works when it is a small fixed slot in the operating week.

1. Monday: pull last week’s comments and DMs, run the standing set of questions, note anything new.
2. Midweek: one competitor or category sweep, rotating targets so the set is covered monthly.
3. Before the planning session: turn the findings into three specific content briefs, not three themes.
4. Monthly: check which briefs were made, and whether the questions they answered stopped appearing.

Step four is the one teams skip, and it is the only one that proves the loop works.

What a listening pass should hand the humans

The output of listening is not a summary. It is a brief a person can act on without asking a follow-up question.

A good one names the audience question in the audience’s own words, quotes two or three real comments as evidence, states the format and platform, and says what a viewer should understand by the end. Anything vaguer becomes another generic post.

An outside read helps here, because a team that has stared at its own comment section for two years stops seeing the pattern in it. NBK’s audit and strategy work starts with the input for that reason.

Where not to trust the machine

A model asked to find themes will always find themes, including in noise. That is the failure mode to guard against, and the guard is simple: make it show its evidence.

• Require real quoted comments behind every theme, and check a sample by hand.
• Do not treat a sentiment score as the headline. It flattens the thing you needed to read.
• Never let it reply in the brand’s voice unsupervised. Triage is a machine job, tone is not.
• Keep the questions narrow. Broad prompts return broad answers nobody can act on.

What good looks like

A team that listens properly can name, without checking, the five questions its audience asks most and where each one is answered.

Its content calendar is traceable. Every item points back to something a real person said, rather than to a theme somebody liked in a workshop.

And its AI effort sits mostly on the reading side, with humans still deciding what to say. That is the inversion. Most teams have it the other way round.

How NBK thinks about AI in a social team

NBK treats AI as an operations tool before a creative one. It is strong at the parts of the job that are large, repetitive and text-shaped, and mediocre at the parts that need judgement about one specific brand.

So the first place it belongs is the input: reading everything the audience has said, sorting it, and handing people a properly evidenced brief. The content still gets made by humans, and it gets made about something.

That is the same argument as running social as a system rather than a content queue. The system decides what is worth making. The making was never the hard part.

Next step

If your team is posting regularly but still feels stuck, the constraint is often that nobody is reading the audience closely enough to brief them properly. NBK can help find it and rebuild the loop behind the content.