Reporting is still the one that get's me. Seems like a task which could be plausibly automated, but I'd hedge a bet that a lot of marketers in this group spend hours doing it every month!

@andrewbond I'd love to hear more. I definitely agree, but this could encompass a huge variety of things!

• Keyword research/clustering to actual topics I care to create content around

• Content briefs that are actually helpful to the user, not just an SEO hitlist

• Selection of the best reviews that addresses both the search experience concern while balancing helpfulness of on-site “shoppers”

• Opportunity sizing of PR opportunities, especially when it comes to selecting and targeting the right publications that will have an audience that is receptive to the product … not just spamming for eyeballs

• Dynamic/Personalized page experiences that can conform to user’s journey

◦ Considers users coming in from the side door of Google vs those who navigate direct to the homepage

I like the thread, but I think AI should actually do both

Creativity is hard for many people, but they know what looks good when they see it

AI is actually really great for helping people express their emotions/thoughts through different mediums where they might have “writer’s block” or just lack the mechanical skills to accomplish what they dream up

@softplus absolutely, but I think AI can help us explore creative latent spaces MUCH more thoroughly. I see the creative role of the future being about directing/curation being primary.

Its already happening I feel. I know who are already trying to automate the entire marketing process right from what content for which persona and the kind of content and channel.

@raghu many of us here work on this extensively, the next 5 years will be transformative and this group will be pioneers in exploring what is possible and how humans/AI work best together and separately in a creative professional field.

This doesn’t answer your question but I wouldn’t mind less automation - in marketing / SEO at least.

The danger is always automating to the point where you’re no longer needed. Kind of a like a be careful of what you wish for situation.

Let’s say an SEO manager automates 90% of their 40 hour work week, should they still be paid for that full 90%? What is that time going to be replaced with? There’s only so many meetings with stakeholders, organic strategies they can develop etc.

Automating the boring stuff to free up time for strategic / creative things is what humans have been doing for the past few hundred years. If we let AI also do the strategic / creative stuff then we’re really not needed.

I think we currently feel it less in the west but AI is already taking work away from people in the sub-continent. The work 5 people did for x $ a day can be done by 1 person clicking a button.

@harpreet This is the way everything that is automatable will go, it is inevitable unless we want to try and stop all AI progress.

The hope is that it frees us up to do things we never before were capable of doing because we now all essentially have endless expert help in any domain we need. Everyone will just start thinking bigger. They wont worry that their old job as a b-roll logger got automated when they can now be the director of their own Hollywood quality feature film without help from anyone else...

Now how people get compensated in this future... that's a whole other debate and one that deserves a lot more attention than it is getting

100% agree with you - automation is inevitable, outside of the marketing like healthcare there’s a lot of really good, better potential use cases of AI.

Compensation is the tricky thing, different countries will have different laws. Worldwide population is growing so someone smart enough will have to figure it out.

If UBI becomes a thing because of AI then why would someone even be incentives to click buttons on an AI tool or get into plumbing, building etc but that’s a much bigger debate.

An AI that recommends and connects to other AI/tools/API for the project you're working on (staying up to date). So you're just having a conversation along the way and it recommends a tool, signs up if you accede to it, and then enter in the info as you need and continue on to the next part. ????

Let’s go deep. Thanks I love this question.

Automation - The scraping of product reviews. The categorization of these reviews by main type of complaint, praise, or feedback. The filtering of fake malicious bot or spam reviews from legitimate ones.

Future looking - The creation of synthetic data to enrich attributes of the reviewer that may be missing to create a virtual crm profile. The ability to infer missing customer/consumer data that does not show up on reviews and create a synthetic panel of people (leaving reviews follows the 1% rule of the internet).

The ability to test different product messaging and product improvements to this virtual panel created from reviews to run business forecasts by country, region, or down to particular stores.

The ability to then ask every virtual panel participant their NPS, net promoter score, of these product changes or improvements. Oh for fun, how each would rate the product, what the review would be, and for those who became promoters to explicitly describe someone close to them who would benefit (crm addition) and expand the breadth of the virtual panel.

Oh yeah - data enrichment with AI exists right now. By AI I mean guessing things with sub-systems checking if the field exists in other systems - like all the ways people can type or mistype addresses but canonicalize and clean the data.

Philosophically - The issue with being predictions out into the future is the scenario in the movie Minority Report shares where instead “pre-crime” we have “pre-mortum” of ideas and creations. We get so used to trusting our AI systems that we forget how to roll it all back. Human/AI hubris turns failure into a statistic over am essential part of learning and growth.

Two AI for marketing applications I love.

First, the thing I’ve been working on for years and never really worked in the way we wanted until now: turning unstructured, messy keywords into structured entities and aggregating associated data, like traffic, impressions, CTR, pages, revenue and anything else around that. You can use that for all kinds of on-site actions. Most named entity vocabularies don’t cover all the things people think about, talk about and search for, so we need the AI to help expand structured vocabulary to do that. I explained this to some AI geek friends — it’s like NER but you don’t know in advance which entities or names you’ll find. They just laughed — “that’s impossible”. That isn’t limited to query data — you could use it for product data or product reviews, like @victor.m.pan is talking about. But I just love query data.

Second, the AI marketing application I’ve been most inspired by: using digital AI twinsumers to build synthetic market research, like talks about.

@a Have you found any frameworks that work well when converting unstructured to structured data?

I’ve played around with Microsoft’s Graph RAG () , Diffbot (), and OpenAI’s latest Structured Outputs () + Pydantic ()

I haven’t really found anything that feels super reliable that really takes unstructured into structured unless it’s some like ecommerce data where I can pass it XPATH

Eric, I know Diffbot well. In a different life Similar AI competed with them and made vertical-specific multi-modal product understanding and query understanding. There were a few companies doing parts of that like Twiggle (closed down), GrokStyle (acquired by Facebook) & others. Nowadays of course you can get great multi-modal embeddings multi-domain or for specific domains.

I don’t think there is a turn-key solution, yet. My suggestion would be to focus on a single use case and domain first, then see if the underlying model can generalise to handle more. What we’re building for keyword structuring combines multiple LLM prompts and a product stack we built over the years. Previously we hand-built knowledge graphs for fashion, furniture, home & garden _etc_. Then we started ‘growing’ these ontologies from search data. Now, we have something using LLMs that works for any site. But applying it to product data is pretty new and likely we will need to do quite a lot more to get it to work well. (In a way, keyword data is harder because of its brevity).

I think structured outputs + PyDantic is generally a good approach. For other projects, we have built various checks to see if the output is in the format we need or to convert it into that format. Maybe worth considering optimising prompts automatically using declarative programming with DSpy?

The pydantic + Structured outputs is the best I know of at this point... The other thing I do having multiple functions that can fix/heal broken outputs.

This may not be what you are talking about, but I like getting a corpus's embeddings, clustering them, and then using an LLM to analyze and understand the clusters. Often times this makes it possible to make very messy data much more structured by leveraging the LLM to do taxonomy creation.

I've played around with graphrag and trying to generate knowledge graphs from scratch with LLMs. Here is an example of what I mean:

and this:

The use case I was working on I thought would benefit from a knowledge based graph for RAG, but ultimately context windows got so large that I didnt need to do RAG at all anymore. That said, the construction of the knowledge graphs with LLMs is EXTREMELY useful, but with confabulation drawbacks. Here's something I did with knowlege graph generation.

> I like getting a corpus’s embeddings, clustering them, and then using an LLM to analyze and understand the clusters. Often times this makes it possible to make very messy data much more structured by leveraging the LLM to do taxonomy creation.

I love this approach!

Using KGs for multi-hop Q&A is also smart. I’ve seen being used for this too.

Have you had any success using DSPy's? They are fascinating, and perhaps a key to creating truly robust automation pipelines (maybe AGI too?), but they are also incredibly dense and difficult to understand/troubleshoot. Maybe I am missing something, but I'd love to hear about it if you've had success.

I should say the same thing has been true for agentic frameworks that allow the agents a lot of autonomy. The only way I've been able to create stable and valuable pipelines thus far is through extremely explicit design of the architecture myself, and careful prompt tuning at each step.

Kristin, I agree. Agentic frameworks seem great, but they rely on foundation model weaknesses, like planning and reasoning. So, in practice with today’s systems, it’s hard going. DSPy’s pipelines look great and doesn’t it make sense to move from coding each of the features like we used to for deep learning nets to declaratively stating what the pipeline should look like? But I found the learning curve steep. We’re not using DSPy in production anywhere.

ITs cool seeing it grow up before our eyes, but im hoping someone can really improve DSPys and prove their value in more explicit use cases and abstractions that make it easier to use

GPT-5 Might suddenly make all these existing agentic frameworks magically work, that would be wild.

Maybe Strawberries are all you need?

btw the thing I talked about above — “turning unstructured, messy keywords into structured entities” — is . You can play with a .