Why buy when you could build?
A majority of the social analytics tools I used in my early career sucked. I often found that the dashboards they created were lackluster, never quite what I needed for the job, and I was better off exporting the data and doing my own analysis in spreadsheets and presentations. This always perplexed me given the astronomical prices these vendors charged the agencies I worked at. Some even pushed premium analytics features that, among other things, let you reorganize the columns in a table *cough* Sprout Social *cough*.
Building my first tools
I was ecstatic when I first heard that GPT-4 could write Python, a programming language used for automating data analysis, creating data visualizations, and cleaning up messy data.
To start, I had ChatGPT write Python scripts that would create charts of organic performance data over time for the monthly presentations I was making for clients. I would explain the chart style and brand, provide the logo and file locations, and then ask for the script. Running the script locally kept all client data on my computer. This allowed me to take full advantage of AI without violating client confidentiality by sending client data to third parties like OpenAI. It also sped up monthly presentations and made the reports look more professional.
After I built those charts, I was hooked. I started building more tools that would automate the tedious parts of my job, improve the quality of the deliverables for clients, or let me do better research.
What AI changed about building
If we fast forward to today, long-running autonomous agents in specialized coding harnesses with higher intelligence models have drastically changed the effort, economics, and ability needed to create software. AI can now also use a computer and navigate annoying, designed-by-engineers-for-engineers dashboards like Google Cloud Console, TikTok for Developers, and Meta for Developers. The barriers to building tools have dropped significantly.
Much of this existed before AI. It was just expensive. You needed to know how to code, or have a team of data scientists and engineers. Now a $100–$200 OpenAI or Anthropic plan, or a $10 OpenCode plan, gives you access to the most advanced AI models available. You can use them to automate many of the tedious parts of data analysis the industry had done manually for years.
With the power of these agentic coding tools, I have been able to build and productize the audience research and content analysis workflows I did manually as a consultant. They are automated, take a few days instead of weeks, and include better data. They also leverage machine learning algorithms (HDBSCAN), embedding models (V-JEPA 2), and graph network algorithms (Leiden) to provide richer and deeper insights to clients. The results still need a human to validate them, since the groupings can shift a lot depending on how the models are set up. These are tools I own. I can run them through my startup, GAR-AI, offer them as part of my consultancy, or bring them to whatever full-time role I choose next.
Judge software by its API, not its interface
In this new world, I've started to evaluate software less on its user interface and more on what data is available via its API. The value of Sprout Social is no longer that it is a social analytics platform. The value is its single API that, once all my social accounts are connected, I can use as a wrapper to access all my data. Sprout's API is only available on its most expensive plan, Advanced. So yes, I still pay Sprout, but for the data, not the dashboards.
Building still has a cost
That said, building your own tools still has a cost. Someone has to make sure it works, validate the outputs, and fix it when something breaks (or know how to direct an AI to fix it). Someone also has to make sure it's secure and hackers can't steal your data, manage who can access it, and keep it reliable. Those are the non-glamorous parts of software engineering. And the privacy question from my first scripts doesn't go away: coding agents send what they read to the model provider, so you have to be deliberate about which client data they can see.
The lines between non-technical and technical have begun to blur as software has been abstracted away from coding languages to human language. Some readers might balk at learning how to get agents to write software, learning about GitHub, or system-designing a data pipeline. But from my seat, I think these skills will become increasingly important, much like learning how to use Excel.
Personal tools vs. organizational tools
For personal tools, the economics of building are cheaper than ever. The best way to get started is to pick one of the coding plans I mentioned and choose one task to automate. It will teach you the ins and outs of coding with agents. For organizations, there is a large question about data governance and ownership. But from my vantage point, the industry is moving towards a hybrid model, where vendors provide the data and internal tools do the analysis.
So, why buy when you could build? At the end of the day, I'm not saying never buy. Vendors like Sprout still have a place, but for me it's for the data they provide, not the dashboards sitting on top of it. The analysis, the part I used to export into spreadsheets because the tools never quite did what I needed, is now something I build and own. If you're a practitioner, the barriers to doing the same have dropped significantly, so pick one tedious task and start building.