This article started a couple of weeks ago based on an article I saw in Venture Beat looking to re-evaluate the build vs. buy debate. If only I had finished it then and beat the rest of the market…
The debate is not old, but the cost dynamics and speed have changed over the last year as generative AI continues to take the world by storm. It is a great tool to accelerate someone who knows what they’re doing, but I’m willing to bet if you are trying to use this as a replacement to buying software, then you will end up spending far more than you expected.
The old belief was to outsource things that are outside of your core expertise. Things like word processing and spreadsheets are the very old examples. ERP and CRMs are far bigger things that power the business, and have also mostly been outsourced. Yet, when it comes to customizations specific to the company, it would be a very expensive proposition. Hiring a development team to build and maintain it, or expensive professional service firms to modify the off the shelf software (from back when you had to buy it at a store and install it).
The creating and maintaining custom software is very expensive, and is why the CFOs for most companies would lean towards the buying over building. Additionally, it would offload the risk to the software providers.
While the original software would be installed on-site was a one-time purchase, there was the option for support contracts to provide the patching and bug fixing required. However, as software moved onto the cloud or hosted with the numerous SaaS providers, the economics changed towards renting the software through a subscription model (per-user, tier, and add-ons). This is where a majority of company IT spend goes these days, and is exacerbated further with the “AI add-on” that can add a not so trivial per user/month fee for a chat bot interface. In short, the SaaS companies have gotten a tad greedy.
Now, we have the ability to “vibe” our own versions of a CRM so why do we need to pay thousands of dollars per user per month when we can make our own version? Additionally, why would I need to pay a SaaS company for special functionality that can be generated using AI that effectively pulls from the same underlying source?
This is why the tech stocks tanked the first week of February, 2026!
Vibe coding is a great performance multiplier for experienced developers, and a great enabler for non-developers who have problems to solve that don’t rise to the level of hiring a developer. When it comes to the fundamentals that went into the build vs. buy adage:
- Cost
- Risk mitigation
The up-front costs part of the equation looks like it could be reduced. After all, you don’t need to hire developers nor do you have to pay for the SaaS version anymore. Granted there is the per-token cost to generate the code, but that is a one-time cost right?
But…
Then you remembered the original reason why you bought that program or subscribed to that particular service to begin with: you didn’t want to develop or maintain it. Maintenance has its own costs, and it will also increase your risk over time. Both of these factors will eat into whatever savings you would have had by firing your IT team or the SaaS company by turning to AI.
For example: vibe coding your replacement CRM also means have to maintain it. All software, including what will get generated, pulls in external code. That code undergoes updates, and needs to be integrated into your software. So now you have to keep track of those changes, and update your vibed software just like a regular dev team. This may sound simple, and you could have an agent take care of it. The challenge is that maintaining it, even by hand, can be challenging due to the underlying software changing how you are expected to use it, and that doesn’t make it a drop in fix. This is the one thing AI cannot account for, nor do I expect it to do so anytime soon.
If anything, the example highlights another major cost that has not been touched by the various media. You must still verify the results produced by the code generated by AI. Unlike other technologies, you cannot tell when things go wrong and the results are generated out of nowhere (the hallucinations). After all, you wouldn’t want to be sanctioned or terminated due to an error in what was generated (the article cited was from last week).
The thing is, reviewing code takes as long, if not much longer. Some of it is piecing together the requirements and whether the new functionality works, but the other part (and this is specific to the AI generated code) is that the code generated is inherently more complex and verbose.
In conclusion…
This is me speaking from an experienced engineer standpoint. For someone with less experience and not knowing what to look for, it would be very easy to believe the results without really knowing or understanding if it is doing the right thing. Luckily, most of the tools we want can easily be verified by playing around with it. The stuff that is a simple proof of concept to verify behavior or try out new processes. Generating code does make this a lot faster as well as a quick way to get immediate feedback.
If this is just for you, that is one thing, but when it comes to your company, then off-loading this risk makes more sense. This is where the debate inside the organization is still alive and you must still apply your own tech diligence.