You don't need AI, you need a person who knows AI
Every company is making a huge shift to add AI to their processes right now, whether or not it's actually the right tool for the task. Almost no one is able to accurately evaluate what benefit AI is having on their work. Most companies are seeing huge spend with no ROI, and overall the answer to "is AI profitable yet?" is a resounding no. But engineers who know what they're doing are seeing huge improvements in their workflow using AI.
Every company is making a huge shift to add AI to their processes right now, whether or not it's actually the right tool for the task. Almost no one is able to accurately evaluate what benefit AI is having on their work. Most companies are seeing huge spend with no ROI, and overall the answer to "is AI profitable yet?" is a resounding no. But engineers who know what they're doing are seeing huge improvements in their workflow using AI.
What's the difference between seeing performance gains or just wasting money? A person who knows how to use AI correctly. There's a strong argument that AI doesn't have ROI at all, based on the costs and output. But individual people who know what they're doing are able to see performance gains. So instead of just slapping AI onto something, look for a person who knows how to avoid the challenges and maximize the impact.
Challenges
I've seen several of my colleagues in the industry charged with adding AI to their workflow, and the only metric the company checks is the number of tokens they use. This is the equivalent of evaluating a farmer by the number of miles they drove a tractor. Did they grow or harvest any crops? Did they even drive the tractor over the area where crops could be grown or harvested? A farmer who backs their tractor in and out of the barn a thousand times would appear to be doing better than a farmer actually growing food to supply their town.
Undirected or poorly directly AI models will still produce trash. The phrase "garbage in, garbage out" gets even more true at the speed at which LLMs allow us to move. I've been able to see direct examples on my teams and in my own work. Senior engineers who understand software architecture will produce good products quicker, and people who don't understand the basics of writing software will produce a huge volume of completely unmaintainable code, also very quickly.
Impact
In software engineering, AI services like Claude Code and Cursor can and do have a huge impact on delivery timelines, quality, and cost. One of the first gains I saw on my teams was test coverage. Engineers famously don't love spending their time writing test coverage, despite knowing the importance of proper testing in software architecture. AI models are capable of evaluating code and generating test coverage, which can quickly raise the overall coverage on a lagging project to an acceptable level. These tests still need human review to ensure correctness, and you can't let people on your team take shortcuts like using AI to rewrite a failing test instead of fixing the underlying problem. But this an easy way to get real system-wide improvement on a codebase with fairly minimal oversight.
One of the other big gains was the design to dev pipeline. Our teams use Figma and Claude primarily, and the new functionality to bring design elements straight into code is a huge time saver. This is something that was theoretically handled with design tokens already, but seems to be much easier with the AI integrations.
Other Concerns
The cost question will come up again in the future. Right now companies like Anthropic and OpenAI are burning money and offering access to models at rates probably much cheaper than the eventual price when these companies try to become profitable. Using local models for smaller tasks is one way to address this. Some companies are even bringing back junior engineers to save money on tokens. The common thread in all the solutions to cost problems is having a person who understands how and when to send tasks to AI, which model to use, and how to orchestrate the work.
I've also avoided talking about the negative impact of current AI models on the environment and local communities. There are real problems there, globally in terms of energy usage and fresh water consumption, and locally in terms of data centers being placed near already disadvantaged communities and causing more harm. But these are all solvable problems with known solutions. We can generate functionally limitless energy between clean renewables like solar and wind, with nuclear to handle sudden changes in demand. We have the technology to process salt water into fresh water, and purify water for drinking. And we have the ability to regulate where and how data centers are constructed so it's fair. All of these require time, money, and competent leadership to accomplish, but they're very much possible.
Takeaway
Most current AI models are specifically engineered to be likable and agreeable in order to increase engagement. They'll happily tell you that bad ideas are good, or that huge mistakes are smart business strategy, if it seems like what you want to hear. If you want to use AI effectively and not just waste money, you need to deeply understand the technology behind it, or you need to bring in people who do.
I've intentionally chosen not to support comments on this site. If you feel a strong need to discuss something, you can use the social sharing links above to write your own comment. If this post has sparked any big ideas, either positive or negative, I encourage you to write your own post and share it on your own website. Feel free to drop me a link in the contact section, and I'll add links here to interesting responses.