AI Observability 2.0: Why Enterprises Need To Measure The Full Cost And Value Of AI
Yuri Gubin is Chief Technology Officer at DataArt, working with organizations integrating AI into complex engineering environments.
gettyAs we’re approaching the end of 2026, it’s interesting to look at what we expected from AI this year and what really happened. If you go back a few years, many of the predictions were based on guesstimates and wishful thinking. Now we do see that AI is everywhere, agentic development is quite good and there are methods evolving with examples of how AI really boosts development. We also see the impact on individual productivity, even outside engineering, in knowledge management and other areas.
What we’re still missing, despite all the efforts, is an industry-proven methodology that has become a standard for AI-enabled development. We have different frameworks explaining the stages of AI adoption, and examples of what works and what doesn’t, but we’re still some way from having an approach that everybody can use, regardless of which LLM they’re working with.
And then there’s the ROI question. We still need better measurements and evidence of the benefits, particularly when it comes to understanding whether the same approach will work in different environments. This is where observability could become particularly interesting.
Organizations will be rethinking the costs of AI, and I’m not talking about subscriptions or tokens alone. It’s the cost of data, infrastructure and everything you put in place to make AI happen in your company. All these expenses add up, and we’ll likely see more scrutiny and greater demands for results.
This is already happening, of course. We’ve been anticipating this and advocating for reasonable use of AI, where you validate results with data rather than relying on expectations. Some organizations will try to spend less without substantial changes in quality or delivery, and others will spend more, but this rethinking is already underway.
If you look at the work companies were doing three or five years ago, they were paying for people, cloud infrastructure and other operational expenses. Now you may be doing the same work, but you’re also consuming AI resources, and the value of that work may be different. So the economics change, including margins and how companies capture the value created. This may affect business models across the board, including outside IT services.
One area we’ll likely see developing is observability, but in a different sense from how we’ve traditionally understood it. Before, we would talk about operational telemetry and system metrics, while with AI, we need to understand where it’s being used and how much value it has created.
I would call it AI Observability 2.0, because there should be a new standard and a new wave of technologies that help us understand the impact of AI.
Right now, I don’t think the ecosystem is there to tell us exactly how much we spend on each action when we consider the end-to-end flow. And it’s not only about pure numbers. You have the technology, the decision-making, the judgment, the quality of the result, the spending and the value created. We need to understand the entire chain, including what happens when something doesn’t work well and how we can troubleshoot, evaluate and react faster. Some of this may come through new technologies, or it may be changes to the same technologies we’ve had in our toolboxes before.
One thing we still haven’t resolved is how to measure ROI in a way that gives us reliable, repeatable results. Let’s say AI works well in one team and you see productivity improvements. If you apply it to another team, even within the same organization, will it work the same? The technology stack might be different, the problem space is different and it’s difficult to say whether you should expect the same benefits.
There’s something similar happening with data. Its importance was already on everyone’s mind at the beginning of 2026, but as you start making progress, you discover new depth to the challenges you already knew about.
You may have created a data lake and put all the data in one place, but then you discover that it doesn’t help you improve the use of AI as much as you expected. Maybe the conventional data lake is too slow, or maybe you’re still missing some data. You wanted to transform the company, and now you realize that having the data in one place doesn’t necessarily solve the problem.
I hope 2027 will be a year of AI growth, but in a more reasonable, pragmatic and mature way, rather than explosive experimentation. Hopefully, as an industry, we’ll make progress on methodologies and standards for AI-enabled development, because those are still to be defined.
I also expect technologies that help us understand the value AI creates will become more important. Observability is an interesting area because we’re still missing that understanding of the complete process, and we’ll need it to make better decisions about AI investments.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

