Improving Your Signal to Noise Ratio

posted on November 20, 2026 by Will Mapp, III

 

Hey! Stop scrolling and put down your PSL!

Welcome to the inaugural article of our new blog series, Signals and Solutions! If you’ve been following our blog writing exploits you would have noticed our series explored how data, software, and analytics impacts our lives in everything between healthcare and the invention of the pumpkin spiced latte.

We’re revamping the blog series to lean heavily on the intersection between data sciencing technology and its impact on how we work, live, and play.

In this first issue, let’s talk about signal.

It’s been a year since I last wrote about turkey data. And at that time, around Thanksgiving, McKinsey reported 88% of companies testing AI, while 40% were deploying agents in production. Most companies were testing the waters. Others were wondering what kind of real ROI could be realized. And frankly, other companies waited out of fear.

A year later, the results are in, and there’s a lot of noise. According to a November Zapier survey, 71% of companies have deployed AI agents in production. However, the ROI isn’t there yet. In McKinsey’s State of Global AI, 80% of their survey’s respondents say AI has improved their productivity, but only 37% can tie their AI usage to their organization’s earnings.

So, what gives?

Productivity gains AI can deliver are easy to imagine: the ability to do more using less effort, agents running around doing grunt work while I work on more strategic tasks, and a robot to wash and fold laundry are appealing. After a glut of half-baked use cases, the set of real problems AI could solve became easier to identify, but a lot of noise has increased with adoption challenging your ROI.

Tokenmaxxing

AI tokensOrganizations turned to our favorite online platforms like ChatGPT and Claude to vibe code and generate new applications. OpenAI’s Codex and Claude Code integrations made it easy for developers to prototype ideas and refactor code. This year’s proliferation of agents made delivery faster, but more costly. Uber famously burned through its entire token budget by May of this year, and had a hard time connecting token spending to customer benefiting features. Most developers are salaried and contractor costs are relatively predictable, throw in a plethora of agents for each developer and you’re having a token party. Excessive tokenmaxxing makes a solid ROI hard to predict.

Security and Accuracy

Security is often assumed at the end of projects, unfortunately, and it’s no different when it comes to AI app development. In most cases, generated code expressly implements the request, and in many cases security additions have to be prompted. Implementation accuracy in many cases is assumed by vibe coders and bugs resulting from app specific input validation and range checking fall into the rework pile. Which requires more human time and…tokens. See above.

Rehashing the Old Using the New

I’ve been using the expression, “this is a hill I’m going to die on,” a lot lately. And one of those hills is the: we-should-be-rethinking-how-we-do-things-in-the-AI-era. AI gives us asynchronous execution, disparate and disconnected information retrieval, and unstructured data source processing. We have a tremendous opportunity to solve problems differently. Reddit groups and Build-in-Beta chats are filled with devs and builders using AI agents to follow human processes and steps. Most of our traditional filing, cataloging, and process systems are built around human-error and are aids to minimize our mistakes. In order to realize ROI using AI solutions, agentic AI problem solving requires new, abstract thinking.

Human-in-the-Loop

Human in the loopWhat?! Will, what are you saying?  What…are you actually saying? Did you spit out your pumpkin spiced latte? No. I’m not saying what you’re thinking. Currently, AI requires our oversight. And that’s a good thing. Many projects are over-integrating HITL where all decisions are gated by humans, and this is like hiring a bad intern to do your job. Other projects are neglecting to implement HITL effectively. Both extremes are costly, and costly to your ROI.

How Do You Realize a Return and Boost Your Signal

Nvidia, hyperscalers like AWS, and a new generation of powerful, open weight models are reducing industry’s need for massive token spends. The latest open weight models have given organizations more choice in using lightweight models for interpreting user intentions, and larger models to reason solutions. Organizations are opting to standup their own servers in clouds they already use to (hopefully) bring more predictability to their AI spending.

Breaking headlines about agentic breaches and ISO 42001 shined a new spotlight on the unique challenges posed by AI systems. I have had more conversations with non-technical leaders about data and system governance than I have in a long time. When someone from accounts receivable is throwing around “AI governance framework” and not making a parlor joke, we’ve arrived. AI governance and upfront data privacy go a long way to minimizing risk.

We take HITL seriously at Qlarant. We take a holistic view of the solution and understand where agents can deliver the most impact while also doing the least harm. On that hill about rethinking problems and solutions, part of smart human-in-the-loop implementation is specifically understanding where human failing can be augmented by machine intelligence, and understanding the risks involved. Can AI redundancy minimize hallucinations and mistakes in tedious areas? Are workflows optimized for AI and HITL? A big chunk of making gains in HITL efficiency are asking the right questions.

We put it out there in our inaugural edition of S&S, Signals and Solutions. I am a believer in AI technology and think we can do great things with it. I also believe AI presents an opportunity to deliver great benefits to our customers and bring about the next level of human capability. Now that we know AI can do the job, it’s up to us to realize efficiency gains by making good choices on our token usage, adapting common sense security approaches for AI, and analyzing problems using a different lens while asking the right questions.

Stay tuned, and switch to the praline latte when you sit down to read the next edition of Signals and Solutions.

about the author

As Chief Technology Officer, Will Mapp keeps a constant eye on the future and ensures Qlarant is at the forefront of the latest and emerging technologies. See all posts from Will Mapp, III.

Let's Talk About Solutions

How can Qlarant help you improve quality and program performance?