Punnam Raju Manthena, Cofounder & CEO at Tekskills Inc. Partnering with clients across the globe in their digital transformation journeys.
Businesses all over have been scaling their AI chatbots, coding copilots and AI agents. AI-related spend is expected to reach $1 trillion by 2027, growing by about 40% year-on-year, per an analysis by Bain & Company—so much so that you are at an advantage if you have an early lead.
But, as with most technologies, AI is growing faster than what businesses would be able to explain, or faster than what’s required. This gap between scaling observability and explainability could well trigger the next wave of AI failures/incidents.
What Is AI Scaling?
When you are scaling your AI, you are enhancing the data volumes, computational ability and model, either to improve your existing AI or to spread it across your business.
With AI, you can potentially speed up workflows, quicken your business cycles, enhance your customer personalization through smart chatbots, improve your decision-making, track performance indicators and proactively predict market changes, among other things.
Importantly, scaling AI can also mean freeing up your people from repetitive, mundane tasks, and letting them work on more strategic and creative initiatives, all of which can make your business better.
What Is AI Observability?
AI observability is a practice of monitoring the AI features of your application—prompts, responses, costs, errors and output quality—in real time so that you can pick out and troubleshoot issues quickly and keep your systems safe and reliable. It gives you a peek into how your AI systems will behave in complex and dynamic environments.
What Is Explainability?
Equally important is explainability. How can you ever trust an AI system that can’t explain how and why it makes its decisions? After all, explainability tells you, in a clearly understandable manner, how the model works internally and how it reasons.
Why does explainability matter? First, it builds trust. When you understand why the AI makes certain choices, you can feel more confident using it. It detects any possible bias, unfair rules or prejudices in the data.
The Myth Of Retrofitting
It’s easy to think, “We’ll add observability once we scale,” but this is a myth that needs to be corrected. Retrofitting—adding parts, features or technologies after their initial manufacture—may seem like an easy workaround, but retrofitting trust into a live system can be way more expensive and risky than building it in from day one.
If you add observability after you scale, when you’re already live and busy, you’re prone to inconsistent supervision, intake of raw and unshaped data resulting in high costs, and rewrites.
It is better to adopt observability while scaling and monitor real-time data across your systems so that they remain precise, cost-effective and reliable even as your workloads grow. By “baking it in,” all it could take to maintain observability and build trust as you scale are some standard SDKs or templates.
Trusting AI
What does it mean to trust AI? It is just the trust, or the confidence, that your users and other stakeholders will have in the dependability, safety and, importantly, ethical integrity of your AI systems.
Your AI systems should be reliable and work correctly all the time, be transparent and explainable about how and why it makes a particular decision or offers a particular recommendation, be fair to all stakeholders without any malicious bias, be accountable for the outcomes and offer data privacy and security against malicious attacks.
This trust matters because trust influences how safely you can embrace, gain from and control transformative technologies like AI. It’s also important to acknowledge that mistakes can cause irreparable damage, especially in vital sectors such as healthcare and finance.
The Three Layers Of An AI System
You can visualize an AI system as having three distinct layers: model level, agent level and decision level.
The model-level refers to the underlying intelligence engine that interprets and processes the text and returns the relevant responses. This is a typical passive text-in, text-out entity without any inherent external awareness, and can erode decision quality in case of any unchecked drift or hallucination.
On the other hand, the agent-level builds a working layer around the model level, invokes the required tools, does the reasoning part and manages memory.
The last one, decision level, does the finer reasoning and selects the action. It decides the tool that needs to be called, the parameters to pass and how to address unexpected errors or ambiguous results.
The Next Move For Business Leaders
As in any design scenario, scaling comes with some uncomfortable trade-offs. Observability may sometimes slow down scaling, aside from adding to the cost. If you run after speed, you’ll outsmart your competitors alright, but may get burned out in the long run. It’s speed versus accuracy, wherein your system could respond faster but with less accuracy, or respond slowly with greater accuracy. Otherwise, you could go in for small, inexpensive models, but they would give you low-quality output.
As a business leader, what would you do? I recommend having traceability as a launch gate for your system to check if your product is fully ready for market release and not as some post-incident fix. Importantly, you’ll have to assign someone who can explain it before the deployment and not after the release. For all this, you need to look at observability spend as a risk-reduction budget, and not some design line item.
The Bottom Line
Let’s face it. You need to scale to keep pace with the market, but you need to fit in observability as well as explainability so that you’re better off facing issues tomorrow. Not being able to explain what your AI is doing would be as unacceptable as not knowing where your data is.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?

