Blog
Essays by Vikas Goel on AI agents, Voice AI, self-evolving systems, and the gap between AI demos and AI in production.
Indian enterprises report strong AI adoption, but the same research points to weaker governance and capability. Here is what the published numbers do and don't say, and a five-question test for whether your first agent is ready for production.
An agent can give the right final answer and still be unsafe, slow or ruinously expensive. Here is a seven-metric scorecard for AI agent evaluation, how to build a test set from real work, and how to set release gates before you ship.
You can't reliably prompt your way out of prompt injection. Treat the model as untrusted, and secure what it is allowed to do. Here is a blast-radius review with eight checks for any AI agent that can take real actions.
An LLM gateway gives you one control point for model traffic: routing, keys, budgets, logging and fallbacks. It won't make an agent safe on its own. Here's what a gateway actually does, when it's worth adding, and how to decide between building, using open source, or buying.
A voice AI demo in a quiet room proves very little. Production voice AI in India has to survive code-mixed speech, noisy phone lines and a tight response budget. Here is a practical latency budget, the failure modes that matter, and a test plan you can run before you sign a vendor.
Most teams blame the model's price for a runaway agent bill. The real driver is re-sent context, and it grows quadratically with the number of steps. Here's a reproducible cost model you can run on your own numbers, and the changes that cut a 10-step agent's input tokens by about 65% without touching the model.
India's 2,100+ Global Capability Centers finally have the board mandate to lead on agentic AI. 83% are scaling GenAI and 58% are piloting agents. But a GCC is structurally optimized for delivery and cost arbitrage, and agentic AI rewards almost the opposite muscle: owning outcomes, buying instead of building, and running eval and governance in production. This is the operating-model climb from cost center to AI-first, and the gap most GCCs fall into on the way up.
A voice AI agent that hallucinates a balance, a policy limit, or a payoff amount hasn't made a cute mistake. In banking, insurance, or healthcare it has created a spoken, recorded, actionable liability. You don't fix that with a better model or a sterner prompt. You fix it with an architecture: retrieve the authoritative record, verify it against policy, then speak or refuse. This is how reliability gets engineered into a regulated voice agent.
Should you build your own AI agents or buy a platform? The honest 2026 answer is neither on its own. Most enterprises should buy the 80% and build only the critical 20%. This is the decision framework, the real total cost of each path, and why ~75% of in-house AI builds fail, from a CTO who's made the call with real budgets.
Ordinary AI is confident, which is fine until the problem is hard, contested, and expensive to get wrong. Then confidence without the ability to doubt itself becomes the danger. This is why autonomous agents hallucinate on high-stakes problems, why a bigger model makes it worse, and what the deterministic gatekeeper architecture does to make an agent trustworthy: grounding, tool-input validation, structured-output enforcement, and confidence-gated escalation.
A voice AI sales or lead-qualification agent fails in two ways: it's too slow, or it's confidently wrong. Both are voice model optimization problems. This is the production playbook: the latency budget, the levers that move it, and how to make the agent fast and grounded enough to trust on a live call.
95% of enterprise AI pilots show no measurable P&L impact, and it's almost never the model's fault. A pilot and a production system are different disciplines. This piece covers why pilots die in the gap between them, and the specific operating model the surviving 5% share: one owner, buy-not-build, eval discipline, a production harness, and ROI baselining.
Tokenmaxxing looked like a vanity metric. The real reason AI agents burn tokens is that they were trained on GitHub's verbose code, then set loose without a harness. From a CTO running production AI in front of 290M+ users: what's going on, and what to build instead of writing longer rule files.
AI agents are not a faster chatbot. They are autonomous systems that can hold a goal across hundreds of steps, recover from their own mistakes, and act in the world. The implications for business are bigger than most leaders realise, and the failure modes are different from the ones they're preparing for.
What I've learned from putting AI voice agents on real customer calls across India, the Middle East, and Latin America. Latency budgets, language switching, and why a single hallucination is a customer-impact event, not a chat-window inconvenience.
Pure-AI customer service is a sales pitch. Pure-human customer service is increasingly unviable at scale. The interesting question is what the right blend looks like, and which decisions about that blend are made by the architecture vs. left for runtime.
Thirty years of building enterprise systems has taught me that the model is the easiest part. The substrate underneath (observability, recovery, queueing, idempotency, identity, budget control) is where AI automation succeeds or fails at scale.
In 2020, with twenty years of software engineering behind me, I enrolled in an M.Tech in AI at BITS Pilani. It was the most useful thing I've done in the last decade. This post is about why, and about why I think more experienced engineers should take their AI education more seriously than they currently do.