Most conversations about AI safety happen in the abstract: alignment, hallucination, guardrails, words that sound important and explain nothing. Then there's the version MITS actually had to solve, which was much smaller and much more concrete: what happens when a nervous new employee asks Priya a question she doesn't actually know the answer to?

The tempting, lazy answer is to let the AI guess. Large language models are extremely good at sounding confident even when they're wrong, and a new hire has no way to tell the difference between a correct policy answer and a fluent, plausible-sounding invention. Getting that wrong on someone's actual first day, about something like leave policy or tax declarations, isn't a minor bug. It's a trust problem. So MITS built the system to do something most consumer AI products don't bother with: know its own limits, out loud, in real time.

How the Answer Actually Gets Built

When a new joiner asks a question during the Open Q&A segment, the system doesn't just hand it to GPT-4o and hope. It runs a full retrieval pipeline first. The question is converted into a vector embedding using OpenAI's text-embedding-3-small model, then matched against a knowledge base of real MITS documentation, indexed in Azure AI Search during development and demo, and in Qdrant, running in Docker on MITS's own on-premise servers, in production. Only after the most relevant chunks of real company knowledge are retrieved does GPT-4o get involved, reasoning over that retrieved context to compose an answer. This pattern has a name in the industry: retrieval-augmented generation, or RAG. It's the difference between an AI that's memorized a topic and one that's actually allowed to open the textbook before answering.

The Threshold

Here's the part that matters most. Every answer the system generates carries a confidence score. Above 70%, Priya answers directly, in her own voice, without interruption. Below that threshold, or if the question touches something sensitive, the system doesn't guess anyway. It quietly flags the moment to the HR representative sitting on the same call, through a real-time event called FlagQuestion, delivered over a SignalR connection (ASP.NET Core's real-time messaging framework) the instant the uncertainty is detected, along with the question itself and the confidence score behind it.

The HR representative sees an amber alert appear in their dashboard within a second or two and can step into the conversation immediately. No dead air, no "let me check and get back to you" that leaves a new hire hanging. The handoff is close to seamless, which is exactly the point, it's supposed to be invisible unless you're the one watching for it.

Why 70% and Not 100%

It would be simple to make the system flag everything below near-total certainty, and safer on paper. But that would defeat the purpose, HR would end up fielding every routine question personally, right back to the original bottleneck this system was built to solve. The 70% line represents a deliberate bet: confident enough that routine, well-documented questions get handled instantly, cautious enough that anything genuinely uncertain gets a human before it gets a wrong answer. It's a tuning problem as much as an engineering one, and it's the kind of decision that only gets made correctly when the people building it actually care about the outcome, not just the demo.

The Quiet Philosophy Underneath

There's a broader idea buried in this one design choice: humility as an architecture decision. Most AI products are built to always produce an answer, because an empty response looks like failure. MITS built the opposite instinct into Priya, sometimes the correct output is "I'm not sure, let me get someone who is." For a system that's literally someone's introduction to a new employer, that restraint isn't a limitation. It's the whole point.

---

MITS Global Consulting has spent over two decades earning trust in banking, insurance and fintech, industries where "I'm not sure" said honestly is worth more than a confident guess said wrong. Learn more at https://www.mitsit.in/