The Four Quadrants Behind Every Agent We Ship
Some weeks, NightWatch — our production-intelligence agent — clusters a fresh batch of recurring errors into an issue family and hands engineering a grouped case instead of a wall of noise. The same week, Bureau Copilot surfaces something in a credit bureau report that a busy underwriter might have missed on a Friday evening. Neither agent was built to the same checklist. That’s on purpose.
Why “Just Build AI” Doesn’t Work for a Regulated Lender
We’re a regulated, data-sensitive entity — RBI and the DPDP Act are first-order design constraints, not audits we retrofit around. Frontier models don’t run on Indian soil, so our architecture has to account for that from the start. And our customers generate almost entirely unstructured data — passbooks, cooperative-bank statements, cash flows nobody’s bureau has scored properly — so the easy “plug a model into clean rows” playbook was never available to us. Put those together and “ship everything, fast” stops being a strategy. We built a four-quadrant map instead: before we write a prompt, it tells us how much care an agent needs.
The Four Quadrants
Two axes decide the quadrant: internal or customer-facing, and aggregated signals only or individual customer data.
Quadrant 1 — internal, no customer data. Build these first. Near-zero regulatory exposure — production monitoring, engineering tooling. A bad output causes inefficiency, not a credit event.
Quadrant 2 — customer-facing, no customer data. The acquisition flywheel — a customer shares their own data at the point of consent; we store nothing linked to them individually.
Get InPrime Finserv’s stories in your inbox
Join Medium for free to get updates from this writer.
Quadrant 3 — internal, uses customer data. The moat. Never leaves our system, only credit and risk teams see individual outputs, and it’s where RBI’s explainability requirements matter most.
Quadrant 4 — customer-facing, uses customer data. Highest stakes. Proactively contacting a named individual with their own financial data means every touchpoint needs explicit consent mapping. We pilot these in one geography before they go anywhere near scale.
A bad recommendation on where to open a branch costs us efficiency. A bad recommendation on a collections call costs us trust.
What’s Actually Live
Nine agents are in production, mapped cleanly across the four quadrants. Bureau Copilot has logged 20,741 runs, turned bureau and repayment signals into 68,707 analyst-facing review items, and seen 10,246 of those actually acted on. Our AI calling-bot suite — EMI reminders, collections, lead qualification — has handled 13,000-plus calls.
On the engineering side, NightWatch’s orchestrator has run 22 reports and clustered 286 issue families, resolving 246 of them; its Autofix Worker takes the fixable families, drafts deterministic patches and tests, and raises the pull request — with a human deciding what’s safe to merge. The Incident RCA Agent has completed 117 root-cause runs and produced 27 AI-synthesis artifacts on top of a deterministic evidence base. The AI PR Review Agent sits on every pull request as a pre-merge gate, and ThreadPulse turns scattered Nexus threads into decision and follow-up summaries so context doesn’t get lost between product, engineering and ops.
The Rule Underneath All of It
Across all nine agents, four things never move: AI sits inside the workflow, not bolted alongside it. Deterministic control — rules, evidence, permissions, audit logs — stays the base layer; AI assists prioritisation and synthesis on top of it. Human accountability stays explicit — credit officers, relationship officers and engineers keep owning the decision. And sensitive data is treated as a constraint the system is designed around, not a risk we hope not to trigger.
For Our Customers, Our Team, and What’s Next
For customers, this is why an AI-assisted decision on their loan comes with the same explainability a human decision would. For our team, it’s why an engineer can trust NightWatch’s queue instead of triaging noise by hand. For InPrime, the roadmap follows the same discipline: an AI Financial Analyst backed by certified metrics, a Graduation Engine for pre-emptive top-up offers, and a vernacular Lifecycle Companion — each one built for the quadrant it belongs to, not the one that would ship fastest.
