Ask a credit head why a good MSME deal didn't get financed, and the answer is almost never "we didn't have the capital." It's some version of "we couldn't get comfortable in time" — the underwriting took too long, the analyst was stretched across too many files, the deal was too small to justify the hours a proper review takes. Capital was sitting on the balance sheet the whole time. What was missing was cheap, fast, trustworthy evidence.
This is worth naming precisely because the industry's own language obscures it. "Cost of capital" is a phrase everyone reaches for, and it genuinely matters for pricing. But cost of capital doesn't explain why a ₹2 crore working-capital deal gets a two-week underwriting cycle while a ₹50 lakh one, financially sound on paper, gets deprioritised or declined outright. What explains that is cost of verification — and verification cost does not scale down with ticket size the way it should for the economics to work.
The gap this leaves behind
The scale of what sits on the wrong side of that gap is not small. Of 8.7 crore registered MSMEs in India, only around 3.6 crore have ever accessed formal credit (SIDBI–TransUnion CIBIL, MSME Pulse, July 2026) — meaning the majority of a formally registered, taxable, GST-filing business population has never once cleared the bar for formal financing. Set against an addressable MSME credit gap the RBI Expert Committee put at ₹20–25 lakh crore, the picture is not "these businesses are uncreditworthy." It's that verifying creditworthiness, at the ticket sizes and volumes this segment operates at, costs more than the deal is worth to underwrite manually.
This isn't unique to India, and it isn't new — but the direction of travel matters. NBFC co-lending AUM reached roughly ₹1.1 lakh crore by 31 March 2025 (CRISIL), and the RBI's Co-Lending Arrangements Directions 2025, in force from 1 January 2026, widened co-lending to every regulated entity and every loan type. More lenders are entering a segment where the binding constraint was never capital — which means more underwriting desks are about to hit the same wall the earlier entrants already know well: verification doesn't get cheaper just because more people are trying to do it.
Why manual triage doesn't fix itself
The obvious response — hire more analysts, build a bigger credit team — runs into a wall that headcount can't solve, because the problem isn't analyst throughput, it's the shape of the work. A pre-credit read on a deal means pulling exposure data, checking capacity against thresholds, screening for related-party concentration, and cross-checking whether the same underlying trade shows up somewhere it shouldn't — each from a different source, each requiring a different kind of judgment, and none of it parallelisable across analysts without duplicating the pull. Add headcount and you add more people doing the same slow, bespoke assembly, not a faster version of it.
What actually changes the economics is separating two things that manual underwriting bundles together by necessity: the assembly of the pre-credit picture, and the judgment on what to do about it. Assembly — pulling exposure, checking capacity, screening for duplication, reading the trade's provenance — is mechanical enough to be instant and cheap once it's built once. Judgment — deciding whether a FLAGGED deal is a hard pass or a story worth a phone call — is exactly the part that should stay with a human, and exactly the part a good pre-credit signal frees up more time for, by taking the mechanical assembly off the desk entirely.
A signal, not a verdict
That's the shape of a pre-credit deal indicator: a fast, structured read — CLEAR, REVIEW or FLAGGED — computed the moment a deal enters a queue, with a durable link back to the full evidence room and, where a complete evidence Pack exists, a short-lived signed link to it. Not a credit score. Not an approval. A triage signal that tells a desk where its limited hours are best spent, computed from the same sources an analyst would pull manually, just not manually.
The distinction matters because a signal that quietly became a decision would be a worse outcome than the problem it replaced — a black box making credit calls nobody can audit is not an improvement on a slow human process, it's a different, harder-to-catch failure mode. The discipline that keeps a pre-credit signal a signal is refusing to let it output anything resembling a verdict: no allow/deny, no numeric score dressed up as an answer, nothing that lets a desk stop thinking once the indicator shows green. CLEAR means "nothing here needs your first hour." It does not mean "approved."
Why an agent changes what "triage" can mean
A human analyst pulling this picture manually, even efficiently, still means one analyst, one deal, one pull at a time. An AI agent working a pipeline can request the same pre-credit read on every deal that lands in a queue, the moment it lands — not instead of a human decision, but so that by the time a human looks at the queue, it's already sorted into what needs attention now and what can wait. That reordering, done at zero marginal cost per deal, is what actually lets a lender extend the same underwriting discipline to a ₹50 lakh ticket that it already applies to a ₹2 crore one — because the expensive part, the assembly, stopped being expensive.
This is why the signal is exposed over MCP under the same scoped-key model as every other family: an agent's key can be narrowed to exactly this read, nothing wider, and every call lands on the same audit trail and the same per-family meter as a human-triggered REST call would. The agent orchestrates the triage. The desk still makes every call that matters.
Underwriting capacity was never really capped by capital. It was capped by how many deals a limited number of expert hours could get comfortable with in a day. Take the mechanical half of that comfort-building off the desk, and the constraint moves to where it actually belongs — judgment, which was always the scarce and valuable part.
Watch the deal-signal step in the guided simulation, or see the full control model.