What is running in production
Bajaj Finance is not describing a pilot. The numbers reported with its Q2 FY27 results are production figures: 65 agentic applications live, 129 autonomous agents behind them, and a headcount inside the dedicated AI unit that rose from 230 to 328 over the period. Another 300 roles sit on the digital platform side.
Customer service shows the shift most plainly. The share of do-it-yourself interactions handled automatically by voice and text bots went from 71 percent in the first quarter to 75 percent in the second. That is a four-point move in three months, or roughly 5.6 percent relative growth, on a base that was already mostly automated.
Underwriting productivity rose by more than 20 percent, according to the company. A risk-modelling test on 500,000 business accounts showed a 15 to 20 percent improvement in outcomes, described as early results rather than a settled figure. The full disclosure is on the BW People report.
The arithmetic behind the loan bots
Two figures matter more than the agent count. In Q2 FY27, AI text and voice bot leads produced ₹2,744 crore in disbursements. A year earlier the same channel produced ₹1,295 crore. That is 2.1 times the volume, on a channel that did not exist at scale two years ago. One crore is 10 million rupees, so the quarterly figure is 27.44 billion rupees.
The half-year number is where the target gets interesting. Bajaj booked over ₹6,500 crore in AI-assisted disbursements across H1 FY27 and is guiding to ₹11,000–12,000 crore for the full year. H1 already covers 54 to 59 percent of that range. To hit the guidance, H2 only needs ₹4,500–5,500 crore, which is less than H1 delivered. Unless the bots stall, the company is closer to its annual number than the guidance implies.
The earlier Q1 figure is smaller but cleaner as a unit economic. ₹517 crore in incremental disbursements came from 400,000 offers generated after analysing 45 million customer interactions. That works out to about ₹12,900 of disbursement per offer and roughly ₹115 of disbursement value for every interaction the model touched. Neither number is a profit figure, and Bajaj has not published the cost side of the pipeline.
How the staffing ratio changes
Divide the AI unit by the live agents. Today that is 328 people for 129 agents, about 2.5 employees per agent. At the stated targets it becomes 400 people for 600 agents, or 0.67 per agent. The ratio would fall by nearly a factor of four.
The caveat is obvious: the AI unit is not only agent builders. It covers data, evaluation, platform and governance work too, and the company has not broken the headcount down. Still, the direction is clear. Bajaj expects each new agent to need less human scaffolding than the first 129 did. That is the actual claim hidden inside the 600 target.
A $4 million bet on Indian model weights
On November 18, Bajaj Finserv Intelligence and Pennant Technologies put $4 million jointly into Soket AI, about $2 million each. Soket builds sovereign foundational language models and maintains an open-source agentic harness called Loop.
The word sovereign is doing work here. An Indian lender building agentic workflows on Indian-trained weights is a data-residency and procurement decision as much as a technical one, and it mirrors what European banks have been asking their vendors for since the AI Act took effect. Whether Soket's models hold up against frontier systems is not something the funding announcement answers.
The rulebook European lenders work under
Creditworthiness assessment of natural persons sits in Annex III of the EU AI Act as a high-risk use. Any European lender running underwriting agents on retail customers lands in that category, with obligations on data governance, logging, human oversight and technical documentation.
The timing slipped. Full obligations for Annex III systems were pushed to December 2, 2027 by the AI Omnibus package, so European risk teams have roughly a year of runway that Bajaj did not wait for. Transparency rules did not slip: Article 50 has been live since August 2, 2026, which means the chatbot layer Bajaj automates at 75 percent must, in the EU, tell the customer it is an AI. And an automated refusal of credit still runs into Article 22 of the GDPR, which gives the customer a route to human intervention.
Collections is the sharper edge. Bajaj lists collections among the agentic domains. A bot that negotiates repayment with a defaulting borrower raises questions the AI Act treats as high-risk when it feeds into credit decisions, and consumer-protection rules add pressure on tone and frequency. No equivalent disclosure exists in the Bajaj numbers.
Open weights as the European option
A European team that wants the same architecture without sending customer data to a US API has one obvious route. Mistral Large 4, released in late October 2026, ships open weights under the Mistral Ecosystem Open Weight License and lists at $0.70 per million input tokens and $2.10 per million output tokens, with cached input at $0.07. For comparison, GPT-6.2 Sol lists at $2 and $10, Gemini 4 Neon at $3 and $15. DeepSeek-V4.1-Flash undercuts everything at $0.15 and $0.60 off-peak under MIT.
The catch is the same one we keep hitting in AI Arena: a model of Large 4's class is not a single-GPU deployment. Self-hosting it means rented H100-class capacity or a small internal cluster, plus the evaluation and logging work the AI Act expects. The API is cheaper until compliance enters the calculation.
Does Bajaj Finance use AI to approve or reject loans?
The company describes agentic workflows in lending and underwriting, and reports a 20 percent-plus productivity gain in underwriting. The disclosed disbursement numbers come from offer generation and bot-sourced leads. Bajaj has not published how much of the final approve-or-decline decision sits with the model rather than a human.
What separates an agentic application from an AI agent?
In Bajaj's own counting, one application can host several agents. The current 65 applications run 129 agents, roughly two per application. The target of 118 applications and about 600 agents implies roughly five agents per application, so each workflow would get more granular rather than the portfolio simply widening.
Can a European lender copy the 75 percent bot-handled service rate?
The automation rate is technically reachable. The legal conditions differ: Article 50 transparency applies to every customer-facing bot, and any automated decision with significant effect needs a documented route to human review under GDPR Article 22. Bajaj's figure is a service metric, not a compliance metric.
What the disclosure does not contain is as telling as what it does. There are no error rates for the agents, no override rates for human reviewers, no figure for how often customers contest an AI-generated offer, and no cost per interaction. The 15 to 20 percent risk improvement is still labelled early by the company itself.