India's IT/BPO sector — which employs 10–15 million people and anchors an entire aspirational middle class — is structurally contracting under AI pressure. TCS shed 23,400 jobs in FY2026. Net hiring across top-5 IT firms turned negative. Analysts project BPO employment could fall from 4 million to under 1 million by 2030. But the bigger problem isn't the death of the old model. It's that the obvious replacement — AI annotation and RLHF work — is the same trap wearing a different costume. This piece argues that India's only durable next export category is the accountability layer: work that exists not because AI can't do it yet, but because someone must be legally liable when it's wrong.
The problem
The old India trade had a simple arbitrage: rich countries had expensive labor for codified, low-judgment, repeatable work; India had cheap labor for the same. That built a $50–280B industry. AI eats exactly that category first — codified, low-judgment, repeatable — so the arbitrage is closing structurally, not cyclically.
The test that matters: capability gap vs. accountability gap
A capability gap closes. Every category pitched as AI-proof on capability grounds — translation, customer support, junior coding — has shrunk as models improved. Training a workforce against a capability gap means racing a target that accelerates faster than the pipeline.
An accountability gap doesn't close. It means: someone must be legally, professionally, or financially answerable when this output is wrong, and that person cannot be a model. This widens as AI improves — more AI output in regulated domains means more volume requiring an accountable human, not less.
FIELD 1: LAW — the floor is enforced, weekly, with money
As of mid-2026, the Charlotin database documents 1,348 cases globally where AI-hallucinated content was submitted to courts — 915 in US courts alone, growing from 2/week in early 2025 to 2–3/day by late 2025. Sanctions escalated 11x in 18 months: from $5,000 (2023) to $55,597 (2025), with a Sixth Circuit order of $30,000 against two attorneys in March 2026 for a brief with 24+ fabricated citations. Courts have been explicit: the duty to verify AI output before filing is non-delegable, regardless of which tool drafted it. This isn't a capability problem. It's a liability structure.
FIELD 2: MEDICINE — the regulator has already drawn the line
FDA's framework for AI-enabled medical devices explicitly tiers by autonomy. The lowest-friction tier — AI surfaces recommendations, clinician makes all decisions — is where nearly all cleared AI diagnostics sit. Higher autonomy tiers carry sharply higher regulatory burden. This means the accountability layer is built into the approval pathway itself: AI that informs a human is the path of least resistance; AI that replaces human judgment is not. This holds even as diagnostic accuracy improves — the bottleneck being regulated is liability, not accuracy.
FIELD 3: FINANCE — FINRA is writing the requirement now
FINRA's December 2025 Annual Regulatory Oversight Report, for the first time, dedicates explicit guidance to AI agents in brokerage workflows. Core risk flagged: agents "acting or transacting without human validation." Required fix: human checkpoints before execution. FINRA Rule 3110 — requiring firms to supervise associated persons — applies regardless of AI involvement. Technology-neutral stance means firms are fully responsible for AI-driven outcomes under existing law. 44% of investment advisory firms currently using AI have no formal validation process for outputs — a gap regulators explicitly flagged as a 2026 examination priority.
The trap: rlhf and annotation are not the answer
India is already building a data annotation / RLHF industry (500,000+ trained annotators, market growing 23–27% CAGR). This is tempting to call "the next BPO wave." It isn't. Annotation exists to make AI better, which means the category shrinks as models improve at synthetic data generation and self-critique — exactly what happened to the original BPO categories. Domain-specialist annotation (a doctor ranking AI medical outputs) sits closer to the accountability line than image tagging, but it's still a bridge, not the destination.
THE DURABLE CATEGORY sits one tier up: humans standing between AI output and a real-world consequence the model isn't allowed to be liable for. Verification that scales with AI adoption rather than against it.
Approaches
Legal AI Output Verification as a Scalable Export Service
Build India-based verification firms specializing in checking AI-drafted legal filings against primary sources before licensed attorney sign-off. 1,300+ documented hallucination cases globally, sanctions escalating, citation-verification now a compliance obligation. India has English fluency, legal-adjacent talent pool from existing LPO (Legal Process Outsourcing) industry, and cost advantage. This is the LPO wave but with AI as the client rather than the law firm.
Feasibility: high
Clinical AI Triage and Second-Read Verification Layer
FDA's "AI suggests, human decides" tier for diagnostic imaging creates structural demand for a verification workforce sitting between AI output and licensed clinician sign-off. High-volume, lower-acuity imaging (chest X-rays, diabetic retinopathy screening) is where AI already operates at scale and the human-in-the-loop requirement is most standardized. India's existing medical BPO infrastructure (radiology reading, transcription) is the closest existing foundation.
Feasibility: medium
Financial Compliance Verification for AI-Assisted Advisory and Trading
FINRA is writing mandatory human-checkpoint requirements for AI agents in brokerage workflows in real time. 40% of advisory firms use AI, 44% of those validate nothing. The compliance verification gap is active, escalating, and being written into examination priorities. India's existing BFSI outsourcing talent base is a direct precursor workforce. The moat: regulators require the human layer, so the category can't be automated away without a regulatory change.
Feasibility: high
What would move this
This is not something one person can move forward alone — and deliberately so. I'm looking for people who want to work on this seriously:
— Regulatory lawyers or legal technologists who understand the AI liability landscape in US/EU/UK courts and can help scope the legal verification opportunity with real numbers.
— Physicians, radiologists, or healthcare AI practitioners who've worked with FDA-cleared diagnostic AI and understand where the human-in-the-loop requirement actually sits operationally.
— Compliance professionals or ex-FINRA/SEC people who understand what the 2026 oversight report means in practice for how firms will actually procure verification services.
— India-based IT/BPO founders or operators who've built outsourcing practices and can stress-test the go-to-market side — what it actually takes to build institutional trust with global clients in a regulated domain, vs. a cost-arbitrage pitch.
— Policy researchers or economists thinking about India's employment transition — especially anyone working on the 60%+ formal-sector automation susceptibility figures and what structural interventions are actually on the table.
If you've pushed back hard on something in this piece and have data to back it up, that's even more useful. Reach out.
Originally published on uditraj.site — one of the Big Problems reports.
Top comments (1)
The capability gap vs. accountability gap distinction is the sharpest framing I've seen on this. It reframes the whole question from "what can AI not do yet" (a race you lose) to "what must a human sign off on" (a race that widens as AI spreads). That's a genuinely different strategic position.
One pushback worth testing: the accountability layer may be more defensible than annotation, but it's also far more concentrated. Legal verification, clinical second-reads, and financial compliance checkpoints are all gated by licensed professionals and regulated institutions — which means the addressable market is smaller than BPO ever was, and entry requires trust and credentials that don't scale the way a headcount ramp does. That's not a reason not to do it; it's a reason the transition will be slower and more credential-heavy than "the next BPO wave" implies. The 10–15 million people currently in the sector can't all move into it.
The second thing I'd push on: liability can be contractually pushed down the chain. If verification firms start absorbing liability for AI output at scale, the first thing their clients will do is write indemnity clauses into the contract — which either raises insurance costs past the arbitrage advantage or pushes the work back to cheaper jurisdictions. So the real moat isn't the human layer itself; it's a regulatory requirement that the liability can't be outsourced past a certain point. Worth checking whether FINRA's tech-neutral stance actually holds that line.
Genuinely the most interesting thing I've read on India's AI transition in a while.