AI in Recruitment: Evidence, Skill Matching & Workforce Productivity | India AI Impact Summit 2026

Contents

Executive Summary

This talk examines the critical decision organizations face when deploying AI: whether to automate tasks entirely or augment human decision-making with AI assistance. Drawing on evidence-based research, particularly an RCT in Ghana on teacher hiring, the speaker argues that the automation vs. augmentation choice depends on measurable impact outcomes—not just cost savings—and that rigorous evaluation is essential before deployment. The panel discussion extends this to India's informal economy and development context, emphasizing that AI's benefits must be designed inclusively to avoid amplifying existing inequalities.

Key Takeaways

Automate or Augment? Neither is universally right. The answer depends on (a) measured impact on outcomes you actually care about, (b) baseline human capability, and (c) whether humans will use AI input. Rigorous evaluation is non-negotiable before scaling.
India's Informal Economy is AI's Untapped Frontier. 500M+ informal workers have massive productivity upside—not from job displacement, but from new possibilities: language barriers removed, geography decoupled from opportunity, skills accessible via voice. This creates new professions, not just efficiency gains.
Design for Inclusion from Day One, or Amplify Inequality. Women, rural populations, disabled users, older workers, and non-English speakers will be left behind if AI systems aren't intentionally inclusive. "Adverse incorporation" means they get integrated into AI-driven systems in harmful ways (data extraction, higher costs, dignity losses) without benefits.
Real-Time Feedback Loops Trump Annual Evaluations. The pace of AI change makes traditional RCTs slow. Use AI itself to collect real-time data, test, iterate, and prove impact quickly—building trust and enabling course correction mid-deployment rather than years later.
Infrastructure > Tools; Race to the Top > Hoping. India's Aadhaar/mobile success, Amul's farmer advisory, and Blue Dot jobs platform show that shared digital infrastructure—designed for participation, not extraction—enables inclusive growth. This prevents the social media trap where technology is deployed first and harms evaluated later.

India AI Impact Summit 2026

Key Topics Covered

Automation vs. Augmentation Framework: The fundamental organizational choice when deploying AI
Evidence-Based Policy: The role of rigorous evaluation (RCTs, process evaluation) in AI deployment decisions
Cost-Benefit Analysis Beyond Economics: Why cost savings alone are insufficient; impact on outcomes (e.g., patient health, hiring quality) must be central
Real-World Case Study: Teacher hiring in Ghana using GPT-4 recommendations
India's Informal Economy & Gig Work: AI's potential for 500+ million informal workers, 10 million gig workers, and emerging digital professions
Inclusion & Equity in AI: Gender gaps in AI adoption, age barriers, rural-urban disparities, and accessibility for disabled users
Digital Public Infrastructure: Blue Dot program as a model for matching jobs, skills, and services at hyper-local scale
Productivity vs. Development: Growth opportunities in AI for low-productivity economies vs. efficiency gains in high-productivity ones
Evaluation Frameworks: Four-stage evaluation model (model → product → use → development impact) and the importance of real-time feedback

Key Points & Insights

The Automation/Augmentation Decision Requires Evidence, Not Intuition: In audience polling, 90% initially favored augmentation (AI as assistant) based on perceived risk, but when confronted with actual impact data showing 30% worse health outcomes with automation vs. 5% worse with augmentation (despite 20% cost savings), preferences shifted. This demonstrates that cost savings alone are insufficient decision criteria—organizations must evaluate actual outcome impacts.
Ghana Teacher Hiring Case Study Defied Expectations: The speaker expected augmentation (human + GPT-4 recommendations) to work best. Instead, full automation (GPT-4 alone) increased successful teacher hiring by 70%, saved money, and improved teaching quality. Crucially, augmentation failed because: (a) human teachers ignored AI recommendations anyway, (b) the process slowed them down without changing decisions, and (c) the humans weren't initially good at hiring. This shows that augmentation's success depends on the baseline human capability and whether humans actually use the AI input.
India's AI Opportunity is Fundamentally Different from Developed Countries: India is the #2 global user of Claude (by volume, due to population) but #101 per capita. However, usage is 15x more efficient (tasks taking 4 hours taking 15 minutes vs. 10x globally) and 45% concentrated in computational/mathematical tasks (highest globally). This suggests India's informal economy and gig workers can leapfrog productivity constraints through AI in ways developed economies cannot—because there's more upside potential.
Inclusion Must Be Designed In from Day One: Women are 20% less likely than men to engage with generative AI at work, risking entrenched AI bias. Barriers include gender, age, disability, rural-urban access, language, and safety concerns. The "adverse incorporation" risk means AI can amplify both benefits and inequalities—requiring intentional design for inclusion, not afterthought remediation.
New Work & Professions Will Emerge, Not Just Automation: The speaker challenges the binary framing. YouTube didn't augment housewives; it created a new profession. Similarly, India's gig economy (zero 20 years ago, 10 million today) shows new possibilities emerging from technology removing friction. AI removing language, geography, and access barriers may create professions we cannot yet imagine—particularly in lower-income regions.
Blue Dot Model as Infrastructure: The "Blue Dot" approach inverts job/skill discovery: instead of job-seekers finding jobs, the system finds people. In one Ghaziabad district, AI-enabled Blue Dot discovered 4,000 jobs vs. 10 shown by traditional platforms. Applying this to scholarships, training, and apprenticeships can reach excluded populations (girls, people with disabilities) in rural areas without requiring them to travel, wait 6 months, or navigate complex bureaucracy.
Evaluation Must Be Rapid, Iterative, and Answer "Should We Use AI At All?": The four-stage evaluation framework (model → product → use → impact) includes a "stage zero" question: should AI be deployed at this decision point? Beyond proving cost-effectiveness, evaluation builds trustworthy AI ecosystems and prevents "shipping and hoping." AI itself can accelerate real-time data collection and feedback loops—enabling faster iteration than traditional impact evaluation.
Growth vs. Efficiency Distinction Matters: In high-productivity economies, automation primarily saves costs (efficiency). In low-productivity economies like India, AI's primary value is enabling growth—offering services, products, and professions that didn't exist before. The small business owner doesn't lose out to automation; they gain capability to compete, expand, and serve new markets.
Digital Infrastructure, Not Just Digital Tools: India's success with Aadhaar and mobile (500M+ users, digital payments ecosystem, now Amul's AI advisory reaching 1,500 farmers including 60% women in Gujarat) shows infrastructure approach trumps top-down solutions. Infrastructure is built without assuming the "best" final form; others innovate on top. This prevents "shipping and hoping" and ensures a "race to the top" vs. social media's "race to the bottom."
Measurement of Who Benefits Is Critical: Recurring question: if productivity rises, who captures the gains? Design must determine whether benefits accrue to workers, employers, consumers, or are captured by platforms. India's experience suggests inclusive infrastructure design (e.g., government/civil society oversight, not just private platforms) helps distribute gains equitably.

Notable Quotes or Statements

"The very next question I bet is going to be: do we automate or do we augment? I'm going to be making basically three points... the first one is I want you to do some introspection and think about in your organization how you would start to think about how to answer this question." — David (JPAL Co-Chair, University of Zurich)

"Automation had a huge impact on success rates. We increased successful hiring rates by 70%." — David, on Ghana teacher hiring study results

"I think the real question is whose work is getting transformed, who's measuring it, and who's deciding what to do." — Murugan Raswadvan (Wise Foundation CEO), reframing the AI impact question

"When they started this off with one or two cows, they had no experience of milk farming... The wisdom of 50 years of Amul, [data from] 30 million cows, is now available to them in their language, voice, when they want." — Shankar (AoStep/Blue Dot), on Amul's AI advisory empowering women farmers

"45% of all usage [of Claude in India] was in computer mathematical tasks, which was the highest percentage anywhere globally... and more efficiency gains than we've seen elsewhere—tasks that would take 4 hours took 15 minutes." — Elizabeth Kelly (Anthropic), on India's AI usage patterns

"India from an extremely low productive base, informal economy... the options for being more productive are now much more because of AI... India can show many more possibilities than what it showed with YouTube or with a mobile phone." — Shankar, on India's unique AI opportunity

"Women are already 20% less likely than men to engage with generative AI technologies at work... risks that what we call a vicious circle of AI bias getting entrenched." — Dr. Becky Faith (FCDO/IDS Sussex), on gender gaps in AI adoption

"It's not just a one-for-one trade-off. Growing the pie... more products and services... it comes back to how do we actually get AI fluency into the hands of more people so more people can have those opportunities." — Elizabeth Kelly, on growth vs. efficiency framing

"You create a basic idea, you have a certain target inclusion number in mind, and then you keep improving it... You don't build a road thinking the best car has already been invented." — Shankar, on India's infrastructure approach to digital scale

"We need to keep learning from the last 25 years of digital in development... how to include people not in an adverse way, in a way that actually improves their rights and development outcomes." — Dr. Becky Faith, on applying lessons learned

Speakers & Organizations Mentioned

Speaker	Title	Organization	Expertise
David	Co-Chair, AI Evidence Initiative; Professor of Economics	JPAL, University of Zurich	AI in development, evidence-based policy, labor economics
Murugan Raswadvan	CEO	Wise Foundation	Tech, AI, policy, social impact intersection
Dr. Becky Faith	Senior Research Fellow	FCDO (UK Foreign Commonwealth Development Office); IDS Sussex	Gender, technology, digital futures for work, development
Elizabeth Kelly	Head of Beneficial Deployments	Anthropic	Responsible AI deployment, safety, public benefit
Shankar (Shankul Mohanty)	Co-founder & CEO	AoStep Foundation	Digital public infrastructure, education access, Blue Dot program
Sead Obermeyer	(Referenced, gave earlier talk)	Implied healthcare/AI	AI applications in healthcare, diagnostics

Other Organizations/Initiatives:

JPAL (J-PAL): Abdul Latif Jameel Poverty Action Lab — conducts RCTs for development interventions
Anthropic: AI safety company; deployed Claude; runs economic index tracking AI usage globally
FCDO/UK Government: Funds AI for development research via AI Ford Program
Wise Foundation: Invests in tech-policy-social impact organizations
AoStep Foundation: Builds digital public infrastructure in India (blue infrastructure, education scale)
ID Insight: Research partner on JPAL/FCDO AI evidence initiatives
Amul: India's largest dairy cooperative (36M farmers); launched AI advisory
Center for Global Development: Partners on four-stage AI evaluation framework
ESRC Center for Digital Futures at Work: UK-based research on technology and work

Technical Concepts & Resources

Concept/Tool	Description	Context
Automation vs. Augmentation	Binary choice: AI makes decisions independently vs. AI recommends, humans decide	Central framework for organizational AI deployment
RCT (Randomized Controlled Trial)	Gold-standard evaluation methodology; Ghana teacher hiring study used this	Evidence-based impact measurement
GPT-4	Large language model used in Ghana teacher hiring experiment (2+ years ago)	Reference point for AI capability at time of study
Claude/Claude Code	Anthropic's AI assistant; used in Anthropic's product development (co-work product built in 10 days by AI)	Example of AI-augmented/automated software development
Anthropic Economic Index	Dataset tracking global Claude usage by geography, profession, task type	Evidence of AI adoption patterns (India #2 globally, #101 per capita)
Four-Stage Evaluation Framework	Model → Product → Use → Development Impact (+ Stage Zero: "Should we use AI?")	Structured approach to AI impact evaluation
Adverse Incorporation	Development economics concept: technology integrated into economy in harmful ways (data extraction, higher costs, dignity loss)	Risk framework for inclusive AI deployment
Blue Dot Program	Geographic job/skill/service matching platform; inverts traditional search (system finds people vs. people find jobs)	Infrastructure model for hyper-local opportunity matching
Aadhaar	India's national digital identity program (billions affected); mentioned as successful scale model	Reference for inclusive digital infrastructure design
AMUL Advisory	Voice-based AI system delivering cow health/farming advice in Gujarati to smallholder women farmers	Real-world inclusion example (1,500 farmers, 60% women, 20-minute scholarship discovery vs. 6 months)
Process Evaluation	Qualitative evaluation of how AI systems are used, barriers, gendered/age/rural-urban adoption gaps	Complements RCTs in understanding real-world deployment
AI Fluency	Skill/literacy for using AI tools effectively; currently concentrated among tech workers	Critical barrier to inclusive adoption
Tokens/Inference Cost	Anthropic Claude: ~$140 per million tokens; inference-time compute increases cost with problem complexity	Cost structure relevant to automation vs. augmentation ROI

Research Projects & Initiatives Referenced

Initiative	Lead/Partner	Focus	Status
JPAL AI Evidence Alliance for Social Impact	JPAL, FCDO, ID Insight, others	RCTs + process evaluation of AI for development	Ongoing; multiple studies in pipeline
AI Ford Program	FCDO	Gender-inclusive, evidence-based AI deployment in development contexts	Active; collaborators present at summit
ESRC Center for Digital Futures at Work	Dr. Becky Faith (UK)	Gender, disability, age, rural-urban barriers in AI adoption; process evaluation	Active research
Ghana Teacher Hiring Study	David (JPAL/U Zurich)	RCT comparing automation (GPT-4 only) vs. augmentation (GPT-4 + human) vs. human-only	Completed; 70% improvement in hiring success with automation
Ethiopia Agri Advisory	(Mentioned, not detailed)	AI-based agricultural advice	Launched Feb 3, 2025
Amul AI Advisory	Amul, India	Voice-based AI advisory for dairy farmers in Gujarati	Launched Feb 11, 2025 (1,500 farmers in Gujarat event)
Blue Dot Infrastructure	AoStep Foundation	Jobs/skills/scholarships matching in Ghaziabad, Karnataka, Ranchi, Bihar	Operational; discovering 4,000 jobs vs. 10 via traditional platforms

Contextual Data Points

India's AI Adoption: #2 globally by total Claude usage (India); #101 per capita
Productivity Gains in India: 15x speedup (4 hours → 15 minutes) vs. 10x globally; 45% usage in computational/math tasks (highest globally)
India's Gig Economy: 10 million gig workers today; ~zero 20 years ago
Amul Cooperative Scale: 36 million farmers; 1,500 attended AI advisory launch
Gender Gap in AI: Women 20% less likely to engage with generative AI at work
Ghana Study Impact: 70% increase in successful teacher hiring with automation
Blue Dot Job Discovery: 4,000 jobs discovered vs. 10 shown by traditional platforms (Ghaziabad)
World Bank Data (Last Week): IT sector productivity in India has risen; companies hiring more (counterintuitive)

Policy & Practice Implications

Stage Zero Question Required: Governments and organizations should ask "Should we use AI?" before designing automation