AI in Work: Humans, AI, or Both?

Abstract

The session began with a 25‑minute research presentation by David Yanagizawa‑Drott, who framed the central dilemma facing organisations today: should AI be used to automate tasks or to augment human decision‑making? He walked the audience through a simple decision‑framework, illustrated it with a live poll, and then shared evidence from a randomized field experiment in Ghana that compared AI‑automated versus AI‑augmented teacher‑hiring processes. The presentation highlighted the scarcity of rigorous evidence on AI‑augmented policies and the importance of systematic evaluation.

The rest of the hour was a panel discussion moderated by Murugan Vasudevan. Panelists from academia, industry, and civil‑society examined real‑world examples (e.g., Amul’s AI‑driven milk‑cow advisory, Anthropic’s “co‑work” product, AI usage patterns in India) and debated broader themes such as equity, productivity, the informal economy, measurement, and policy design. A recurring thread was the need to generate and disseminate high‑quality evidence so that AI tools can be deployed responsibly and inclusively.

Detailed Summary

1.1 Opening & Framing

Introduced himself as co‑chair of J‑PAL’s AI Evidence initiative and highlighted the ubiquitous decision organisations now face: “Do we automate or do we augment?”
Emphasised that the answer depends on organisational context (government, NGO, private sector) and that evidence is scarce; most practitioners rely on intuition or a handful of reports.

1.2 Three‑Step Structure

Introspection – ask participants to close their eyes, picture their own organisation, and identify one task they would like AI to help with. Then decide whether that task is best suited for automation or augmentation and list the evidence supporting the choice.
A Simple Decision Framework – presented a two‑dimensional matrix (cost vs. outcome quality) and suggested evaluating:
- Cost side (labor savings, AI implementation cost)
- Impact side (error rates, health or performance outcomes)
Empirical Evidence – described a randomised field experiment in Ghana that compared three hiring pipelines for rural teachers: (i) human‑only, (ii) AI‑augmented (GPT‑4 generated shortlists reviewed by teachers), and (iii) AI‑automated (GPT‑4 made final selections).

1.3 Live Poll Simulations (Uttar Pradesh health‑diagnosis scenario)

Used a QR‑linked Mentimeter poll to let the audience choose between automation (≈ 20 % labor cost saving) and augmentation (≈ 5 % saving).
Initial poll: ~90 % chose augmentation after hearing a brief cost–benefit sketch.
After a short debate (participants tried to convince neighbours) the second poll showed the automation share rising to ~10 %, illustrating how preferences shift when people discuss trade‑offs.

1.4 Key Findings from the Ghana Teacher‑Hiring Experiment

Automation outperformed augmentation: the automated pipeline raised successful hiring rates by 70 % relative to human‑only, cut costs, and shortened processing time.
Augmentation added latency without improving decision quality; teachers still relied heavily on tacit knowledge that the AI could not capture.
The result contradicted the presenter’s initial intuition that augmentation would be superior, underscoring the value of field experiments.

1.5 Take‑aways & Call for Evidence

Highlighted three “unknowns” that organisations must grapple with:
1. Effectiveness of AI in the specific task (error vs. gain).
2. Implementation cost (including token‑usage pricing, compute, and training).
3. Human‑AI interaction dynamics (how people interpret and act on AI recommendations).
Encouraged participants to partner with evaluators (e.g., J‑PAL) to design pilots that generate robust impact data before scaling.

2. Transition to Panel Discussion

Moderator Murugan Vasudevan introduced the four panelists, highlighting each person’s blend of “Samaj, Sarkar, Bazaar” (civil‑society, government, market) experience.
Brief logistical note: audience questions would continue to be submitted via the QR‑code Mentimeter poll.

3. Panel Discussion – “AI, Work, and Society”

3.1 Opening Reflections (Murugan)

Framed the central question: “Whose work is being transformed, who measures it, and who decides the direction?”
Prompted panelists to share a recent moment that re‑shaped their mental model of AI in work.

3.2 Recent “Aha” Moments

Speaker	Insight
Murugan	Launched Amul’s AI‑driven cow‑health advisory (feature‑phone, Gujarati voice) – realized AI can be a direct empowerment tool for women farmers, not just a productivity booster.
Elizabeth	Anthropic’s “co‑work” product (Claude‑Code‑style assistance) was built entirely by AI in ten days, illustrating a paradigm shift where AI creates AI‑powered tools. Also cited a physiotherapist interview showing age‑based resistance to AI adoption.
Becky	Noted gender‑ and age‑based barriers in health‑system AI roll‑outs; highlighted ongoing RCTs with J‑PAL & ID‑Insight to assess process and equity outcomes of AI deployment.
Shankar	Stressed that AI can create brand‑new gig‑economy opportunities (e.g., YouTube‑style content creation) that did not exist a decade ago, especially when language or access barriers are removed.

3.3 AI & the Indian Context

Anthropic data: India is 2nd globally in total AI‑tool usage but 101 / 116 per‑capita, meaning usage is concentrated in a small tech elite.
Task mix: 45 % of Indian AI queries involve computer‑mathematical tasks – the highest share worldwide, indicating strong demand for technical problem‑solving.
Productivity gains: Indian users reported 15× speed‑ups (e.g., a 4‑hour task reduced to 15 minutes) versus a 10× global average.

3.4 Equity, Inclusion & “Blue‑Dot” Concept

Blue‑dot metaphor (Shankar): an AI‑powered location‑aware service that finds users (jobs, scholarships, services) instead of users having to search. Examples given:
- Job discovery in Ghaziabad/Karnataka – from 10 listed jobs to 4,000 after AI‑enabled mapping.
- Disability scholarships reaching beneficiaries within minutes rather than months.
Discussion of who captures productivity gains: panelists agreed that benefit‑sharing mechanisms (fair wages, revenue‑sharing, skill‑upgrading) are essential to prevent widening inequality.

3.5 Policy & Measurement

FCDO/Becky: Emphasised inclusion (gender, disability, digital safety) as a pre‑condition for AI uptake; cited Harvard data showing women are 20 % less likely to use generative AI at work.
Evaluation Framework (Becky): Four‑stage model – (i) Model assessment, (ii) Product testing, (iii) Usage monitoring, (iv) Development‑impact evaluation.
J‑PAL & ID‑Insight: Ongoing RCTs to test AI tools in education, health, and labour markets; stress the need for real‑time feedback loops.

3.6 Growth vs. Efficiency

Elizabeth: AI can deliver efficiency gains (cost‑saving, speed) and growth opportunities (new services, markets).
Shankar: For low‑productivity economies like India, AI‑enabled “new possibilities” (e.g., language‑localized skill training) may outweigh pure efficiency arguments.
Murugan: Argues that building AI “infrastructure” (like roads) is crucial; iterative improvement will enable a “race to the top” rather than a “race to the bottom.”

3.7 Audience Q&A Highlights

Blue‑dot question: Could AI mediate matching of workers to jobs within a 30 km radius, especially for women or people with limited mobility? Panelists agreed this is feasible and already being piloted in a few districts.
Beneficiary identification: Elizabeth asked who truly benefits from productivity gains; consensus: small business owners, frontline workers, and ultimately end‑users if gains are shared via higher wages or lower service prices.
Evaluation speed: Panelists warned against “shipping and hoping”; instead, they advocated continuous A/B testing and rapid‑cycle evaluation enabled by AI‑generated data logs.

4. Closing Remarks

Moderator thanked panelists and the audience, reminded participants to keep submitting questions via QR code, and introduced the next session of the AI Impact Summit.

Key Takeaways

Evidence matters: The Ghana teacher‑hiring experiment showed automation can dramatically outperform augmentation, contradicting common intuition.
Cost vs. outcome trade‑offs are central; simple cost‑saving calculations can be misleading without understanding impact on quality (e.g., health outcomes).
AI usage in India is high in absolute terms but low per‑capita, concentrated in tech‑heavy jobs; nevertheless, Indian users achieve larger productivity gains (≈ 15×).
Inclusion is a prerequisite: gender, age, language, and disability barriers must be addressed before AI can deliver equitable benefits.
“Blue‑dot” services illustrate a new class of AI‑enabled matching platforms that can surface previously invisible opportunities (jobs, scholarships, services).
Policy frameworks need a four‑stage evaluation pipeline (model → product → usage → impact) to build trustworthy AI systems.
Growth vs. efficiency: AI should be leveraged both to expand the economic pie (new gig‑economy roles, skill‑training) and to streamline existing processes.
Rapid‑cycle pilots: Because AI evolves quickly, real‑time data collection and iterative testing are essential to keep policy decisions evidence‑based.
Benefit‑sharing mechanisms must be designed so that productivity gains translate into higher wages, better services, or broader social welfare rather than concentrated profit.
Collaboration across sectors (academia, industry, civil‑society, government) is vital to ensure AI for work is responsible, inclusive, and impactful.

See Also:

India AI Impact Summit 2026

Explorer

ai-in-work-humans-ai-or-both