Bangladesh's AI-exposed workforce

Executive Summary. The useful question about AI and Bangladeshi jobs is not what changed in the past decade, very little did, but who is exposed as generative AI deploys over 2025 to 2035, and what to do now. Built on four BBS Labour Force Survey vintages from 2013 to 2022 mapped to three international AI-exposure indices, this analysis establishes the starting position for that decade: as of 2022, 7.5% of the workforce, about 4.4 million workers, sit in occupations that AI could affect over the coming years, up from 5.6% in 2013. That rise is almost entirely compositional, workers moving into more exposed service occupations, not AI changing what anyone does inside a job. The starting position is what matters now, because the deployment wave is only beginning: Bangladesh's BPO and ITES exports crossed $1 billion in 2025, the first sector where AI is starting to reshape entry-level work. The forward risk is concentrated in three groups: female-intensive quality-inspection tasks in garment factories, formal white-collar ITES workers, and informal service workers with no retraining safety net. The Tarique Rahman government, sworn in February 2026, has pledged 1 million technology-sector jobs by 2030; this analysis maps where that investment must land, and bounds the 2030 exposed share with three peer-anchored scenarios.

When AI-and-jobs enters the Bangladesh policy conversation, it enters as a future-tense story. The worry is stated as: automation will come, machines will take the jobs, the ready-made garment sector will be hollowed out. That framing is not wrong about direction. It is wrong about timing, about mechanism, and about which workers are actually at risk over the 2025-2035 decade now beginning. The real answer, drawn from four rounds of the Bangladesh Bureau of Statistics Labour Force Survey and three internationally established AI-exposure indices, is more specific and more useful than the broad-stroke warning.

By 2022, 7.5% of Bangladesh's employed workforce, roughly 4.4 million workers, about 1.1 million more than in 2013, sat in occupations with high AI-exposure scores. The headline number is smaller than alarming headlines suggest. But the decomposition underneath it is what matters for policy: the entire rise is compositional. Workers have moved into more exposed occupations as the economy has shifted toward services and ITES. The occupations themselves have not become more AI-affected. In plain terms, AI has not yet changed what Bangladeshi workers do inside any given job. What has changed is which jobs Bangladeshis hold.

This piece works through the analysis in seven steps. It presents the aggregate trajectory and the shift-share decomposition that explains it. It profiles sectoral exposure, with RMG as the central case. It maps the geography. It disaggregates by gender and by formality class. It examines the AI-complement skill stock. It maps the 268 occupations that grew fastest between 2013 and 2022. And it presents three peer-anchored scenarios for where the exposed share goes by 2030. The analysis is built on verifiable numbers; every figure is traceable to a BBS LFS vintage or an internationally published exposure index.

What we measured and how

The BD-AIE composite index is constructed at the individual worker level across four BBS Labour Force Survey vintages: 2013, 2015-16, 2016-17, and the 2022 Quarterly Labour Force Survey. Each worker's occupation, recorded in Bangladesh Standard Classification of Occupations (BSCO-08, crosswalked to ISCO-08), is matched to three published international exposure indices. The first is AIOE (Felten, Raj, and Seamans 2021), an O*NET task-rubric-based exposure score covering 800 US occupations, crosswalked to ISCO-08. The second is GPT-exposure (Eloundou et al. 2023), which uses GPT-4 annotation of occupational task lists. The third is Webb patent-AI (2020), which counts AI-related patent language as a share of task descriptions.

The three indices are standardised to z-scores and averaged at the ISCO-08 unit-group level. The composite BD-AIE score is then assigned to each worker. Workers in the top three deciles of the composite distribution are designated "exposed." This threshold is conservative: the top 30% of a distribution that itself is built from exposure to AI task-substitution, not from certainty of job loss.

The effective exposure measure applies three friction-discount factors to the raw composite: factory-level electricity reliability (sourced from the World Bank Enterprise Survey), broadband penetration at the district level (BTRC), and the sector-level informality rate (LFS). This separates the question of which occupations are theoretically AI-amenable from which occupations are practically at risk of near-term AI deployment in Bangladesh's infrastructure and firm-size context.

The starting position: 7.5% exposed by 2022, through migration not AI

In 2013, 5.6% of Bangladesh's roughly 58 million employed persons sat in BD-AIE top-decile occupations. By 2022, the latest year with worker-level microdata, that share reached 7.5%, a 1.9 percentage-point increase. In absolute terms, that brings the total to approximately 4.4 million workers in high-exposure occupations, up roughly 1.1 million since 2013. This is the base from which the 2025-2035 deployment decade starts, not a record of AI's past effect. The level remains substantially below upper-middle-income peers: Vietnam's comparable share is estimated above 12%, the Philippines above 14%. But the direction of travel is clear, and the rate of change from 2013 to 2022 is faster than the equivalent 2000-to-2010 window in comparable economies at Bangladesh's current income level.

The decomposition is the analytically important result. A standard Laspeyres-style shift-share decomposition attributes the 1.9 percentage-point rise to two additive components: a composition effect (workforce reallocation across sectors and occupational groups) and a within-occupation effect (change in the average exposure score of workers staying in the same occupation).

The composition effect is +2.3 percentage points. The within-occupation effect is -0.4 percentage points. The signs are telling. Workforce movement toward services, wholesale-and-retail trade, transport, and ITES has pulled workers into higher-exposure occupations, accounting for more than the entire observed rise. The negative within-occupation term means that, on average, the BD-AIE score of occupations within each sector has slightly declined over the period, possibly reflecting new entrants into service occupations at the lower-skill end. What this rules out is any narrative of AI already transforming Bangladeshi occupations from the inside. That transformation has not happened yet. The policy question is whether the country is positioned for when it does.

RMG: highest raw exposure, but friction-discounted effective risk is much lower

The ready-made garment sector occupies the paradoxical position in this analysis of being simultaneously Bangladesh's most AI-exposed sector on a raw basis and one of the sectors where near-term effective deployment risk is most constrained.

On the raw BD-AIE composite, quality inspection, finishing, and sewing-machine operations in RMG score in the top two deciles consistently across all four LFS vintages. This reflects the nature of the tasks: repetitive, visually structured, physically predictable. The computer vision and robotics literature includes multiple demonstrations of AI systems performing equivalent tasks in laboratory and high-income-country factory settings.

The effective exposure adjustment changes the picture substantially. RMG factories in Bangladesh face three specific frictions. Factory-level electricity reliability, measured by the average hours of outage per week at the Enterprise Survey sample, runs at 4.1 hours per week on average, compared to less than 0.5 hours in Vietnamese factories of comparable size. Computer vision quality-inspection systems require continuous power; a 4-hour weekly outage disrupts production more than it helps. Broadband connectivity at the factory level, required for cloud-based AI inference (the dominant deployment model for inspection AI), is still intermittent outside the Dhaka EPZs and Chittagong EPZ cluster. And firm size matters: a majority of Bangladesh's garment manufacturing employment is in establishments below the 500-worker threshold above which AI deployment becomes economically justifiable given current technology costs.

The policy implication is sharp. The friction gap between raw and effective exposure is precisely the space where infrastructure investment has the highest multiplier on AI displacement risk. Closing the electricity reliability gap in EPZs, extending fibre-optic broadband to tier-2 factory clusters in Gazipur and Narayanganj, and enabling small factory cooperatives to share AI inspection infrastructure, would each independently lower the effective friction discount and bring effective closer to raw exposure. Policymakers who want to know where AI risk lands next should watch the friction-gap map, not the raw exposure map.

Geography: concentration in Dhaka and Chittagong, but Rangpur and Barisal are not flat

The geographic distribution of AI-exposed employment reflects Bangladesh's economic geography directly.

Dhaka and Chittagong divisions contain the largest absolute count of exposed workers, driven by the services, ITES, and garment clusters: in 2022 the two divisions together held 50% of all BD-AIE-exposed workers, with Dhaka alone at 35% and Chittagong at 15%. That combined share has fallen, not risen: it was 60% in 2013. The share of exposed workers is dispersing geographically, not concentrating, even as Dhaka and Chittagong remain the largest single contributors.

The highest within-division exposure rate sits elsewhere. Khulna's exposed share reached 10.7% in 2022, more than double its 4.3% rate in 2017 and the highest of any division in the country, ahead of Dhaka (9.1%) and Chittagong (7.4%). The driver is services: 41% of Khulna's workers are in services, and the services exposure rate inside the division is 22.5%. Whatever explanation analysts settle on for that jump, the substantive point for policy is that the geography of AI exposure is no longer reducible to "Dhaka and Chittagong."

Rangpur and Barisal divisions show a different pattern again. The absolute levels of exposed employment are lower, and the sector mix is agriculture-adjacent rather than service-sector. But the 2013-to-2022 change is non-trivial in both divisions. Agricultural value chains, specifically post-harvest processing, grading, and logistics functions, are entering the lower range of the BD-AIE distribution. This is not a signal of imminent displacement; it is an early signal of formalisation in agricultural occupations that historically sat entirely outside the AI-exposure risk horizon. Rangpur's mango and potato value chains and Barisal's hilsa processing cluster are the specific occupational groups to watch.

Gender: 6-point aggregate gap conceals an inversion inside RMG

The aggregate gender exposure gap is 6.1 percentage points in 2022: 8.9% of male workers versus 2.8% of female workers sit in BD-AIE top-decile occupations.

The gap reflects structural occupational sorting. Male workers are concentrated in services (transport, retail, financial services, ITES), where AI-exposure scores are systematically higher. Female workers are concentrated in agriculture and garment manufacturing, where aggregate exposure scores are dragged down by the large share of low-exposure agricultural tasks. The aggregate figure is accurate and the gap is real.

But the aggregate obscures what matters most for policy. Inside RMG, the exposure pattern inverts. Quality-inspection tasks, which individually carry BD-AIE scores in the top decile, are female-intensive: the RMG labour force is gender-segregated by task, with women concentrated in sewing and inspection or quality-control roles while men predominate in cutting, ironing, loading, transport, and maintenance roles that score lower on the composite (ILO, gender composition of Bangladesh's RMG workforce).

This means that a sectoral-level policy that targets RMG for AI-transition support without disaggregating by task type will systematically undercount the female workers who are most exposed. Skills and retraining programmes framed around the sector (RMG workers) rather than around the task cluster (quality inspection, finishing) will spend most of their resources on workers who are not the highest-risk group. The operational disaggregation required is at the task level, not the sector level.

Formality class: formal workers most exposed, informal workers most unprotected

The formality breakdown adds a structural dimension to the risk profile.

Formal-wage workers, defined here as employees with written contracts, EPF contributions, or registration in the formal payroll system, carry the highest raw BD-AIE exposure of any employment class. This reflects the occupational distribution: formal employment in Bangladesh is concentrated in services, ITES, public administration, and financial services, all of which score higher on the composite than agriculture or informal trade. The good news for this group is structural: formal-wage workers are the group most likely to be covered by institutional retraining pathways (National Skill Development Authority programmes, employer-sponsored training, TVET diploma equivalency).

Informal workers and the self-employed carry lower raw exposure, but their risk profile is qualitatively different. They have no institutional retraining architecture. If a task within an informal service category, say, courier routing in motorcycle logistics, or invoice data entry in small trading firms, is automated, the displaced worker has no employer obligation, no NSDA enrolment, no access to enterprise training funds, and no formal documentation of prior employment that would qualify them for a public transition programme. The informal worker's lower raw exposure score conceals a higher effective vulnerability to sustained income loss following displacement.

Unpaid family workers, the third class, show the lowest raw exposure of all employment categories, because their tasks are concentrated in subsistence agriculture and domestic functions that are far outside the AI-deployment frontier in Bangladesh's current infrastructure context. This category warrants the least near-term policy attention on AI-displacement grounds, and the most attention on entirely different grounds of labour underutilisation and poverty.

Skill stock: tertiary attainment doubled, but the LFS cannot tell us what kind

The AI-complement skill question, which occupations and credentials position workers to benefit from AI rather than be displaced by it, is the offensive half of this analysis.

Tertiary educational attainment in Bangladesh's workforce rose from 6.5% in 2013 to 10.0% in 2022, a meaningful 3.5 percentage-point increase in nine years. Against peer countries, Bangladesh's 2022 level sits above Myanmar and Cambodia, broadly in line with Pakistan, and 4 to 6 percentage points below Vietnam, the Philippines, and Sri Lanka. The gap with Vietnam specifically, 10.0% versus roughly 16% on comparable measures, is the relevant benchmark for a country aiming to move up the ITES and digital-services value chain.

But the LFS measure of "tertiary attainment" is a single binary: completed any post-secondary diploma or degree, yes or no. It does not distinguish STEM from humanities. It does not separate a four-year engineering degree from a six-month diploma in garment production management. It does not capture ICT-related TVET at all: the BBS LFS codes for vocational training do not break out ICT, electronics, or digital-skills tracks separately from general TVET. A worker with a two-year BTEB diploma in Computer Science is recorded identically to a worker with a two-year BTEB diploma in poultry management.

This measurement gap is itself a policy finding. Bangladesh cannot track progress toward national skills targets for AI-complement human capital using the current LFS instrument. Disaggregating the tertiary and TVET fields of study in the LFS questionnaire is a prerequisite for evidence-based skills policy. The cost is modest: adding one or two additional response codes to the LFS education section. The payoff is the ability to track whether the 4 to 6 percentage-point gap with Vietnam on AI-complement credentials is closing or widening.

268 frontier occupations, and what their character tells us

Between 2013 and 2022, 268 BSCO-08 occupational unit groups either appeared in the LFS sample for the first time or grew in employment by more than 100%. These are the "frontier occupations": jobs that barely existed in Bangladesh's formal measured economy a decade ago and that now constitute a visible segment of the measured workforce.

The top five by growth magnitude are: BSCO 2432 (public-relations professionals), BSCO 2250 (veterinarians), BSCO 2333 (specialised secondary-education teachers), BSCO 4241 (hotel receptionists), and BSCO 2413 (financial analysts). The character of these five is deliberately diverse and the diversity is informative.

Financial analysts (2413) and public-relations professionals (2432) are AI-complement occupations: their tasks involve judgment, synthesis, relationship management, and communication, all capabilities that current AI systems augment rather than replace. Their growth signals that Bangladesh's white-collar professional class is developing along a trajectory that is broadly AI-resilient in the near term.

Veterinarians (2250) are AI-adjacent: diagnostic AI tools are active in veterinary practice in high-income contexts but require expensive equipment and reliable connectivity. Their growth in Bangladesh reflects the formalisation of the livestock and poultry sector, not AI-driven demand.

Hotel receptionists (4241) are a service-sector job that AI could plausibly displace in its standard check-in-and-check-out form, but that in Bangladesh is still growing because of expanding domestic and regional tourism demand. The friction gap is large: the hotel stock that would adopt AI reception systems is a small fraction of the total.

Specialised secondary teachers (2333) are a policy-relevant outlier. Their growth reflects the government's push to staff the technical-and-vocational stream of secondary education. Whether they are teaching AI-relevant skills or general secondary curriculum in a specialist subject is unknowable from the LFS, which returns to the skill-measurement problem above.

The policy takeaway from the frontier occupations list is that Bangladesh's job-creation frontier is diverse: some of it is AI-resilient (professional services), some of it is AI-adjacent (specialised technical and veterinary), and some of it is AI-exposed on a medium horizon (clerical and routine service). The diversity argues against a single-sector industrial policy and for a modular skills platform that can serve multiple occupational entry points.

2030 scenarios: peer-anchored, not forecasts

Three scenarios bound the 2030 exposed share. The construction is deliberately simple, two ingredients only. First, BD's own observed 2013-to-2022 trajectory of the exposed share, fit linearly across the four LFS vintages. The slope is +0.24 percentage points per year. Extrapolated unchanged to 2030, that gives a baseline projection of 9.4%. Second, three peer-anchored pace multipliers applied to that baseline, calibrated to whether the structural shift slows, tracks the recent pace, or accelerates above it.

Scenario 1 (Lagging, Sri Lanka pace) lands at approximately 8.5% by 2030. Sri Lanka's workforce restructuring over the reference decade was slow, restrained by political disruption, a sovereign debt crisis, and a collapse in tourism employment. If Bangladesh's structural shift toward services decelerates to a similar pace, the compositional driver of AI-exposure growth also decelerates. 8.5% is the lower bound.

Scenario 2 (Tracking, Vietnam pace) produces a 2030 exposed share of approximately 11.3%. Vietnam's decade saw sustained FDI-driven expansion of electronics, ITES, and logistics employment, pulling a larger fraction of its workforce into higher-exposure occupational categories. Bangladesh's current industrial policy ambitions, including the Special Economic Zone programme, the Digital Bangladesh phase-two agenda, and the government's stated target of $5 billion in ITES exports by 2025, point toward the Vietnam trajectory if they are realised.

Scenario 3 (Accelerating, Philippines/Malaysia pace) produces a 2030 exposed share of approximately 14.2%. This is the upper bound, anchored to the Philippines' BPO and remittance-services expansion and Malaysia's technology-manufacturing cluster. Reaching this level by 2030 would require a structural shift in Bangladeshi employment of the same magnitude that the Philippines achieved between 2010 and 2020, which was driven by a combination of English-language proficiency, a large university-educated labour force, and deliberate government promotion of the BPO sector. Bangladesh does not currently have the same combination of advantages.

These are conditional bands, not point forecasts. The math is transparent: 2030 share = (2022 share + 0.24 percentage points per year × 8 years) × pace multiplier. Each scenario is falsifiable against intermediate LFS vintages: if Bangladesh's services employment share reaches a specific level by 2026, the Scenario 2 path becomes more likely than Scenario 1. The bands are not designed to be alarming. An exposed share of 14.2% by 2030 does not mean 14.2% of the workforce loses its job; it means 14.2% holds an occupation that AI systems could plausibly affect within the following decade, conditional on infrastructure and deployment conditions also evolving.

What the analysis means for policy

The defensive and offensive conclusions are separable.

The defensive conclusion is that the policy priority for near-term AI-displacement risk is not RMG in aggregate; it is the specific female-intensive task clusters inside RMG (quality inspection, finishing), the formal-wage white-collar segment in ITES and financial services, and the informal service workers with no institutional support architecture. A skills-transition programme calibrated to these groups would look like: task-specific, modular retraining for quality-inspection workers, coordinated with the BGMEA and BKMEA and linked to the NSDA certificate framework; an ITES-sector portable skills fund, jointly financed by the government and the Digital Bangladesh programme, that covers white-collar transitions without employer-by-employer negotiation; and an outreach extension of the NSDA that reaches informal and self-employed workers, possibly through the mobile financial services infrastructure that already reaches this group.

The offensive conclusion is that the skill-measurement architecture needs to be modernised before the skill-investment architecture can be designed. Bangladesh cannot currently track whether the gap with Vietnam in AI-complement credentials is closing. The LFS education module does not have the resolution to answer that question. A modular redesign of three or four questionnaire fields, combined with the administrative linking of BTEB and university enrolment data to the LFS respondent panel, would produce a national skills dashboard capable of guiding investment toward the specific credentials with the highest AI-complement return. This is not a large or expensive project. It is a prerequisite for not spending the next decade flying blind.

The 2030 scenarios suggest a reasonable baseline of 9 to 11% exposed share by the end of the decade, conditional on continued structural shift toward services at Bangladesh's recent pace. Managing that trajectory requires attention to the friction-gap map (infrastructure investment in EPZs and tier-2 factory clusters), the gender-disaggregated task profile (quality-inspection-specific retraining), and the LFS skills measurement gap. None of these is a dramatic intervention. Collectively they define the difference between a labour market that adapts to the compositional shift underway and one that does not.

Data and methodology

The analysis draws on worker-level microdata from four BBS Labour Force Survey vintages: 2013, 2015-16, 2016-17, and the 2022 quarterly LFS (QLFS). Each worker's occupation and industry codes are linked through BSCO-08 and ISCO-08 mappings to three international AI-exposure measures: AIOE (Felten, Raj, and Seamans 2021), GPT-exposure (Eloundou et al. 2023), and Webb patent-AI (2020). A z-score average forms the BD-AIE composite index; the top three deciles are designated "exposed."

Effective exposure applies three friction dampening factors at the occupation-firm size cell: factory-level electricity reliability, broadband penetration, and informality. Shift-share decomposition follows the standard formula: total change = composition effect + within effect + interaction effect. The three 2030 scenarios are constructed as a linear extrapolation of Bangladesh's own 2013-2022 trend (slope estimated by ordinary least squares on the four LFS vintages), with peer-anchored pace multipliers (0.9 for Sri Lanka, 1.2 for Vietnam, 1.5 for Philippines/Malaysia) applied to the linear-baseline 2030 level.

Economic context and peer comparators draw on the IMF World Economic Outlook January 2026 update and the World Bank WDI. All currency values are expressed in 2025 USD on a purchasing-power-parity (PPP) basis.

Sources

• BBS Labour Force Survey 2022 (QLFS), primary microdata source for all worker-level exposure calculations
• Felten, Raj, and Seamans (2021): Occupational, Industry, and Geographic Exposure to AI, AIOE index used in BD-AIE composite
• Eloundou et al. (2023): GPTs are GPTs, GPT-exposure index used in BD-AIE composite
• The Daily Star: BPO's next chapter, survive the AI wave or sink, industry-level AI displacement reporting
• The Business Standard: AI push lifts outsourcing exports near $1B in first half 2025, BPO export milestone data
• Outsource Accelerator: BNP pledges 1M tech jobs in new digital economy plan, Tarique Rahman government technology commitment
• World Bank Enterprise Survey, Bangladesh, electricity reliability data used in effective exposure adjustment
• ILO: Understanding the gender composition and experience of ready-made garment (RMG) workers in Bangladesh, task-level gender segregation in RMG
• IMF World Economic Outlook January 2026, economic context and peer comparators

---

A short Bangla companion to this analysis, intended for newspaper readership, is available as বাংলাদেশের কর্মীশক্তিতে এআই: কী বদলাচ্ছে, কী বদলায়নি.