Pakistan’s BPO exports crossed $322M, up 19%. The same workforce training AI models for frontier labs is indistinguishable from a well-run call centre.
Pakistan’s call centre and BPO sector crossed $322 million in export revenue in the first eleven months of FY2025-26, up 19 per cent year on year.
More than 1,000 operations are registered with the Pakistan Software Export Board, with an estimated 500 additional medium and large centres operating independently. Nearly 90 per cent of them serve overseas clients. The workforce is young, English-proficient, and embedded in the process-managed, quality-graded, shift-based delivery model that international technology companies rely on.
What most of those companies have not yet recognised is that this infrastructure, which they built for customer support, claims processing, and back-office outsourcing, is almost identical to what the world’s largest artificial intelligence labs need to train their models.
The work is called data annotation and Reinforcement Learning from Human Feedback. Far from being exotic, it is systematic, high-volume, quality-managed evaluation work, exactly what a well-run call centre quality team already does. And the market for it is growing at compound annual rates of 26 to 32 per cent, with the RLHF segment alone projected to grow from $2.8 billion in 2025 to $18.6 billion by 2034. Pakistan’s BPO sector is structurally positioned to capture a significant share of that growth. The question is whether it moves before the opportunity consolidates around the countries that recognised it first.
What the Work Actually Is
Every response that ChatGPT, Claude, or Gemini generates was shaped, at some point, by human beings reading model outputs and deciding which ones were better, which ones were wrong, and which ones were harmful. This process, Reinforcement Learning from Human Feedback, is not a minor step in how large language models are built. It is how they learn to behave, and it runs continuously as models are refined, updated, and adapted for new domains.
In practice, an evaluator is presented with two or more model responses and asked to rank them against a rubric, flagging factual errors, assessing tone, and identifying outputs that violate guidelines. Rather than creative work, it is systematic, high-volume quality assessment, conducted in shifts, reviewed through layered checking, and graded for consistency across annotators. Any call centre team leader in Pakistan who runs quality monitoring on agent calls is already doing the cognitive work that RLHF annotation requires.
Beyond RLHF, the annotation market includes text classification, named entity recognition, audio transcription, and data cleaning tasks. These require English fluency and reliable process governance more than technical expertise, placing them within the existing competency profile of Pakistan’s BPO workforce. The higher-value end of the market, covering legal, medical, and coding annotation requiring domain knowledge, is accessible with deliberate investment in specialist recruitment.
Why Pakistan’s Workforce Profile Is the Right One
The global AI training data market did not develop around a new category of worker. It developed around a workforce that is English-speaking, process-disciplined, cost-competitive, and available at scale in countries with established relationships with American technology companies.
Countries with large BPO sectors absorbed AI annotation work organically, not through deliberate government policy but because the labour profile matched when the demand arrived. India’s $54 billion BPO industry, built over three decades on the same foundations Pakistan’s call centre sector rests on, now hosts companies like iMerit, with 5,500 annotators and over 2,000 language experts, running RLHF contracts for frontier AI labs. Infosys BPM and Wipro have added formal data annotation service lines. Specialised firms like Surge AI, founded in 2020 and profitable on a bootstrapped basis, have exceeded a billion dollars in annual revenue supplying training data to the same labs Pakistan’s IT companies serve.
None of this happened because India has something Pakistan lacks. It happened because India’s operators were positioned when the demand arrived and recognised what they were being asked to do. Pakistan’s call centre operators are, right now, in the position India’s were in roughly four years ago. The incumbency gap is real but it is not structural. The AI training data market is still growing fast enough that new entrants with the right workforce profile can secure meaningful contracts in the current cycle.
What the Pivot Requires
Pakistan’s call centre operators do not need to build new infrastructure from the ground up to enter this market. They need three things.
The first is deliberate repositioning. An operation that currently sells customer support should understand that it can immediately offer supervised AI evaluation, text annotation, and quality assessment to the same North American technology companies that are its existing client base. The delivery model is the same; the task specification is different. Firms that begin actively bidding for annotation and RLHF contracts, rather than waiting for inbound demand, establish themselves before the market settles on preferred suppliers.
The second is investment in annotation-specific quality protocols. AI labs that commission annotation work track inter-annotator agreement, error rates, and batch consistency with precision. This is not more demanding than call centre quality management; it is differently specified. The investment required is in calibration tools and rubric-design capability, areas where the existing quality management experience in Pakistan’s call centre sector transfers directly with targeted training.
The third is connectivity reliability. Annotation work is data-heavy and requires consistent broadband access. Pakistan’s dependence on a small number of submarine cable landing points at Karachi, underscored by the SMW5 fault in July 2026, represents a genuine service-continuity risk when marketing uninterrupted delivery to international clients. Operators need documented redundancy arrangements, and the government’s push to fast-track the Pakistan-Saudi digital corridor, with its mandate for alternate submarine routes and terrestrial links, directly addresses this constraint.
What This Means
Pakistan’s BPO industry stands at the same inflection point its software export sector stood at when it first began selling offshore development to American companies. The capability is there, the market is growing, and the only question is whether Pakistani operators pursue the opportunity deliberately or discover it years later when the obvious entrants have already locked in the major contracts.
The $322 million in call centre exports is the foundation, not the ceiling. As AI compresses the volume of routine customer interactions that need human handling, the category ceiling for conventional customer support will face pressure. The category ceiling for AI training data will not face the same constraint, because the demand for human evaluation of model outputs grows with every new model released and every new domain those models are adapted for.
The government’s enabling role includes PSEB registration frameworks that explicitly classify AI training data as an IT export vertical, fiscal incentives aligned with annotation and RLHF service companies, and the connectivity investment that makes continuous delivery credible to enterprise clients. Industry’s role is equally clear. Pakistani BPO operators should be packaging their quality management capability, their English-language workforce, and their existing North American client relationships into annotation service offerings and taking those offerings to market now, the same way Pakistan’s software houses took their first offshore development proposals to market two decades ago.
The infrastructure exists. The workforce profile matches. The market is open. What Pakistan’s BPO sector does next will determine whether it is building the next phase of its export story or defending the last one.
