The Machine That Learned the Meeting

This is the third article in "The Shape of the Next Decade," a series mapping what the next ten years of AI actually look like when you follow the dependencies instead of the headlines. Part one: The Half-and-Half. Part two: The Three Tiers.

On a Wednesday morning in Copenhagen, a product manager at a direct-to-consumer (DTC) furniture brand opens her sourcing dashboard and finds the work already done. Three Vietnamese factories, ranked by unit cost, compliance history, and projected lead time. Carbon estimates included. She reads the summary, checks the margin on option two, and clicks approve. The whole thing takes four minutes.

Six months ago, this took a week. She would send a brief to her sourcing coordinator in Ho Chi Minh City, a man named Minh, and he would disappear into the problem. Three days of phone calls to factory owners he'd known for years. Sample inspections. Negotiated lead times based on who owed whom a favor, who was overcommitting capacity, who had just lost a contract and was hungry. He'd come back with a recommendation and a paragraph of context she couldn't have gotten anywhere else: Option two is cheapest but they missed a window in March. Option three costs more but the owner answers the phone at midnight when customs holds a container.

The dashboard gives her the same three options. It does not give her the paragraph.

Minh's decade of supplier relationships left a data trail: purchase orders, delivery timelines, QA pass rates, pricing histories, compliance audits. That trail was structurable, and it got structured. The model learned which factories Minh kept recommending and reverse-engineered the quantifiable reasons why. What it couldn't absorb was the rest: which owners cut corners when they're under pressure, which ones answer the phone on a holiday, which relationships survive a missed payment. The knowledge that lived in phone calls and factory visits and ten years of showing up.

Copenhagen doesn't know what it lost. The dashboard doesn't have a column for things Minh knew that weren't in the data. The output looks complete. Three factories, ranked, with metrics. It looks like the same recommendation Minh would have sent. The difference only surfaces when something breaks and there's no one to call.

Meanwhile, Minh refreshes his inbox for the third time that morning. He has been doing this for three weeks, not because he expects a brief but because the habit hasn't died yet. He still has a title and a salary, though his team has shrunk from eight to three.

Neither scene is dramatic. The product manager isn't callous; she used to send Minh briefs directly, but the dashboard replaced the workflow and she stopped thinking about the person behind it. Minh isn't destitute. What happened between them is ordinary, which is precisely what makes it worth examining.

The first piece in this series described how artifact work, the reports, analyses, and comparisons that once filled knowledge-work days, compresses into prompts and reviews. The second described how access to AI splits into tiers, with different users gaining different kinds of leverage. This piece is about what happens when you connect those two observations across a supply chain. Compression doesn’t stay inside one office. It propagates outward, and the people it reaches last are often the ones who can absorb it least. But there’s a subtlety the displacement story misses: the knowledge transfer is lossy. The system that replaces a person captures the pattern of their decisions, not the judgment behind them. It looks more efficient. It is also more fragile in ways that don’t show up on a dashboard.

Same wire, both directions

The development path that the World Bank has called "services-led growth" runs through exactly the work that AI compresses fastest. Outsourced analysis, offshore quality assurance (QA), contracted translation, managed service desks: these were never glamorous jobs, and the wages were often exploitative by the standards of the countries that sent the contracts. But they moved money, and not just wages. The Philippine business process outsourcing (BPO) industry alone employs 1.8 million people and accounts for roughly 9% of gross domestic product (GDP). Across Africa, BPO generates $8.85 billion in revenue and supports more than 1.2 million jobs. These industries also transferred process knowledge, supplier relationships, and institutional competence, the operational scaffolding that lets economies build the next thing.

AI systems were trained, in part, on the outputs of that labor. Kenyan data labelers working for outsourcing firms like Sama, a San Francisco-based AI training data company, earned $1.32 to $2 per hour tagging the content that taught models to classify, summarize, and generate text. The models now reduce demand for some of the work that created the training data. Data flows toward model owners, and value flows toward the firms with the integration capacity to deploy those models. But the transfer is degraded in both directions: the data that trains the model strips out the contextual judgment that made it useful, and the model that returns the output strips out the human relationship that made it trustworthy. The International Monetary Fund (IMF) estimates that 40% of global jobs are exposed to AI, but exposure is not evenly distributed: roughly 60% of jobs in advanced economies face it, compared to 26% in low-income countries. The gap sounds reassuring until you notice it means low-income economies are less exposed because they have fewer of the jobs AI affects, and the jobs AI does affect are disproportionately the ones those economies were building their next decade on.

Meanwhile, a basic-tier user in Nairobi or Dhaka gains real capability from the same system: a nurse gets better diagnostic support, a student gets a tutor that never sleeps. Help and harm travel through the same wire, and the capital that might fund a local AI business was supposed to come from the services revenue that AI is now compressing.

The asterisk on Jevons

The first piece in this series invoked Jevons paradox, the observation that when a resource gets cheaper to use, total demand often rises. Coal-fired engines got more efficient; Britain burned more coal, not less. The argument maps neatly onto AI: if custom software becomes ten times cheaper to build, ten times more of it gets built.

Jevons works when demand is elastic: make software cheaper, and latent demand materializes. The app that was never worth building at $200,000 becomes viable at $20,000. Make graphic design cheaper, and every small business gets a brand identity it couldn't previously afford. In these cases, the expanding pie genuinely creates new work, new roles, new firms.

Jevons breaks when demand is inelastic. Legal document review that costs a tenth as much doesn't produce ten times more legal document review. The same contracts get reviewed. Supplier comparison that takes four minutes instead of four days doesn't generate ten times more supplier comparisons. The same three factories get ranked. Translation of a technical manual into twelve languages at a fraction of the cost doesn't create demand for translation into a hundred and twenty languages. The same manuals get translated, and the budget line shrinks.

And there's a second problem Jevons doesn't account for: quality degradation that doesn't register as degradation. The supplier comparison that took four days came with judgment. The one that takes four minutes comes with metrics. Both look like answers, but only one knows which factory owner picks up the phone at midnight. The efficiency gain is real, but part of it is illusory. The output looks the same while quietly losing the context that made it reliable. That gap doesn't show up in a productivity chart. It shows up six months later in a missed delivery or a quality failure that nobody saw coming because nobody was left who would have seen it.

The work that developing economies absorbed through outsourcing sits disproportionately on the inelastic side. Data entry, document processing, back-office compliance: these are tasks with fixed upstream demand. The International Labour Organization's (ILO) own analysis finds that augmentation dominates over outright displacement in aggregate, but also that clerical support is the most exposed occupational category, with most tasks facing medium or high exposure to generative AI. Kenya's national strategy targets one million digital workers, even as the broader industry confronts significant automation pressure in business services.

There is a popular version of the ATM story that gets cited in these debates: ATMs were supposed to kill bank tellers, but the number of bank branches actually grew. Paul Kedrosky, a technology analyst, has shown that the branch expansion was driven by banking deregulation, not by automation-induced demand.

Tightening

Each cycle of the loop removes an intermediate step, and intermediate steps, once removed, rarely return. The pattern has a ratchet quality: the first pass compresses the task, the second pass eliminates the role, the third pass restructures the team so the role has no obvious place to reappear. But there's a fourth pass the organizational chart doesn't capture: the institution forgets what the role knew. Not just the person, but the tacit knowledge they carried, the relationships they maintained, the judgment calls they made that never appeared in any system. That knowledge isn't archived or saved somewhere recoverable. It's gone, and the institution doesn't know it's gone, because the dashboard still produces answers.

Policy, historically, arrives after the third pass. The General Data Protection Regulation (GDPR) took six years from initial proposal to enforcement. In the United States, the social media policy fight remains unresolved years after the harms became visible. This pattern of disruption first, regulation later is a structural feature of how democracies process novel technologies, and it means the first wave of disruption is absorbed by individuals and communities without institutional support.

AI also genuinely improves crisis response. Disease surveillance models and satellite imagery analysis accelerate epidemic detection and disaster assessment. Refugee logistics platforms optimize resource allocation in ways that save lives. These are not hypothetical benefits, but the loop means the system that improves crisis management can also intensify the upstream conditions (labor displacement and institutional weakening) that create crises in the first place.

What would break it

Here is what to watch over the next two to three years.

Evidence that would weaken the loop thesis: BPO employment in the Philippines, India, and East Africa grows in absolute terms despite AI adoption, suggesting Jevons dynamics dominate even in outsourced services. AI-powered microbusinesses in developing economies generate durable income at scale, not just isolated success stories but measurable shifts in regional employment data. Global South firms capture meaningful share of AI value, building models, not just labeling data for them. Basic-tier AI adoption measurably strengthens local institutions rather than substituting for them: clinics that use AI diagnostics also invest more in training, not less. Demand elasticity turns out to be higher than expected across outsourced work categories. The tacit-knowledge gap turns out not to matter much: quantifiable metrics prove sufficient for most sourcing, compliance, and service decisions, and the fragility predicted by the lossy-transfer thesis doesn't materialize in practice.

Evidence that would strengthen it: entry-level outsourced work contracts by 15% or more in major BPO markets while integration-tier firms report rising margins. Leapfrog stories (the nurse with a diagnostic app, the farmer with a market-price tool) remain individual anecdotes while the institutional ladders those individuals might have climbed weaken. Policy responses to AI labor displacement consistently lag three or more years behind measurable impact. AI training data demands increase while compensation for data laborers stays flat or falls. Supply chain failures increase in sectors where AI replaced experienced intermediaries, and post-mortems trace the failures to judgment gaps the automated systems couldn't detect.

Direction, not speed

The thing about this kind of displacement is that nobody has to want it. A product manager clicks approve because the dashboard looks complete, and a sourcing coordinator on the other side of the world loses his team because the data trail he spent a decade building turned out to be structurable without him. The decisions are rational at every step, and lossy at every step, in ways that the people making them can’t see from where they sit.

Markets optimize for cost without anyone coordinating the effort, which is why compression travels through supply chains so fast. Pushing back requires the opposite: funding, political will, institutional investment, all of which take longer to organize than a pricing algorithm takes to run. The knowledge that gets stripped out along the way doesn’t come back on its own, either. You have to decide it matters before it’s gone.

The rest of this series covers what that compression reshapes next: manufacturing patterns, trust architectures, energy systems, surveillance capabilities, security risks, transition paths, and crisis dynamics.

The product manager finishes her coffee. The sourcing coordinator refreshes his inbox one more time.