On 24 July 2026, on two continents, the least-visible workers in the AI economy stopped being invisible to the law. In Nairobi, Kenya's Ministry of Information, Communications and the Digital Economy opened public consultation on a draft AI policy carrying the first labour provisions we can find written specifically for data annotators, content moderators and AI quality evaluators. In Bengaluru, the platforms deposited the first money ever collected under Karnataka's gig-workers Act. Neither side was watching the other. Both were answering the same question: who pays for the human work that makes AI work.
Be exact about what Kenya has done, because the coverage is not. This is not a rulebook and the rules are not in force. It is a draft policy, open for comment until 4 August, that proposes a fair-pay reference framework benchmarking annotation and moderation wages against international rates, duty-of-care obligations on employers whose staff face prolonged exposure to disturbing material, and a commitment that the government will later publish occupational protection guidelines on written contracts, psychosocial support and grievance procedures. A draft that promises guidelines is two removes from an enforceable floor. It is still the furthest any government has gone.
The number underneath it is the reason it matters. Kenyan data workers earn between $1.46 and $3.74 an hour, against $21 to $27 for the same work in the United States, on figures carried in Kenyan government-linked reporting. A KICTANet-affiliated study found 27% of surveyed moderators reporting significant mental-health distress including anxiety and post-traumatic stress. And 183 former moderators, Daniel Motaung plus a group action of 182 more, are still waiting on a ruling against Meta in Kenya's Employment and Labour Relations Court, four years after the first case was filed.
The Bengaluru half is smaller, harder and further along. Karnataka's Act is real law, passed in 2025 and notified in November. On 24 July the Internet and Mobile Association of India and its members Eternal, Zepto, Swiggy, Urban Company and Meesho deposited the welfare levy the High Court had ordered three weeks earlier. No amount was disclosed, and we will not invent one. The levy is 1% of the worker payout, capped at 50 paise per two-wheeler ride, 75 paise per three-wheeler and one rupee per four-wheeler, with 1% on food and grocery. Karnataka's own estimate of the workforce it covers is about four lakh.
The capped welfare levy on a two-wheeler ride that Karnataka's platforms started paying on 24 July, into a court registry rather than a welfare fund, for a workforce the state estimates at about four lakh -- the same day Kenya opened consultation on pay floors for the people who label AI's training data at $1.46 an hour.
On 24 July 2026 Kenya's Ministry of Information, Communications and the Digital Economy opened public consultation, closing 4 August, on a draft AI policy proposing a fair-pay reference framework that would benchmark data-annotation and content-moderation wages against international rates, duty-of-care obligations on exposure to disturbing material, and occupational guidelines the government commits to publish later; the same day the Internet and Mobile Association of India and its members Eternal, Zepto, Swiggy, Urban Company and Meesho deposited the first welfare levy ever collected under Karnataka's gig-workers Act, in an amount nobody has disclosed, into the High Court registry rather than a welfare fund, after Justice M Nagaprasanna refused a stay on the ground that the money is not demanded as charity but in terms of promulgation of a legislation.
For three years the argument about AI and work has been conducted almost entirely in the countries that buy AI, not the ones that build its inputs. Issue 016 read the West's narrative about job losses. Issue 018 printed the corridor's data, AI revenue rising while hiring fell from Hyderabad to Dubai. This fortnight the third move arrived, and it came from the supply side: the geographies bearing the cost started writing the institutions that price it. That is a different kind of event from a policy paper. It is the beginning of a bargaining position.
Three governments, three postures, one object. Kenya proposes to price AI's human inputs, setting a reference wage for the labelling and moderation work that sits underneath every frontier model. Karnataka taxes the platforms that manage work algorithmically, a paise at a time, and has a court enforcing the levy while the constitutional challenge runs. The UAE does neither: it uses AI to enforce labour law rather than to protect labour from AI, with MoHRE reporting 95% of mandated companies hitting their H1 Emiratisation targets, AI tools flagging fake-Emiratisation schemes, and nearly 16 million digital transactions run through what it calls an AI-powered service ecosystem in the same half-year.
The Western cycle this fortnight was about none of this. It was about model security and openness: an agent breach, an open-weights letter, a new Chinese release, an alliance formed to argue about who may open-source what. That is a real argument. It is also an argument among the people who own the models, about the models. The Global South spent the same fortnight arguing about the people. Both arguments are about power over AI. Only one of them is being had by the side that supplies the labour.
The honest shape of the shift is small and early. A draft in consultation for one week more. A levy sitting in a court registry. An enforcement regime that automates the state rather than protecting the worker. None of it has yet changed a single payslip that we can verify. But institutions are how bargaining power gets stored, and until this month the supply side of the AI economy had almost none to store it in.
Kenya proposes to price AI's human inputs by benchmarking annotation and moderation pay against international rates, Karnataka taxes algorithmic platforms a paise at a time with a court enforcing the levy while the constitutional challenge runs, and the UAE does neither, using AI to enforce labour law rather than to protect labour from AI, with MoHRE reporting 95% of mandated companies meeting their H1 Emiratisation targets, AI tools flagging fake-Emiratisation schemes and nearly 16 million digital transactions run through what it calls an AI-powered service ecosystem; three governments, three postures, one object, and none of it yet visible in a single verifiable payslip.
The counterargument is strong and we should put it first. Kenya has published a draft that promises guidelines, and the same ministry launched a BPO policy one week earlier aimed at growing the very sector these protections would make more expensive. Karnataka's money is in a court registry, not a welfare fund, with no disbursal scheme yet built, an amount nobody has disclosed, and a bench that has already conceded on a peripheral examination it does look like the state Act runs repugnant to the Centre's Social Security Code. Fifty paise a ride against a workforce the state puts at four lakh is not redistribution. On this reading, nothing has happened.
The reply is that something did happen, and it is worth naming precisely. A statutory hook now exists where none did. Justice Nagaprasanna's refusal to stay the levy turned on a distinction that will outlast this case: the amount is not being demanded as a charity, it is demanded in terms of promulgation of a legislation. He also told the platforms that if the money is not going to gig workers he would stay the Act immediately, which is a judge putting the state on notice, not just the companies. And the platforms are now complying and contesting at once, which is what it looks like when an obligation becomes real enough to litigate rather than ignore.
The tension inside Kenya's own month is the fairest test of all of this, and we do not think it resolves cleanly. On 16 July the government launched a BPO policy to grow digital outsourcing work. Eight days later it opened consultation on rules that would raise the cost of exactly that work. Job growth and job quality, attempted in the same month by the same state. Most governments pick one and pretend the other follows. Kenya is at least attempting both in public, which is more than the buyers of this labour have done.
For a worker or a board reading from either end, the practical point is the same. Rules follow power, and power follows organisation. The fact that AI-economy labour now has statutory hooks in Bengaluru and a draft with a consultation clock in Nairobi, however thin both are, changes who the next negotiation starts from. Watch 4 August, when Kenya's consultation closes, and 14 August, when the Karnataka challenge is heard. Both fall inside the next fortnight, and both will tell us more than any of this month's announcements did.
Nothing here has reached a worker yet, and we should say that first: Kenya's provisions are a draft promising guidelines, Karnataka's money sits in a court registry with no disbursal scheme built, and the bench has already conceded the state Act may run repugnant to the Centre's Code. But a statutory hook now exists where none did, the platforms are complying and contesting at once, and Kenya is attempting job growth and job quality in the same month rather than pretending one delivers the other. Rules follow power. The supply side of the AI economy just acquired somewhere to store it.
This issue's panel is a scorecard rather than a layoff split, because the fortnight's headcount news was steady rather than dramatic: Infosys closed the June quarter down 532, now confirmed against the company's own fact sheet at 328,062 from 328,594 rather than carried as search-corroborated reporting as it was in 018, and Cognizant closed the June quarter down about 900 to 356,700, with attrition rising to 13% and most large deals characterised by its finance chief as AI efficiency plays. Tech Mahindra cut 863 to 146,760 while lifting profit 28.4% and putting Perplexity across its own sales organisation. The pattern 018 named has not reversed. It has become unremarkable, which is its own kind of news.
One correction belongs on this page, because we printed it wrong. In 018 we attributed Wipro's Q1 FY27 headcount of 243,044 to the company's press release, which does not carry the figure, and we carried a three-issue open question about whether the underlying move was organic. Wipro's own datasheet resolves the arithmetic: the closing series runs 243,044 against 242,156 the previous quarter, so the quarter-on-quarter move is plus 888. The widely reported plus 9,812 is the year-on-year change against Q1 FY26's 233,232, mislabelled as a quarterly addition. The organic-versus-acquired split cannot be closed at all, because Wipro publishes no components. We are calling that permanently ambiguous and will stop chasing it.
Every rising role this issue comes from the same place: somebody has to operate the rules now being written. Kenya's draft would make employers benchmark annotation pay against international rates and disclose their structures, which is a compliance file nobody currently owns. Its duty-of-care provisions would make psychosocial support an operating requirement rather than a benefit. Karnataka's levy has to be turned from collected paise into claimable benefits by an administrator who does not yet exist, and reconciled against a central Code that may or may not override it. MoHRE's inspection stack has to be read by someone who understands what it actually detects. Three shores, one job family, none of it staffed.
The individual move is unglamorous and reliable. Governance work is counter-cyclical to the thing it governs: the more contested the rules, the more valuable the person who can operate them. If you sit in trust and safety, in gig-platform operations, in outsourcing delivery or in HR compliance anywhere on these three shores, the next eighteen months will reward reading the instruments themselves rather than the coverage of them. Start with the two documents that resolve inside a fortnight, Kenya's draft policy and the Karnataka pleadings, and you will be ahead of almost everyone advising on either.
6 rising role categories, each with a sourced hiring signal.
Kenya's draft would make employers benchmark annotation and moderation pay against international rates and disclose their pay structures, while 183 former moderators still wait on a ruling against Meta in the labour court. Somebody has to build that disclosure file before it is demanded.
Named Meta
A KICTANet-affiliated study found 27% of surveyed moderators reporting significant mental-health distress, and Kenya's draft would put a duty of care on the employer. Psychosocial support stops being a benefit and becomes an operating requirement.
Named KICTANet
Eternal, Swiggy and Zepto have deposited Karnataka's levy into a court registry, and the court was told no scheme for disbursal exists yet. Turning collected paise into claimable benefits is an unfilled job with a judge watching.
Named Eternal, Swiggy, Zepto
IAMAI argues the state levy duplicates the Centre's Code on Social Security and the bench conceded it may run repugnant. Reconciling one payroll against two overlapping statutes, before 14 August, is finance work nobody staffed.
Named IAMAI
MoHRE reports 95% of mandated companies meeting H1 targets, with AI tools flagging fake-Emiratisation schemes rather than policing quotas generally. Reading what the inspection stack actually detects, and what it does not, is now a compliance skill.
Named MoHRE
Ghana published a 10-year AI implementation plan and a $250m compute centre, and Google opened an Africa Applied AI Lab in Accra, in the same fortnight Kenya opened its consultation. Reading three national AI policies at once is becoming one person's job.
Named Google
Printed, not charted. These figures are not measured the same way, on any of the four counts that would let them share a scale. Drawing them together would suggest a comparison the sources do not support, so the numbers are set out instead.
Prev week (w/e 24 Jul)
This week (w/e 31 Jul)
This week's signal through the India, Middle East and Africa lens.
ACCELERATING
Signal
Kenya put the AI economy's input labour into a policy document for the first time. The draft Artificial Intelligence and Other Emerging Technologies Policy, open for comment until 4 August, proposes a fair-pay reference framework benchmarking data-annotation, content-moderation and AI quality-evaluation wages against international rates, employer disclosure of pay structures, duty-of-care obligations on prolonged exposure to disturbing material, and occupational guidelines on written contracts, psychosocial support and grievance procedures that the government commits to publish later. The wage gap it addresses is $1.46 to $3.74 an hour in Kenya against $21 to $27 in the United States. Its parent policy proposes an AI Safety Institute, a public AI registry and a National African Languages and Voice Programme. Ghana, meanwhile, announced a 10-year AI implementation plan and a $250m compute centre at its Accra summit on 29 and 30 July.
BUILDING
Signal
Karnataka is the one place where money actually moved. On 24 July the Internet and Mobile Association of India and its members Eternal, Zepto, Swiggy, Urban Company and Meesho deposited the welfare fee the High Court ordered three weeks earlier, in an amount nobody disclosed. The levy is 1% of the worker payout, capped at 50 paise per two-wheeler ride, 75 paise per three-wheeler and one rupee per four-wheeler, with 1% on food and grocery. Justice M Nagaprasanna refused to stay the Act, holding the amount is not demanded as charity but in terms of promulgation of a legislation, and warned that if the money is not going to gig workers he would stay the Act immediately. The state's own estimate of the workforce covered is about four lakh. Separately, Cognizant closed the June quarter down about 900 to 356,700 and Tech Mahindra down 863 to 146,760.
EMERGING
Signal
The Gulf runs the third variant: a state not protecting labour from AI but enforcing labour law with it. MoHRE reported more than 190,000 Emiratis employed in the private sector across 32,000 companies, with 95% of mandated companies meeting their targets at the 30 June half-year checkpoint, and no softening language anywhere in the release. Fines of Dh10,000 a month per unfilled skilled Emirati role, Dh120,000 annualised, have been live since 1 July for firms with 50 or more staff. The inspection stack that polices it uses AI tools specifically to flag fake-Emiratisation schemes, and the ministry's own service layer ran nearly 16 million digital transactions in the first half through what it calls an AI-powered ecosystem. Ajman completed the UAE's first fully autonomous government transaction, a trade-licence renewal, on 23 July.
Short read · this week's signal across the nine sectors we cover
The AI supply chain's labour layer is the object being regulated: annotation, moderation and evaluation vendors serving frontier-model buyers now face a proposed pay benchmark and a duty of care in their largest African delivery market.
Food, grocery and quick-commerce delivery is where the levy actually bites, at 1% of payout, and where the platforms that pay it are also the ones litigating the Act that created it.
Outsourcing and BPO delivery sits between two Kenyan policies pointing opposite ways, one growing the sector and one raising its cost, eight days apart from the same government.
Welfare-fund administration is a financial-services problem wearing a labour-policy hat: a levy collected into a court registry needs custody, reconciliation and a disbursal rail before it is a benefit.
Service-sector employers in the Gulf sit directly under the quota and the fine, with AI-assisted inspection now flagging circumvention rather than merely counting compliance.
Occupational-health provision is being written into AI supply-chain contracts for the first time, with psychosocial support proposed as a duty rather than a benefit.
Least touched by this fortnight's instruments, but the same logic that priced annotation labour will reach any workforce whose output trains a model.
Compute infrastructure is the physical form these policies assume: Ghana's $250m AI compute centre is a bet that hosting the capacity buys a say over the work.
Furthest from this issue's instruments; the Gulf's automation of its own service layer is the closest thing to a signal, and it is about the state, not the sector.
For Editor reAImagine · curated to this issue's signal · 90-day horizon
Why now
Most reporting called a draft policy a rulebook, and the person who reads the document itself becomes the AI data-work compliance lead.
Do this
Why now
Kenya's draft would require employers to benchmark annotation pay against international rates and disclose their structures, and whoever builds that file first becomes the Moderator welfare officer's natural counterpart.
Do this
Why now
Karnataka's fee reached a court registry and stopped, and the person who designs the rail from collection to claim becomes the Gig welfare-fund administrator.
Do this
Why now
IAMAI's repugnancy plea under Article 254 could void the levy or duplicate it, and the analyst who can model both outcomes becomes the Platform-levy reconciliation analyst.
Do this
Why now
Nairobi and Bengaluru regulate the AI economy's labour while the Gulf automates labour-law enforcement, and the person who keeps the two apart becomes the Labour-inspection data lead.
Do this
Nothing scores this week, and no open entry falls due before the next issue. One watch is worth printing without scoring it: LEDGER-004-03, which bets MoHRE softens Emiratisation enforcement before 31 December, is trending clearly against us. MoHRE's 20 July half-year release reports more than 190,000 Emiratis in the private sector and 95% of mandated companies meeting their targets, with no softening, delay or waiver language anywhere in it, and six independent enforcement pieces checked this week carry none either. We also correct the record from 018: Wipro's Q1 FY27 headcount of 243,044 belongs to the company's datasheet, not its press release, and the quarter-on-quarter move is plus 888 from 242,156, with the widely reported plus 9,812 being a year-on-year change mislabelled as a quarterly one. The organic-versus-acquired bridge cannot be closed from any Wipro document, and we are recording it as permanently ambiguous rather than chasing it a fourth issue. Ledger 005 opens after the carried entries, native to this issue: it bets that Kenya's pay provision survives to enactment, and that Karnataka's Act survives its constitutional challenge.
Anthropic's ID-verification policy takes effect and, whatever its stated intent, functions in practice as a citizenship-sorted access path: US consumers regain restricted-tier access first, with no announced parity path for Indian or GCC passport holders. Anthropic says the change is an unrelated appeals update; we forecast the observable outcome and will score it.
Scored 9 July 2026. Fable 5 came back for every consumer on earth on the same day, 1 July, because the US Commerce Department lifted the export controls on 30 June. The restoration ran through diplomacy, not identity checks, and it landed a week before the ID policy took effect on 8 July. The policy itself verifies identity and age for flagged consumer accounts, carries no nationality component at all, and exempts Team, Enterprise and API customers. The disconfirming evidence we carried inside the entry, Anthropic's statement that this was an unrelated appeals update, held up better than our forecast did. To score this a hit we needed restricted access re-sorted by passport through the verification flow. It was not.
At least one further US frontier-model release goes through government pre-release review rather than open launch, extending the pattern already visible in June.
Scored 16 July 2026, a day early, because the pattern resolved ahead of the date. OpenAI previewed GPT-5.6 with the US government for about a month, released it on 26 June as a limited preview to around 20 government-approved organisations, and only opened it to the public on 9 July after a federal evaluation window under Executive Order 14409's voluntary pre-release framework. That is government pre-release review rather than open launch, exactly as forecast. The honest complication belongs on the record: the White House publicly denied giving any green light, approval or clearance, and EO 14409 explicitly bars mandatory licensing or preclearance. The claim required review, not approval; review demonstrably happened, so the hit stands on the wording as published.
At least one of TCS, Infosys, Wipro or HCLTech publicly announces a formal multi-model or sovereign-fallback architecture policy as strategy, not as a procurement footnote.
At least three of India's top four IT firms disclose a named AI-revenue metric, in whatever form each chooses, in their Q2 FY27 results.
India's top four IT services firms, TCS, Infosys, Wipro and HCLTech, in aggregate add net headcount over FY27's first half, April to September 2026, while each scales AI-attributed revenue, confirming the reroute: the work returns offshore even as the Western rhetoric softens.
The first senior role explicitly titled for AI sovereignty or model continuity, distinct from CISO or Chief AI Officer, is publicly posted by a GCC entity or Gulf sovereign-linked employer.
At least one multinational publicly names the Philippines, Romania or Poland, India's closest challengers on this index, as the lead location for a new AI-delivery or engineering hub, chosen over India, in a 2026 announcement.
On the next annual refresh of this index, India retains first place on the outsourcing-led composite while staying outside the top three on the capability-weighted view, confirming that its lead rests on delivery scale rather than AI preparedness.
MoHRE publicly adjusts, delays or waives an element of Emiratisation enforcement, citing market conditions, before 31 December 2026.
trending against. MoHRE's 20 July half-year release carries no softening language, and six enforcement pieces checked this week carry none either.
At least one company that attributed 2026 layoffs to AI is publicly reported to have rebuilt the same function in India, the Gulf or Africa, directly or through a capability centre or outsourcing partner, within twelve months of the cut.
A Gulf sovereign-linked or government entity publicly launches an initiative to position the UAE or Saudi Arabia as an AI-work delivery hub, not only a buyer or funder of AI, consistent with the capability-strong, labour-light profile the index assigns the Gulf.
The Karnataka Platform Based Gig Workers Act survives its constitutional challenge, meaning validity upheld, or the petitions dismissed or withdrawn, by 31 March 2027.
Basis: Justice Nagaprasanna's refusal to stay the levy and his statutory-mandate reasoning, plus platforms already depositing, which weakens their own irreparable-harm argument; against it, the bench's own concession that on a peripheral examination the Act does look like it runs repugnant under Article 254.
Kenya enacts its AI policy, or an AI Bill, with the data-worker pay provision substantively intact, meaning pay for annotation, moderation or evaluation work calibrated against international rates for equivalent work, by 30 June 2027.
Basis: the provision has ministry authorship rather than only advocacy backing, and consultation closes 4 August 2026; against it, Kenyan legislative timelines are slow and the same ministry launched a BPO growth policy eight days earlier.
Accenture's new bookings decline year on year again, confirming the June repricing as structural rather than sentiment.
Challenger's AI-attributed US job-cut count for the second half of 2026 exceeds the first half's 101,743, despite the softened executive rhetoric. The narrative and the number diverge further, not less.
Data & AI remains a top-two growth sector in every remaining 2026 quarterly Cooper Fitch Gulf Employment Index, even if total GCC hiring stays flat or negative.
A monthly board-ready brief tying AI to the workforce and location decisions in front of you. Built by reAImagine.work and InGovern Research Services. Bengaluru and Dubai.
What you receive
A board-ready brief tied to your sector and the workforce decisions in front of you this quarter, this fiscal year, and twelve-plus months out.
Who it's built with
Editorial by reAImagine.work, founded by Debu Mishra. Board governance practice from InGovern Research Services, founded by Shriram Subramanian. Bengaluru and Dubai.
How it lands
Approve. Appoint. Commission. Separate. Deploy. Decide. Establish. Stakes in INR or USD ranges. No theatre.