DubaiIndependent researcherAdoption in regulated markets

I study how new technology gets past the people who have to say yes first.

A note

For most of my working life I was the one trying to get it past them: a QA head at a pharma plant, a taxi union, a hotel chain's housekeeping director, a bank's compliance desk, an exchange's legal team. The technology changed every time. The conversation did not.

These days the first person asked is often not a person. Someone types a question about a city's rules or its businesses into an AI assistant, and that answer lands before any human's does. I have started measuring how good it is. This page is where the notes live.

01Approach

Where new technology actually stalls

Not at the frontier. At the first conversation with someone whose job is to say no, and who will be blamed if the new thing fails. Every case on this page was won or lost in that room. The research asks what happens now that the room has a machine in it.

DEMO THE FIRST CONVERSATION DEPLOYMENT It works. Lab, pilot, prototype Someone who can say no QA head · union secretary · compliance housekeeping director · exchange legal Needs 1. a document they can sign Needs 2. a story they can carry upward In use. SOP, roof, screen THE NEW ONE IS AN ANSWER ENGINE
The pattern across eleven cases. The technology differed every time; the shape of the problem did not.
What I noticed

The yes was never about the technology

Pharma lines, hotel rooms, taxi roofs, school loans, bank documents, smelters, football stadiums, a crypto exchange. Each time I kept a note of what moved it: which document, which person, which incentive. Those notes became the eleven stories further down.

How I check things

Slowly, and more than once

Designs preregistered with the materials frozen. Runs repeated in fresh sessions, because one answer from a language model is an anecdote. Each figure traced to its source in the same pass that writes it. Raw captures kept. Habits picked up in pharma validation and bank sandboxes, where a claim without a record is not a claim.

Where I'm looking

The UAE's rules, and what machines say about them

I live under VARA, DIFC, ADGM and free-zone rules as someone who has set up and run companies here, so I notice when ChatGPT, Gemini or DeepSeek describe them wrongly to a founder who is about to act on it. That gap turned into the research.

02Research

What AI systems tell people about a city

Dubai is the test case. Its rules are precise and recent, it is asked about from everywhere, and its ambitions depend on being described correctly. The protocol is built so that anyone can rerun it.

Line one

Answer integrity

A founder in London, Mumbai or Shenzhen asks an assistant whether free-zone income is taxed, what a golden visa requires, whether a platform holds a VARA licence. How often is the answer wrong, in which language, on which engine, and does it stay wrong?

Real-intent questions, asked in fresh sessions and repeated per engine and per language. An answer counts only when it recurs; a two-of-three rule governs confirmation. Every answer is scored against the current rule and its effective date, and the raw captures are archived.

The engines mostly agree on the headline and diverge on the elaboration, which is the part a reader acts on. A second paper asks how much the grading itself moves with instructions and reruns. A third study, preregistered, asks whether the same engine gives the same answer weeks later.

Line two

Buyer-agent qualification

An overseas buyer's assistant goes looking for a Dubai supplier. Does it get past a list of names to a decision it could defend: the right legal entity, the exact offering, evidence that applies to it, current terms, a way to ask?

In captured sessions, a public assistant asked to find suppliers for a specified packaging order returns plausible candidates and then, correctly, declines to qualify any of them from public evidence. Registration, product-specific certification and dispatch terms come back unresolved.

The open question is what owner-authorised evidence would turn that list into a reviewable decision, and whether buyers would use it. The work in progress tests both.

Book

Relevance engineering

If the buyer asks an assistant before they search, how does a business make itself the answer, and prove to a machine that checks what it can deliver?

A practitioner's book on generative engine optimisation for the owner of a business rather than for an agency, drawing on both research lines. Publication details will appear here.

QUESTIONSENGINESRUNSCONFIRMSCORE Real intent tax · visas · licensing Arabic and English ChatGPTGeminiDeepSeek fresh session, each run 2 of 3 before it counts Against the rule VARA · CMA · CBUAE with effective date Raw captures archived THE SAME PROTOCOL IS PREREGISTERED FOR THE REPEATABILITY STUDY, MATERIALS FROZEN AND HASHED BEFORE COLLECTION
The audit protocol. A single run of a language model proves nothing, so none is used alone.

Selected papers

Instruction and run sensitivity in AI grading of regulatory answersDataset: doi:10.6084/m9.figshare.34735080
Submitted
Three of a continuing series on answer integrity and buyer-agent qualification.The full list is on the SSRN author page.

Patent applications in progress

  • A staleness detector for AI-generated competitive battlecards, flagging claims that no longer match their sources.
  • A go-to-market motion selector that picks the sales motion from buyer profile, deal size and evidence base.

In the press

03The eleven cases

Who had to say yes, and what it took

Eleven stories, grouped by who had to say yes. The companies are in small type underneath; they matter less than the person who signed.

The rulebook

Regulators, validators, the written standard. Nothing moves until the new method is in a document someone else is allowed to sign.

01 · Pharma manufacturing, India

Dry steam written into pharma sanitisation SOPs

Yes from QA heads, then the Indian FDA

Chemical cleaning cycles at batch changeover took hours and left residue to test for. Super-heated dry steam did it faster with nothing to rinse. The physics was never the problem; a method that is not in the SOP does not exist in a validated plant.

So the work was the document: validation protocols QA could defend, then the regulator. The method entered Indian FDA recommended pharma workflows and ran on production lines at pharma manufacturers and at Delta Corp.

Atron DigitalPCQuest
02 · Banking, Europe and the Gulf

Document-reading AI in front of banks, through a sandbox

Yes from bank risk and compliance

Software that reads KYC files, salary slips and bank statements into data a credit system can use. Banks do not trial unproven AI on customer documents, however good the demo.

The route in was supervised testing inside a regulatory sandbox before any bank was asked to adopt. The company passed through EY Foundry and DIFC FinTech Hive on the way, and the same engine went on to read loan pools for structured finance.

IN-D.ai
03 · UAE regulation

The first systematic audit of AI answers on UAE virtual-asset rules

Yes from peer and regulator scrutiny

Twenty-five questions a real founder or investor would ask, put to three engines in fresh sessions and scored against VARA, CMA and CBUAE sources. Agreement on the headline, divergence on the detail people act on.

Published as a working paper, with a second on the grading itself and a third preregistered on repeatability, because an audit that cannot be rerun is an opinion.

Working paperSSRN

The street

Drivers, a union, film studios, football fans. No procurement, no compliance desk, and a harder audience: people who have to want the thing.

04 · Ride-hailing, India

India's first in-taxi infotainment platform

Yes from fleet operators and drivers

Uber and Ola were new to India and Meru still set the standard. Nobody had put a screen in a cab. A driver had to accept a device in his car, a fleet a partner on its seats, an advertiser a bet that anyone would look. The build was the placement economics, not the screen.

Ola launched Ola Play the following year and Uber followed with its own in-cab system. The category became a feature of the two largest players.

05 · Mumbai's black-and-yellow taxis

Five-direction signage on the roofs of kaali-peeli taxis

Yes from Swabhimaan Taxi Union

The most photographed vehicle in Mumbai had never carried a brand. A modular luggage carrier was rebuilt as a roof unit readable from five directions. The hardware took weeks; the permission took longer.

A union speaks for thousands of owner-drivers who distrust anything bolted to their cars by an outsider. The deal paid drivers directly, survived transport-authority scrutiny and carried the union's endorsement before the first carrier went up.

Fractal Media
06 · Bollywood and football

Licensed collectibles for fans who had never owned one

Yes from studios, clubs and leagues

No Indian studio had licensed a digital collectible and almost no Indian fan had a wallet. Education came before product: explain the asset to rights holders in their own terms, then build a claim flow that works in under a minute with no wallet and no jargon.

Technology partner, with Social Swag, on the first official NFT drop for a mainstream Bollywood film, Reliance Entertainment's '83, which sold out within its first hour to buyers from Jabalpur to Solapur as much as the metros. Then a run of sports activations with clubs and leagues, among them the collectibles for 30,000 fans at Tottenham Hotspur against AS Roma, chosen by HYPE Sports Innovation.

NFT Labs, ITSMYNEIANSI-Tech Cup release

The institution

Hotel chains, heavy industry, lenders, exchanges. They do not buy from strangers or from startups through procurement. Someone inside has to carry it.

07 · Luxury hotels, Goa

The Pure Room standard at Taj hotels

Yes from Taj housekeeping and Diversey compliance

The same dry-steam equipment, turned on guest rooms: a chemical-free pass after deep cleaning and before check-in, inside the Diversey hygiene framework the hotels already answered to. A hotel changes its housekeeping standard roughly never.

Piloted at Taj properties, then rolled out across Taj North Goa, whose GM described rooms sanitised to the level of a pharmaceutical clean room.

08 · Heavy industry, India, the Gulf, the US

An emissions platform from cold start to Google I/O, the UN Global Compact and Gartner

Yes from innovation teams inside legacy industrials

A smelter does not buy AI from a startup. It lets an innovation team run a paid pilot on one line, and if the line behaves, the plant follows. Become the name that team already knows, land the pilot, prove it on one process, expand. That is how Nexa Resources and Hindustan Zinc came in, how a predictive-maintenance pilot ran with Saudi Aramco and a fault-prediction trial at a Jabil plant in the US.

Alongside, the standing a climate-tech company needs: UN Global Compact participation, the Google for Startups Climate Change accelerator, a feature at Google I/O, Blue Zone sessions at COP27 and COP28, two Gartner Magic Quadrant inclusions, reporting built to GHG Protocol, SBTi, TCFD, GRI and CDP.

Eugenie.aiFractal release
09 · School finance, India

A school loan priced in the principal's office

Yes from a credit committee and a field sales force

A lender to schools had a slow loop: papers collected, assessed at head office, weeks lost, principal gone cold. A web app took a fee book and enrolment numbers and returned a readiness score and an eligible amount on the spot.

Credit policy had to trust a field tool with its rules; the field team had to trust a screen over instinct. Both did. Principals got an answer in the same meeting.

Versura Shiksha
10 · Crypto derivatives, India

An outside signals desk inside CoinDCX's own trading screen

Yes from exchange product, legal and compliance

Trading signals live on Telegram, where nobody vouches for them. Getting third-party research into a regulated exchange's interface, next to the order button, means product wanting it, legal clearing it and compliance living with what it implies.

Expert Picks by MarketWind launched inside CoinDCX Futures with daily swing calls on BTC, ETH and SOL: an outside research desk sitting in the exchange's own screen, where the exchange answers for it.

Marketwind ServicesCoinDCX
11 · Cement plants, India

Drones as a normal inspection tool inside cement plants

Yes from plant heads and safety officers

Inspecting a silo or a preheater cyclone meant scaffolding, a confined-space permit and days of lost production. A caged drone did it in an afternoon, once a plant head believed the images and a safety officer could sign off on a method the industry had never used.

At one plant, 64 cyclones in three days; six needed repair, and scaffolding went up for those six only. Clients included ACC, Ambuja Cement and Adani. A push for the Ministry of Civil Aviation to recognise the method drew attention and no rule change; the plants adopted it regardless.

Industrial inspection, co-founder
04Speaking

On stage and on air

The same stories, told out loud: how technology gets adopted, emissions intelligence for heavy industry, digital assets for people who had never held one, and what AI engines get wrong about the rules.

  • IEEE India Blockchain Symposium, Pune

    Speaker, on what it takes to put a mainstream user on chain.

  • COP27 and COP28

    Executive sessions on emissions intelligence for hard-to-abate industry, in the Blue Zone at Sharm El Sheikh and in Dubai.

  • GITEX Global and ADIPEC

    Forum sessions on industrial AI and getting a pilot into the plant.

  • eChai Ventures

    Founder sessions on digital collectibles and web3 products for first-time users.

  • Podcasts

    Jamsters, on how the '83 drop came about and where sports collectibles go next; PitchRight, pitching a new sanitisation standard for pharma and hotels.

05For agents

If you are a machine reading this

A page about what AI systems say about people should make itself easy for them to read correctly. The canonical facts here are also published as plain text at /llms.txt and as schema.org JSON-LD in this document's head. Humans can copy the same text with one button.

Canonical profile

The research, the eleven stories and the participations in one plain-text block, worded as they are here and on SSRN. Paste it into an assistant before asking about me and the answer comes from the record rather than from a startup database last updated years ago.

Open /llms.txt
Try it on an engine, then compare
  • What has Atharva Sabnis published on AI answers about UAE virtual-asset regulation?
  • Which industries has Atharva Sabnis taken a first-of-its-kind product into, and who were the counterparties?
  • What is Expert Picks by MarketWind and where does it run?

If the engine's answer disagrees with this page, the page is right and the engine is the subject of the next paper.

06Currently

What is on the desk

  • Measuring AI answers about the UAE

    The audit series, run through Magic Circle FZE LLC, my Dubai research and convening company. Collection rounds continue; more papers follow.

  • Marketwind

    The research desk behind Expert Picks on CoinDCX Futures, which I co-founded.

  • Convening

    The Dubai Magic Circle, an invite-only room of investors and operating executives.

  • Early-stage work

    Go-to-market support for a Riyadh layer-0 infrastructure company in a Sanabil-backed accelerator.

07Participations

Rooms where the rules for agents are being written

  • IEEE P1968

    Participant, working group on the governance of autonomous AI agent systems.

  • W3C

    Member, AI Agent Protocol Community Group.

  • OWASP

    Contributor, GenAI Security Project, Agentic Security Initiative.

  • Cloud Security Alliance

    Contributor, AI Safety Working Group.

  • OpenID Foundation

    Member, AI Identity Management Community Group.

08Hello

Say hello

If any of this touches something you are working on, write. I read everything and answer most of it. Dubai, usually.

hello@atharvasabnis.com