August 28, 2026

Courts Cleared AI Shopping Agents. India's Payment Rails Have Not.

You tell your AI browser to find the cheapest 65-inch TV under Rs 60,000 and buy it. It finds one in about eleven seconds. Then it stops. Not because Amazon blocked it, and not because a judge did. It stops because nothing in India's payment plumbing knows how to let a piece of software hold your wallet.

Courts Cleared AI Shopping Agents. India's Payment Rails Have Not.

TL;DR: A US appeals court has ruled that when you send an AI agent shopping, you are the one visiting the store. That settles one fight and leaves the harder one untouched. In India, no payment rail yet exists that will let the agent pay.

Why It Matters

On 9 March 2026, a federal judge in the Northern District of California ordered Perplexity to keep its Comet browser off Amazon. On 4 August the Ninth Circuit vacated that order in Amazon v. Perplexity, No. 26-1444, and the reasoning is the part worth keeping. Where the AI company's servers never speak to the retailer directly, and every request routes through the shopper's own machine, the panel held that it is the shopper who "accessed" Amazon's computers under the Computer Fraud and Abuse Act. Not the software vendor.

Read the coverage and you would think agentic shopping just won outright. It didn't, quite. The panel left Amazon's contract and tort claims standing, terms-of-service breach included, and went out of its way to say it was not foreclosing liability for agents with greater autonomy, or for designs where the vendor's own servers hit the retailer. So the decision protects one narrow shape: a local agent driving your session, on your hardware, with your login. Build it any other way and you are back in front of a judge with worse facts. There is a second thing the ruling did not do. It did not stop Amazon from blocking agents by other means. Nothing in the opinion obliges a retailer to serve a request it can detect and refuse, and detection is a product problem rather than a legal one. A platform that loses on the statute can still rate-limit, fingerprint, challenge or quietly terminate the account, and none of that needs a judge's permission.

None of which is why your agent stalls at checkout in Bengaluru. The Reserve Bank of India's Digital Payments E-Mandate Framework, in force since 21 April 2026, requires your issuer to notify you at least 24 hours before any recurring debit lands. Sit with that for a second, with an agent in mind. The one rail built for automatic payments comes with a mandatory day of warning attached. It is the same instinct now shaping how platforms decide what software may act on your behalf, applied to money instead of apps, and it lands harder here. E-mandates were designed around a repeating charge whose amount and merchant you already know, on a date fixed well in advance. An agent's entire value is the opposite of that: an unpredictable amount, at a merchant you have never used, on the day it finally finds the thing. The rail and the use case were built for different worlds.

Four numbers frame the standoff: how long the ban actually held, what the retailer stands to lose, how fast agent buying is climbing, and how much of the open web is already machine traffic. Together they explain why Amazon spent five months fighting this, and why losing once has not ended it.

Injunction Held

148 days

Before the panel vacated it

Ad Revenue At Risk

$19.8 bn

Amazon, one quarter of 2026

Agent-Driven Orders

3x

Year on year, second quarter

Machine Traffic Share

1 in 30

Of all web visits, 2026

Take the order growth. Shopify's second-quarter 2026 earnings set it against a figure that matters more: 75% of AI-attributed purchases fell outside the company's top hundred product categories. Agents are not winning the things people already know how to buy. They are finding the awkward, badly-named, three-pages-deep thing you would have given up on by the second search.

"

A payment rail that must warn you a full day before it moves your money was built for gym subscriptions, not for an agent that just found your shoes.

Three Routes, One That Pays

So where does this leave an Indian shopper who actually wants the thing? Three routes exist right now, and the gap between them is not about how clever the model is. It is about who holds the authority to move money and who answers when the money moves wrongly. The third column deserves a note, because it is the one people picture when they hear the phrase. A delegated-payment agent is not an agent holding your password. It is an agent holding its own credential, one your bank recognises as separate from you, spending inside limits you fixed beforehand, leaving a trail that records which instruction came from a human and which from software. That is a great deal of new infrastructure, and not one piece of it is a model-quality problem.

Dimension Local Agent On Your Device Retailer's Own Assistant Delegated-Payment Agent
Legal Status Cleared on 2 statutes, CFAA and California's CDAFA Never at issue, you are the merchant's logged-in user Untested, no agent has settled on UPI rails
Contract Risk 2 claim families survive on remand: contract and tort None, the retailer wrote the terms it is enforcing Undefined until a protocol publishes its terms
Payment Rail Your saved card or UPI, you press the final button Retailer wallet or saved instrument, one tap Reported UPI extension, nothing in production
Auth Interrupt Second factor required above Rs 15,000 per transaction Same threshold, prompt raised on the retailer's screen No published exemption, assume the threshold applies
Purchase Latency Seconds, capped by how fast you type the OTP Seconds, instrument already on file 1 day floor, set by the pre-debit notice rule
Dispute Route Card or UPI chargeback, unchanged by the ruling Retailer grievance desk, then your issuer None defined, chargeback rules still to evolve
Setup Steps 3 steps: install browser, sign in, grant site access 1 step: open the retailer's existing app Not installable, no consumer-facing build exists
Live In India Yes, for search, comparison and cart building Yes, bounded to that one retailer's catalogue No, stakeholder consultation stage only
Best Suited For Comparison hunting where you still approve the buy Repeat orders inside one retailer you already trust Nobody yet, watch the consultation instead

Notice what the table does not contain. There is no column where an agent both chooses freely across the whole market and pays without you. That combination is the product everyone is describing, and in India it currently exists nowhere. The court fight was about the first half. The second half is a payments question, and payments questions in this country are settled by the regulator, not by the Ninth Circuit. It is tempting to read all this as a delay, as though the parts are on order and the launch is a scheduling matter. That reading is too generous. Granting payment authority to a non-human actor is a genuinely hard design question, and the countries working on it are not converging on a shared answer.

1 2 3 4 Agent identity Delegated mandate Payment authority Dispute route Reported, unspecified No published spec Held by the notice rule Still to be written

Four things have to exist before software can pay on your behalf in India: a way to identify the agent, a mandate you actually granted it, authority to move the money, and a route to complain when it goes wrong. Not one of the four is finished.

Friction Points

India does have an answer in progress. Business Standard reported in July 2026, citing industry sources, that the National Payments Corporation of India is building a Unified Agent Protocol to authenticate agents and set transaction limits without rebuilding UPI underneath. That is the right shape. It is also, as of today, a reported development rather than an announced product: no official NPCI statement, no timeline, no pilot, no published limits. The protocol is real, or at least the consultation is; the protocol itself is still slideware.

Which brings up the thing that irritates me about the current commentary. Specific per-transaction caps have been circulating as though they were policy. They are not. They come from one writer's proposal about what NPCI ought to do, and they have been repeated until they read like a specification. If you are planning around numbers nobody at the regulator has published, you are planning around fiction, and the correction will be expensive.

Then there is the hole nobody wants to own. An agent buys the wrong size, the wrong variant, the wrong seller. Under a card payment you dispute it. Under an agent-initiated debit, who is the counterparty: you, because a court just said the agent is your hands, or the vendor, whose model picked the listing? Reporting on the protocol work concedes that chargeback and dispute mechanisms will need to evolve. That concession is doing a lot of work. The same framework already obliges the issuer to send a post-transaction notification carrying its grievance redressal details, which tells you plainly how the regulator pictures recourse: a named human at both ends of every debit. An agent-initiated purchase breaks that assumption at the first step, and no amount of protocol design makes the question of who authorised the spend disappear.

  • Check where the agent actually runs. If it drives your browser on your machine, the Ninth Circuit's reasoning covers you. If it calls the retailer from a vendor's cloud, that protection was explicitly not extended.
  • Read the retailer's terms before you point an agent at it. Contract claims survived this ruling untouched, and account termination needs no court at all.
  • Keep the final confirmation yours. The moment you hand over the button, your chargeback story gets harder to tell.
  • Treat any agent asking for a standing payment mandate as premature. Nothing in India authorises it yet.

Three questions that decide your exposure

Where does it run? The machine the request leaves from is what determines whose legal problem an agent's shopping becomes.

Who presses confirm? Your finger on the last button keeps the dispute path you already understand. Delegating it swaps a known process for one that has not been designed.

What do the terms say? A retailer that bans automated access can close your account tomorrow, and no appellate reasoning about statutes will reopen it.

Use an agent to shop, not to pay. Let it hunt and fill the cart, then check the total yourself and press the button with your own thumb. That is not caution for its own sake, it is the only configuration where you keep both the legal cover the Ninth Circuit just described and the dispute rights India's payment rules already give you. Revisit it the day NPCI publishes an actual specification. Until then, the agent works for you right up to checkout, and that is genuinely useful on its own.

August 25, 2026

Android Sideloading Rules 2026: What Actually Changes For India Now

A friend sends you an APK on WhatsApp. It is a small utility, the kind that never made it to Play because the developer could not be bothered with a store listing. You tap it, Android asks its usual question, you say yes, and it installs. That flow is changing, and most of the coverage has handed Indian readers the wrong date for it.

Android Sideloading Rules 2026: What Actually Changes For India Now

TL;DR: Google's developer verification goes live on 30 September 2026 in Brazil, Indonesia, Singapore and Thailand. India is not on that list and will not be until 2027. What already reaches Indian phones is the advanced sideloading flow, with its one-time waiting period.

Why It Matters

Start with what the rule actually checks. Google is confirming who published an app, not what the app does. No code review, no content pass, no judgment on whether the thing is any good. It is a name-and-document check attached to a signing key, and the practical effect is that anonymous distribution stops being an option on certified devices. Whether you read that as overdue hygiene or as another gate on hardware you already paid for probably depends on how you felt when Apple kept self-service repair out of India using a similar safety argument.

India sits outside the first enforcement wave, and that is doing a lot of work in the coverage. Nothing on an Indian phone becomes uninstallable on 30 September. But look at which stores signed up: Google Play, Samsung's Galaxy Store, Xiaomi's GetApps, OPPO's App Market, vivo's V-Appstore, HONOR and Palm. That is nearly every app store that ships preloaded on a phone sold in this country. StatCounter put Android at 92.44% of India's mobile OS share in July 2026. When the rule does arrive here, it will not arrive at the margins.

The security case is real, and worth stating at full strength rather than waving at. Google's own Android Developers Blog post from 25 August 2025 said its analysis found over 50 times more malware from internet sideloaded sources than from apps available through Google Play. That is Google measuring its own platform, so read it with the appropriate squint, but a gap that size is not a rounding error. The counter-argument is not that the malware is imaginary. It is that identity checks catch the lazy operator and inconvenience the hobbyist, while the organised fraud rings buying stolen KYC documents in bulk carry on. Nobody has published a clean before-and-after on that yet, and until someone does, anyone claiming to know which way it nets out is guessing.

Cooling-off wait

24 hours

Before the first unverified install

Developer fee

$25

One-time full account registration

Hobbyist device cap

20 devices

Free account, no government ID

Play apps auto-registered

99%

Handled without developer action

The waiting period is the number that changes behaviour, and not because it is long. It cannot be completed in the moment. Someone on a phone call telling you to install a file right now, this minute, before your account is frozen, has just been handed a delay he cannot argue you past. That is the same pressure pattern running through the AI voice cloning scams working Indian phone numbers, and a forced pause is a surprisingly rude interruption to it.

"

A one-day pause is not really a defence against malware. It is a defence against urgency, which is what every coercion scam actually runs on.

What Changes And What Does Not

Strip out the speculation and the rule set is small enough to hold in your head. Here is the whole thing, as published on Google's developer verification pages and its Android Developer Console help documentation.

Category Detail What it means
Deadline 30 September 2026 Four countries only, India not included
First wave Brazil, Indonesia, Singapore, Thailand India's turn arrives in 2027
Stores Seven participating stores, Play included Indian phone brands are all inside
Sideloading Still allowed through the advanced flow Buried in developer options by design
Pause A one-time wait before the first install Built to break live coercion scams
Cost A paid full account, or a free limited one Hobbyists trade reach for privacy
Checked Developer identity, tied to a signing key Verification is not app content review
Escape hatch adb installs stay outside the rule Needs a computer and a cable

Read that table twice and the shape of the thing shows up. Google did not close the door. It moved the handle to a place ordinary users will never look, added a delay to the one path that leads there, and left a developer tool untouched for people who own a laptop. Every one of those choices is defensible on its own. Together they describe a device that is a little less yours than it was.

Aug 2025 · Mar 2026 · Aug 2026 · Sep 30, 2026 · 2027 · Rules announced · Console opens · Advanced flow · Four countries · Global rollout ·

Five stops on one line: the policy was announced in August 2025, verification opened to all developers in March 2026, the developer tools and the power-user install path shipped in August 2026, four countries hit enforcement on 30 September 2026, and everything else follows from 2027.

Friction Points

The advanced flow lives inside Developer options, which you reach by tapping the build number seven times. Google is not hiding it out of embarrassment. Burying it is the design. But it produces an odd result: the people most likely to need an app that never reached a store, a regional language keyboard, a niche utility, a local college's attendance app, are frequently the people least equipped to go digging through settings that warn them off at every screen.

Then there is the paperwork gap. An individual developer supplies a government ID. An organisation needs a D-U-N-S number from Dun and Bradstreet first, and that request can take up to 28 days to come back. For a two-person studio in Pune shipping a niche app, the fee is not the obstacle. The month of waiting and the registered-entity requirement are, and they arrive before a single line of the rollout touches India.

One more thing worth flagging, because it gets lost in the outrage cycle. The advanced path is not a one-time switch you flip and forget. Keep an eye on these before you decide the whole thing is overblown:

  • The unverified-install setting can be enabled temporarily for seven days or left on indefinitely, so a temporary grant will lapse without warning you again.
  • The adb route still installs unregistered apps, but it needs a computer, and plenty of Indian users have only ever treated a phone as their whole computer.
  • Google's help text covers new installs only, and says nothing about apps already sitting on your device, so do not assume either outcome.
  • Your phone brand's own preloaded store is on the participating list, so "I never use Play" is not the exemption people think it is.
  • The rollout reaches handsets in stages, so two identical phones in the same house can behave differently this month.

Certified devices only

The rule binds phones that ship with Google services, not every Android build.

Direct distribution survives

Google says developers keep the freedom to ship straight to users or via any store.

Two ways to register

Play Console, or a separate Android Developer Console for anyone who avoids Play.

So: nothing on your phone breaks next month. Use the time. If you rely on an app that lives outside a store, message the developer now and ask whether they have registered, because the answer in 2027 is going to be a lot less negotiable than the answer today. And if you are the sort of person who already keeps a second browser around rather than trusting one company's default, apply the same instinct here and get the file you actually care about onto your device while the door is still easy to open.

July 22, 2026

AI Voice Cloning Scams in India: Your 2026 Survival Guide

The call comes at 2:40 on a Tuesday afternoon. It is your son's voice — the exact pitch, the exact pause before he says "Amma," even that slight nasal edge he's had since childhood. He's been in an accident, he's at a private hospital, the desk won't admit him without an advance, and a "hospital administrator" comes on the line with a UPI handle. Ten minutes later the money is gone. Your son, it turns out, was in a lecture hall the whole time, phone on silent.

AI Voice Cloning Scams in India: Your 2026 Survival Guide

That is not a scene from a Netflix thriller. It is the standard fraud playbook of 2026, and it runs on two rails India built better than almost anyone: cheap AI tools and instant payments. The same UPI stack that lets you split a dinner bill in four seconds also lets a stranger with a cloned voice drain a retirement account before the chai gets cold.

TL;DR: Fraudsters now clone a familiar voice from a few seconds of public audio, fake an emergency, and push you to pay over UPI before you think. No app setting fully protects you. A family code word, a callback rule, and the 1930 helpline are your real defences.

Why It Matters

Start with how little the attacker needs. A wedding video on Instagram, a voice note forwarded to the wrong group, a podcast clip, a birthday reel — any short sample of someone speaking is enough raw material for today's open-source cloning models. The scammer doesn't hack your phone. They harvest your family's public audio, build the clone in minutes on a free tool, and then engineer panic: an accident, an arrest, a stuck visa, a kidnapping. Panic is the product. The voice is just packaging.

The scale has stopped being a niche cybercrime story. Industry tallies put global losses to AI-enabled fraud at roughly $442 billion for 2025, and India sits near the centre of that map because of how thoroughly daily life here runs on WhatsApp and UPI. This has climbed the food chain too — a cloned voice of Bharti Airtel chairman Sunil Bharti Mittal was used in an attempt to move funds through a senior executive. If a scammer will build a deepfake to fool a boardroom, the version aimed at your parents costs them almost nothing.

Here's my first unfashionable opinion: the "be aware, stay alert" poster campaigns banks love are mostly theatre. Awareness assumes the victim has time to think. These scams are engineered to remove thinking time — the fake son is crying, the fake policeman is shouting, the clock is always ticking. Defences that require calm judgment in the moment will keep failing. Defences that were agreed on before the call — a code word, a hard rule to hang up and dial back — actually hold, because they don't depend on you being clever while terrified.

The numbers below are why this deserves space on your family WhatsApp group tonight rather than someday. Each one describes a different edge of the same machine: how little input it needs, what it took last year, how many people reported it, and how fast it is growing.

Audio Needed
3 sec
enough to clone a voice
Reported Losses
₹22,495 cr
Indian cyber fraud, 2025
Complaints Filed
2.81 mn
cybercrime cases, 2025
Growth Rate
+24%
complaints, year on year

Sit with that first cell for a moment. It's less audio than the greeting most of us leave on a voicemail. It means the question "has my voice ever been recorded in public?" now has the same answer for everyone: yes. Planning your defence around keeping voices private is already a lost game — the defence has to work even when the clone is perfect.

The Anatomy of the Scam, Row by Row

Before the table, one piece of context. What police cyber cells describe from 2025-26 case files is not one scam but an assembly line, and every stage has been made easier by an off-the-shelf tool. Reading it as a system explains why no single tip — "check the number", "listen for robotic tones" — is enough on its own.

CategoryDetailWhat It Really Means
Voice sourceReels, voice notes, wedding videosYour family's audio is already public
Cloning toolsFree and open-source appsZero cost, zero skill required now
Face fakesOne photo, under a minuteVideo calls no longer prove identity
Delivery channelWhatsApp calls and fake UPI appsThe scam rides trusted, daily apps
Pressure scriptAccident, arrest, "digital arrest"Urgency is the actual weapon here
Payment railUPI, instant and irreversibleSpeed favours the scammer, not you
Recovery pathHelpline 1930, bank, cybercrime.gov.inReporting fast beats reporting perfectly

The row that deserves your attention is the payment rail. UPI's four-second settlement is a genuine engineering triumph — and in a fraud, every one of those seconds works against the victim. Once money moves through two or three mule accounts, recovery odds fall off a cliff. That is why the practical fight is won or lost in the minutes before you approve a payment, not the days after.

The four-stage pipeline above is the whole crime: only stage three ever touches the victim, which is why interrupting that single moment — with a code word or a callback — collapses everything before it.

Friction Points

Now the uncomfortable part: what still doesn't work. Caller ID apps flag unknown numbers, but these calls increasingly arrive on WhatsApp from freshly created accounts wearing a stolen profile photo. Banks send warnings, but a warning SMS competes with a screaming voice claiming to be your child. And the apps themselves keep getting faked — imitation UPI apps that look pixel-identical to the real thing have travelled across India through WhatsApp forwards. I wrote about how telecom-grade bot support fails customers in an ordinary billing dispute; imagine that same support maze when you're trying to reverse a fraudulent transfer with your hands shaking. In field studies of these cases, one number keeps surfacing: the window between the first ring and the first debit averages around 19 minutes. That's the entire battlefield.

There is also a genuine grey area nobody has resolved, in India or anywhere else: who eats the loss when a victim authorised the payment? Bank rules protect you reasonably well from unauthorised transactions. But a voice-clone victim taps "pay" themselves — tricked, yes, fully authorised, also yes. Regulators are still arguing about where liability should sit, and until that settles, assume the answer is: you do. Which is one more reason the AI tools now embedded in our daily software deserve scrutiny before trust — the same caution I argued for when comparing AI browsers against Chrome earlier this month.

Watch for these tells before any rupee moves:

  • Urgency plus secrecy. "Don't tell anyone, pay now" is the signature of every script. Real hospitals and real police have paperwork, not UPI handles.
  • A refusal to be called back. Hang up and dial the person's actual saved number. A genuine caller survives this test every single time; a clone never does.
  • Payment to an individual for an institutional need. Hospital deposits, court fines and customs fees do not land in a personal UPI ID.
  • Slightly-off audio behaviour. Clones handle interruptions badly. Ask an unexpected question — the pet's name, last Sunday's lunch — and listen for the stall.
Key Takeaways
  • Agree on a family code word tonight — it costs nothing and defeats a perfect clone.
  • Make "hang up, call back on the saved number" a house rule, not a suggestion.
  • Install UPI apps only from official stores; a forwarded APK is an automatic no.
  • If money moves, call 1930 and file at cybercrime.gov.in within the hour — speed decides recovery.

Do one thing before you close this tab: message your family group and set the code word. Not tomorrow. The scammers already have the voice samples — the only question left is whether your family has a script of its own when the phone rings.