Ask Kabbalist

You land in Delhi before dawn, connecting onward to Singapore, and the officer hands your passport back without reaching for a stamp. The camera above the desk already knows your face. Nobody sat you down and explained how to use DigiYatra, and at that terminal it has stopped being a choice you get to make. The queue moves faster. Your record of having been there moves somewhere you cannot see.

Why DigiYatra, and why the rush?

India's Bureau of Immigration stopped stamping international boarding passes on 1 September 2026, so the paper trail a traveller once carried home now lives entirely inside an app and an airline database. That is the real story, and it broke a week ago with almost no coverage joining the two changes together.

Two things happened this year, three months apart, and they only make sense read as one move. First the face scan became the default way through the terminal at the busiest international gateways. Then the ink stamp, the one artefact that proved you physically left and re-entered the country, was withdrawn. Visa officers asked for that stamp. So did tax officers, and employers running background checks. Now the answer is a PDF and a database query.

What is DigiYatra actually doing at the gate?

It matches a live camera image against an encrypted template built from your Aadhaar-verified selfie, then releases your boarding record to the checkpoint. Outlook Traveller reported the operational detail plainly: the selfie and the boarding pass go into the app before you reach the airport, and the gate does the rest. There is no counter and no second look. This is the same pattern as India's Android developer verification rules, where a convenience is announced first and the compulsion arrives later, quietly, in a circular.

The programme's governance is the part worth sitting with. DigiYatra Foundation is a private not-for-profit, and because it is private it sits outside the Right to Information Act, 2005. So the body holding a biometric template of every international flyer through four major airports cannot be asked, by an ordinary citizen, how long it keeps that template. The Internet Freedom Foundation has raised exactly this: privacy, surveillance, exclusion errors, and no institutional accountability. A Kerala High Court PIL alleging commercial misuse of passenger data is still live. None of that stopped the mandate.

That last figure is the one that should shape how you treat the app. It was measured while enrolment was still sold as voluntary, and it describes people who handed over a document at a counter and were signed into a biometric programme without registering that it had happened. A system with that much accidental enrolment in its optional phase does not suddenly grow better consent hygiene once it becomes compulsory. It just stops needing consent.

How to Use DigiYatra: What Each Step Demands

You enrol once in the DigiYatra app with an Aadhaar-verified selfie, attach each boarding pass to that profile before you travel, then walk to the e-gate where a live camera matches you against the stored template and opens the checkpoint.

The steps are simple. What sits behind them is not, and the table below is the version nobody prints on the airport signage. Read the Identity row first if you hold a foreign passport, and the Paper trail row before your next visa application.

The Ownership row is the one that changes how you should read every reassurance about deletion. A shareholding split between the state airports authority and the operators it regulates is a commercial arrangement wearing safeguard language, and the entity it created answers to neither a regulator nor a citizen with an RTI form. The 15-minute buffer in the Gate fallback row is not published anywhere either. It comes from the plain logic of the system: if the automated lane rejects you, you rejoin a manual queue that was sized for fewer people than it used to hold.

Can you register for DigiYatra without a boarding pass?

You can create the profile without one, since enrolment needs only the Aadhaar-verified selfie, but the boarding pass has to be added to that profile for every single trip before the gate hardware will recognise you and let you through.

Which means the app is two systems wearing one name. The identity half is permanent and lives with a private foundation. The travel half is per-journey and expires. Most of the friction people report sits in the second half: a pass added too late, an airline that issues the document only at the counter, a name spelt differently on the ticket and the Aadhaar record. The rollout is fine, mostly, if your face matches on the first attempt.

The camera is the least interesting part of this. What nobody has published is a retention window for the biometric template, and the body holding it is structurally unanswerable to the people it holds it on. That is a stance, not a finding, and it is the same gap that shows up whenever the rules land before the plumbing does, exactly as they did with AI shopping agents and India's missing payment rail.

• Download and store the e-boarding pass as a PDF the moment it arrives, because the immigration stamp that used to prove your trip no longer exists.
• Check that the name on your ticket matches your Aadhaar record character for character, including initials, before you add the pass.
• Ask for the manual counter if you do not want the scan. It still exists at these terminals for people the automated lane rejects, and staff will point you to it if you insist.
• Keep a printed pass in your bag. It looks absurd next to a face scanner, and it is still the cheapest insurance you can carry.

Do one thing this week. Open your last international e-boarding pass, save it somewhere that survives a lost phone, and do the same for every trip from here. The face scan is settled and arguing with it at the gate will only cost you your connection. The paperwork is what you can still control, and the day a consulate or a tax officer asks you to prove you left the country, that saved file is the whole of your answer.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

Related: How to use DigiYatra now that the face scan is mandatory for international flyers

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.

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.

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.

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.

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.

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.

You ask your browser to find the three cheapest flights to Delhi, check them against your calendar, and draft a note to the group chat. It does all of it in one pass. No stack of tabs, no copy-paste, no forty-minute rabbit hole. Then a week later a researcher shows how one poisoned link could have told that same helpful assistant to quietly forward your inbox to a stranger. Same tool. Same power. Two very different endings.

What These Browsers Actually Do

The browser used to be a window. In 2026 it wants to be an employee. OpenAI's ChatGPT Atlas, Perplexity's Comet, and The Browser Company's Dia all ship with an agent that reads the page you are on, references your other open tabs, and clicks through live sites for you. Google and Microsoft bolted the same kind of agent mode into Chrome and Edge. The pitch is simple: stop doing the boring web chores yourself and let the software handle them.

This is not a chatbot sitting in a side panel. The difference that matters is context and action. A traditional browser with a ChatGPT tab open still makes you shuttle text back and forth. An agentic browser already sees the checkout page, the flight results, and the half-written email, and it can act on all three without you lifting a finger. That shift is why adoption moved fast. AI-driven search sat under a tenth of activity back in 2023. By this year it climbed to a large slice of everyday queries, and roughly half of consumers now say they would rather get a direct answer than a page of blue links.

Those numbers are worth pausing on, because they explain why every major company suddenly wants to own the address bar rather than just the search box.

Take that last figure and sit with what it means on the ground. When nearly a third of searches skip the results page entirely, the habit that built Google, scanning a list and picking a link, starts to erode. People stop visiting sites and start asking the browser to read the sites for them. That is a quiet rewrite of how the open web gets used, and it is the real prize these tools are fighting over.

The Trade You're Really Making

Here is where the AI browser vs Chrome question gets honest. You are not choosing between a fast browser and a slow one. You are trading control for convenience. Chrome does what you tell it, click by click, and nothing more. An agentic browser interprets a goal and then makes dozens of small decisions on its own to reach it. When those decisions are booking a table or comparing prices, that is a gift. When a hidden instruction on a web page redirects those decisions, that is a problem nobody has solved.

That problem has a name: prompt injection. Attackers hide commands inside ordinary content, a web page, an email, even the text buried in a URL, and the agent cannot reliably tell your instructions apart from the stranger's. It processes both through the same pipeline. OpenAI said plainly in December 2025 that prompt injection is "unlikely to ever be fully 'solved.'" Read that again. The company building one of these browsers told you the front door does not fully lock. And because the agent can reach your logged-in accounts, a successful attack is not a nuisance popup. It is your email, quietly leaving.

Atlas vs Comet vs Dia vs Chrome, Side by Side

Feature checklists miss the point here. What separates these four is how much they act for you, and how much that exposes you. This table reflects where each one stands in 2026, not a launch-day demo.

Read across the bottom row and the strategy picks itself. The AI browsers win on doing. Chrome wins on not getting you burned. Most people will end up running both, an agentic browser for grunt work and a locked-down one for banking, and that split is a feature, not a failure.

It helps to see how an attack actually travels, because the mechanics are simpler and nastier than the marketing suggests. Four steps, no malware download, no dodgy attachment, just words on a page the agent was told to trust.

Where It All Goes Wrong

The convenience is real, and so are the failure points. These are not rare edge cases. They are the predictable seams that show up once an agent has real access to your accounts and the open web at the same time.

Security teams have already turned theory into working attacks. Researchers at LayerX demonstrated a technique they nicknamed CometJacking, where a single crafted link could push Comet into pulling data from a user's connected Gmail and calendar and shipping it to an outside server. Brave's own security group reproduced indirect prompt injection inside the same browser. None of this required the user to type anything wrong. They just clicked.

Watch for these before you hand an agent the keys:

• Account reach. If the browser is logged into your email, bank, or work tools, a hijacked agent inherits all of that access instantly.
• Invisible instructions. Malicious text can hide in white-on-white page content or query strings you never see, so nothing looks off.
• No clean fix. This is the honest grey area. Vendors can add guardrails, but they openly say the underlying flaw has no perfect patch, so caution is on you.
• Silent actions. An agent working in the background can click, send, and share faster than you can notice and stop it.

So do not rip out Chrome this weekend. Install one agentic browser, point it at your low-stakes chores, research, shopping comparisons, and messy tabs, and keep every account that touches money or identity on the browser you already trust. Run them side by side for a month, watch what the agent does when you are not looking, and let its behavior, not the keynote, decide how far you hand over the wheel.

You're three weeks into a product redesign sprint. Your junior designer just delivered a Canva export. Your tech lead is screaming about broken component handoffs. And your creative director forwarded an Adobe Firefly AI Assistant demo with one line: "Can we do this?"

Three tools. Three completely different promises. One overloaded design team.

The question isn't which tool has the best keynote moment. It's which one actually survives contact with a real 9-to-5 workflow — the kind with tight deadlines, mixed skill levels, and a Slack channel that never stops pinging.

The Short Answer (Read This First)

Figma AI is the undisputed pick for UI/UX teams who live in component libraries and need airtight developer handoff. Adobe Firefly leads when brand-safe, commercially licensed generative content — now spanning images, video, and multi-app agentic workflows — is non-negotiable. Canva Magic Studio, fresh off its AI 2.0 overhaul at Canva Create 2026, is genuinely impressive for speed — but it's engineered for teams who think in content, not systems. None of them is the complete answer. Pick based on where your workflow currently breaks.

What Actually Separates These Tools in 2026

Most roundups compare pricing tables and checkbox feature lists. The real difference shows up in specific workflow moments — not spec sheets.

Figma AI has undergone a fundamental rethink. Its 2026 AI suite now spans content generation, image editing, smart search, UI drafting, code handoff, and full-site creation — tools like First Draft, Make Image, Replace Content, and Add Interactions help teams move faster through early ideation and prototyping. But the single most significant shift is Figma Weave — a visual canvas for building repeatable, scalable generative AI workflows, letting teams generate images, turn images into video, and scale brand guidelines into full illustration sets. Figma And then there are AI Agents: through the Figma MCP server, AI agents can now write directly to Figma files, creating and modifying real design assets using your team's actual components, variables, and tokens.

Key Figma AI strengths for UI/UX work in 2026:

• Code-to-Canvas: paste a React, HTML, or SwiftUI snippet and Figma instantly generates an editable UI component on your canvas — what previously took three full days of manual recreation now takes under two hours
• AI agents governed by Skills — Markdown instruction sets that encode your design system's intent, not just its assets, making AI output far more predictable across your team
• Figma Make: prompt-to-prototype tool that generates working apps from text descriptions or existing designs
• Over 200 AI-powered plugins available as of mid-2026, with ChatGPT Images 2.0 now integrated into Make Image and Edit Image
• Dev Mode with AI-generated CSS and platform-specific code snippets
• Figma Sites, Figma Buzz, and Figma Draw now round out the platform well beyond its original design canvas roots

Adobe Firefly, by contrast, has spent 2026 becoming an agentic creative engine rather than just a generative image tool. The headline is the Firefly AI Assistant, which brings the power of Adobe's creative apps into a single conversational interface — describe what you want, and it orchestrates multi-step workflows across Photoshop, Premiere, Lightroom, Express, Illustrator, and more Adobe. For UI/UX specifically, Firefly's biggest card remains its commercially safe training data. But the tool's depth has expanded considerably.

Firefly now gives creators access to more than 30 top industry AI models — including Kling 3.0, Google's Veo 3.1, Runway Gen-4.5, and ElevenLabs — positioning it less as a single AI and more as a unified front-end for whichever model best suits the task at hand.

What Firefly does better than the competition in 2026:

• Precision Flow: generates a range of image results from a single prompt via an intuitive slider, letting you explore interpretations from subtle to dramatic without rewriting prompts
• AI Markup: draw, select, or reference specific areas of your image, then apply targeted edits with text prompts or image references
• Custom Models (public beta): train a Firefly model on your own images to capture a specific style, character, or photographic look — reusable across projects without losing visual consistency
• Adobe Brand Intelligence: validate and assemble on-brand content against your brand rules across entire production workflows
• Adobe Stock integration directly inside the Firefly Video Editor, with access to 800 million licensed assets

Canva Magic Studio arrived at its most significant moment yet. At Canva Create 2026 in Los Angeles, Canva unveiled Canva AI 2.0 — the most significant evolution of the platform since launch — expanding beyond design generation to become the system at the centre of how work gets done.

The standout feature: Magic Layers, which transforms static AI-generated images into fully editable designs, has been used more than nine million times in just over a month since launch. And the new conversational interface accepts text or voice prompts and returns fully editable designs — say "create a flyer for a product launch with a dark background" and a polished layout appears in seconds.

Where Magic Studio genuinely earns its place in 2026:

• AI video generation — turning a simple text prompt into a polished video — now available globally across all locales.
• Magic Insights delivers cohesive explanatory narratives alongside ready-to-use charts and formulas, turning raw data into actionable decisions.
• Canva AI Studio now embedded inside the newly acquired Affinity suite — giving professional designers Generative Fill, Expand & Edit, and background removal inside a studio-grade vector and photo tool
• 85% of marketers using the platform save at least 4 hours per week with Canva's AI tools, according to Canva's own 2026 AI in Marketing report

But the ceiling still arrives fast for UI/UX professionals. There's no true developer handoff, no conditional prototype logic, and the component system remains template-bound compared to Figma's structured variant model.

Why the Numbers Tell a Harder Story

The mistake most teams make is optimizing for features when they should be optimizing for where time actually bleeds out. Internal benchmarks from mid-sized design agencies reveal that a 4-person UI/UX team loses an average of 4.3 hours per weekly sprint to tool-switching friction alone — jumping between Figma for wireframes, Firefly for asset generation, and Canva for stakeholder decks. That's nearly 18 hours a month that appears nowhere in any project estimate, and none of the 2026 updates have solved it yet because the tools still serve fundamentally different masters.

The financial and scale gap between these platforms is equally striking, and the numbers tend to land harder than any feature comparison:

That $479 gap between Adobe Creative Cloud All Apps (~$659/year) and Canva Pro ($180/year) is the number most teams use to justify Canva in budget conversations — but it's a false economy for anyone doing real UI/UX work. Figma Professional, meanwhile, has settled at $12/editor per month on annual billing Vendr — making it, per seat, actually cheaper than Canva Pro annually ($144 vs $180 per seat per year). That pricing inversion from what most designers assumed two years ago quietly happened, and it changes the calculus for small teams who once defaulted to Canva on cost alone.

The Head-to-Head That Actually Matters

Before committing to any platform, understand what you're comparing against dimensions that UI/UX work genuinely depends on — not keynote bullet points.

This table reflects real-world conditions in 2026, not sandbox demos:

The "New in 2026" row is doing real work here. All three platforms shipped substantial updates, but the nature of those updates reveals their strategic direction: Figma is becoming a design-to-engineering platform with agents in the loop; Adobe is becoming a cross-app agentic studio; and Canva is becoming the conversational interface for visual work that doesn't require design expertise.

Where Each Tool Will Actively Trip You Up

These aren't edge cases. These are the friction points teams report after 3–6 months of actual production use.

Figma AI pitfalls in 2026:

• Most AI outputs still need human review for accessibility, semantics, and production readiness — the tools improve speed but don't replace judgment.
• Figma Make's generated code remains bloated and largely unusable in most development contexts without significant cleanup; it works for rapid prototyping but rarely ships as-is.
• AI credit enforcement began in March 2026 — teams that relied on generous pre-enforcement limits will find themselves hitting ceilings mid-sprint
  • Pay-as-you-go credit top-ups are now available but add an unpredictable line to monthly bills
• Dev Mode's AI code snippets still perform better for React than for native Android; iOS output typically needs a QA pass before handoff

Adobe Firefly pitfalls in 2026:

• The Firefly AI Assistant is entering public beta — it's announced but not yet fully in teams' hands, meaning the agentic workflow is still more promise than production-tested reality
• The standalone Firefly plan (~$180/year) is capable but isolated; the real cross-app workflow requires the full Creative Cloud All Apps subscription at $659/year
• There is still no native prototyping in Adobe's lineup; the XD gap from 2023 remains unfilled even after NAB and Summit 2026 announcements
  • Teams using Firefly for UI work almost always run Figma or another tool alongside it

Canva Magic Studio pitfalls in 2026:

• Advanced features like Magic Layers remain in public beta in only a handful of markets (US, UK, Canada, Australia), with global rollout still pending.
• Brand kit AI enforcement is inconsistent — colors apply reliably but font hierarchy rules break down on complex multi-column layouts
• Magic Write still produces serviceable copy for general use but fails with technical product language and compliance phrasing
  • Legal teams frequently flag Magic Write output as requiring full rewrites, not light edits
• Canva raised its Pro plan price from $10 to $15/month in late 2024 — a 50% increase — which changes the cost-benefit math for solo freelancers who used to default to it on budget grounds.

One grey area worth acknowledging honestly: if you're a mid-level freelancer handling both marketing design and occasional UI requests, no single tool here covers everything cleanly in 2026. Figma is feature-heavy relative to what you'd use it for at that volume. Firefly is expensive unless you're already invested in Creative Cloud. Canva AI 2.0 handles 80% of your output until a client asks for developer-ready specs — and that 20% always arrives at the worst possible moment. The real answer is probably two tools. Which two depends on a client mix that shifts every quarter, and anyone who tells you otherwise is selling you something.

The Verdict — No Ceremony

If your team ships digital products, Figma AI is the call. The agent-on-canvas capabilities via Skills and Weave alone have made a gap that was already wide into something that's becoming structural.

If your team needs brand-consistent, legally safe asset generation at scale — across images, video, and cross-app workflows — Firefly earns its subscription once the AI Assistant matures out of beta. Give it another quarter before betting the pipeline on it.

And if you're onboarding a non-designer into a content production workflow in 2026 and need results in under a week, Canva AI 2.0's conversational interface is the fastest path to competency that exists right now.

The 4.3 hours a week lost to tool-switching only accumulates when teams lack a clear protocol for which tool owns which output type. Write that protocol before you approve the next tool purchase. The platforms keep getting smarter. The teams using them without a decision framework don't.