Blog

If you have come across a modern browser, you must have noticed that it has changed. It has not remained passive just to open tabs or search bars. It has become active and intelligent. Browsers such as Comet (Perplexity), Atlas (ChatGPT) or Aria (Opera) perform actions for you. Almost like a personal assistant facilitating everyday tasks.

AI browsers unlock productivity gains. They alert us towards security risks. Conventional browsers simply display content. AI browsers interpret, summarize, autofill and even execute tasks for you users. They provide autonomy layer and alert us to risks. Attackers assault this thinking layer, using advanced techniques such as prompt injection to influence browser behaviour.

OpenAI’s Atlas puts the assistant at the centre of your digital life. It has access to all authenticated sessions (SaaS, email, banking and corporate systems). In an AI browser, attackers do not need to trick the user. They instead trick the browser.

In prompt injection, they send malicious instructions on a webpage. It is barely visible or off-screen text. It overrides user’s commands. AI is unable to distinguish between what you typed and the instructions which secretly tells it to do. The assistant can be coerced to read emails, scrap calendars or move data between apps. The browser gets converted into an insider threat.

AI-powered browsers extend an LLM’s decision-making to web actions and authenticated sites. It creates new attack surfaces. Attackers can deliver malicious content through compromised web pages, deceptive emails. There are flaws in Comet (Perplexity). There could be data leakage.

There is an issue of privacy. On an agent which digests an injection, it can pivot to other tabs or SaaS. It behaves as if you are doing it. Another issue is tool misuse. An agent passes its own instructions. The third use is persistence. Attackers contaminate long-term memory. Bad introductions survive across sessions and trigger later.

Education and government are the most targeted sectors. There is silent data exfiltration from SaaS and Cloud apps.

AI browsers are candidates for high-risk attack surface. There should be strong AI governance. The risks cannot be managed simply by human oversight. There should be AI-powered defences, protecting you at machine speed!

As you know, TCS is India’s largest IT services company. It has planned to become the ‘world’s largest AI-led technology services company’. TCS has recorded an AI revenue of about $1.5 billion per annum.

About 54 of the top 60 clients use TCS for AI. Of all clients, 85 per cent leverage TCS for their AI work. Its quarter-on-quarter growth on AI has gone up by 16.7 per cent.

The company has executed over 5,500 AI projects and completed 209 platform deployments.

The first shift in TCS was from the mainframes to the web. The current shift is to generative AI. It is a fundamental shift due to the unprecedented speed and scale of its impact. This shift distinguishes itself from being mere technology upgrades of the past.

There is a strategic roadmap having five pillars for the company — internal transformation, redefining services, reimagination of customer value chains, having a future ready talent pool and expansion of ecosystem partnerships.

TCS has a vision of being the world’s largest tech services company. It values customer context and customer relationships. It also values the experience it has acquired. It focuses on strategic investments made. All this make it move towards the goal it has set.

There is a rigorous AI-first culture. In every project undertaken, the basic question the company asks if AI could do the job better. At times, it means cannibalization of its own revenue.

Among the world’s premier consulting firms such as PwC, EY and BCG, McKinsey & co is eminently placed. It marks its 100th year and is planning to celebrate it in Chicago, its birthplace. McKinsey is a go-to advisor for companies and countries. It has an honourable roster of clients from blue chip companies such as Coca-Cola and Goldman Sachs to the governments across the globe. It is the flag bearer of consulting industry. Its genesis can be traced to James McKinsey, a University of Chicago accounting professor, who set up this firm in 1926. Its first client was a meatpacker.

The partners of this company gathered in late October 2025 for confab which also served as a kickoff to its centennial celebration. Among those who were in attendance were Rio Tinto (Chairman, Dominic Barton), Ryan McInerney (Visa Inc. Chief), Condoleezza Rice (former US Secretary of State) and Oprah Winfrey (talk show legend).

Bob Sternfels, McKinsey’s global managing partner and a de facto leader, delivered a rallying cry of manpower reduction to the thousands of attendees. The non-client facing departments need to cut about 10 per cent of headcount across their business. That is equivalent to a cut of a few thousands. It could be done over the next 18 to 24 months.

This is the type of approach that McKinsey advises to many of its clients. It is the time to say ‘physician heal thyself’ by taking your own medicine. McKinsey’s revenue has plateaued in the last five years, after rapid hiring over the last decade. It has to improve its effectiveness and efficiency of support functions. Between 2012 and 2022, the head count climbed to 45000 from 17000. Since then, it declined to around 40000. The revenue earned is around $15 to $16 billion in the last five years.

Of course, the company would like to hire more consultants, but would like to reduce the support staff. It introduced Project Magnolia in 2023 to axe about 1400 jobs, but has avoided giving a code name to this year’s plan.

The company wants to escape years of lackluster growth. The present period is challenging for the industry as clients are turning cost-conscious, and there is a slow down in demand. Besides, the advent of AI and automation forced id to reduce 200 global tech jobs. Accenture asked the US government to reduce spending on consulting. Saudi Arabia was a major client till 2024 providing in fees to McKinsey at least $500 million per annum. It has started reducing the consultancy payments.

The company is ready to move ahead. It reconsiders its engagement with China and opioid makers. It cost the firm heavily in terms of penalties and legal settlements. Perhaps, they have set the ship on the right path.

Gen AI models may have wowed boardrooms, but a lot of effort may be required to manage the transition from GPT (generative retrained transformer) to enterprise grade AI. LLMs show some early utility (say summarizing documents) but they are far from production or enterprise ready state.

First of all, there are structural issues — integrating AI with the legacy IT systems, ensuring data generation, managing hallucination risks, regulatory compliance and maintaining explainability. AI roll out, thus, is tedious.

The IT firms can be enablers here. They are the ones who bridge the gap between AI models and usable enterprise applications. AWS and Microsoft are hyperscaler AI players. OpenAI, Google and Anthropic are model builders. These develop core technology — this core technology has to work in commercial organizations, say a bank or a pharma company or a logistics company. It is the last mile integration or operationalization. The real value lies here.

Enterprise solutions are thus like enterprise resource planning ( ERP) deployment. If an AI agent has to take approvals for HR procurement, it requires access to data (both structured and unstructured) across multiple systems and each with its APIs, security protocols and formatting challenges. Indian IT firms are well-versed in such integration.

Another important issue is customization of the models. LLMs are too generic. and there are hallucinations and context drift. these models need adaptation. the first step is fine tuning the model. there could be retrieval-augmented generation or hybrid architecture. Such architecture combines deterministic and probabilistic reasoning. All this means building complex orchestration layers. ML is applied as pipeline. There should be testing of outputs. Few enterprises could do all this in-house and there is a role to play for IT services companies here.

After deployment, these systems are to be governed and maintained. There has to be performance monitoring. There has to be retraining of the models with new data. Prompts will be refined, and feedback attended to. The outputs should comply with standards (both internal and external regulations).

Enterprise AI is beyond generic chatbot models or internal assistants. There are domain-specific applications. In banking, it can manage fraud detection. In manufacturing, it can manage supply chains.

There are monetization opportunities across the entire AI life cycle. There could be pre-implementation cosultation. Next, there is design and integration. Here there is data integration, prompt engineering, security compliance and system alignment. AI solutions must be embedded into existing IT system, Salesforce and Workday.

A new asset class comes into play — AI models and agents, monitoring tools, retraining pipelines, audit trails, governance dashboards and long-term support.

IT firms will not be able to do this with just talent pools. They should also build proprietary frameworks, accelerators and tools. Infosys has launched Topaz, TCS has launched Wisdom Next, Wipro is building ai360 and HCL has introduced AI Foundry. These are infrastructure wrappers. They deliver AI-as-a-service.

Only disciplined, methodical work will make AI usable and sustainable.

India will host The Global AI Summit in New Delhi in February 2026. Niti Ayog has released a report AI for Viksit Bharat drafted by McKinsey & Co. It spells out how AI could potentially reduce the gap between GDP today and the aspirational GDP under Viksit Bharat. Mainly, there could be automation of routine tasks and smarter decision making. AI could drive innovation, particularly in manufacturing, financial services, pharma and automotive. AI has the potential to boost GDP by $ 500-600 billion through productivity improvements for manufacturing and banking. Another increase of $280-475 billion can come through AI-driven R&D in pharma and auto.

Nasscom and BCG have drafted another report which examines AI’s impact on the tech sector. There would be job losses but there would also be new job opportunities. IT and BPO jobs are subject to automation. AI in education from schools onwards will create opportunities for talents, domestically and abroad. There could be reskilling programmes and building of infrastructure.

Deloitte has drafted another report AI for Inclusive Societal Development. India’s informal workers contribute (who constitute 90 per cent of workforce) neraly nearly half the GDP. This informal workforce should be transformed.

There should be collaboration between policy makers, academia and industry. AI should be used as a transformative power.

Apart from the four sectors pointed out in the first report drafted by Niti Ayog, there should be focus on real estate and professional services, other than banking. There are areas such as infra construction, textiles, hospitality, tourism and agriculture. AI must be integrated to these areas. AI can be helpful in food processing, logistics and storage.

The displacements in real world workforce must be mapped. AI’s impact on white collar jobs must be considered, apart from tech level and BPO jobs. Jobs lost cannot always be secured in the same sector. AI must access MSMEs without causing attrition. MSMEs contribute 60 per cent to India’s employment.

Rural India must be given enough attention.AI-related transformation must trickle down to the rural sectors.

The Summit must demonstrate how cutting-edge technology can be harnessed for inclusive growth.

In 2006, Yahoo! was ahead of Google in traffic but was slowly losing talent war. Yahoo! was run by Hollywood executive Terry Semel. It chased the ad deals. Google onboarded engineers. Yahoo! rejected engineers who had no computer science background. One such reject was Sparsh Gupta, a mechanical engineer from Delhi College of Engineering. The same Gupta set up a startup — Wingify , a visual website optimizer. His DCE batchmate Paras Chopra was a partner. Their startup was leading A/B testing platform serving Microsoft, Disney World, World Economic Forum, UNICEF, Vodafone Group, HBO and eBay.

The startup earlier launched one product which flopped. It was posted on Hacker. There was on brutal comment — ‘Don’t replicate what Google does for free. Find what Google is bad at. Become the best at that.’

Googles website optimiser required coding knowledge. Marketers hated writing the code. So Wingify built a visual version. Drag and drop. No developers were needed. They christened it Visual Website Optimizer. They gathered clients and kept growing steadily. They earned profits between Rs. 40- Rs.60 crore.

They are new competing with Optimizely and Adobe. In January 2025, Wingify stake of $ 200 million was acquired by Everstone. Chopra left to explore other avenues. Gupta has stayed back as cofounder. He retains most of equity.

The first chatbot introduced was OpenAI’s ChatGPT and remained a market leader. Since then, Google’s Gemini is trying to catch up and now the gap between them is narrowing. The jury is still out on who finally leads.

ChatGPT defined this category by its first-mover or pioneer advantage. It witnessed one of the fastest adoptions and near-total mindshare. In fact, it is accepted as another name of AI. However, since then Google’s Gemini 3 has no longer remained a chaser but has become a genuine challenger.

Google’s Gemini is powered by Gemini 2.5 Flash including Nano Banana, the image variant and Pro. It has been integrated to Google’s ecosystem (say Search, Gmail and productivity suites). It acquires contextual relevance.

It has 346 active users monthly. Amongst the downloads in India. It holds 50 per cent market share. Googles overall figures of monthly users are 650 million. In absolute number, ChatGPT’s monthly active users are 810 million, but its growth rate is slowing down. Even in desktop visits, Gemini is ahead of ChatGPT.

Google has added features such as Nano Banana and has launched Gemini 3. It has accelerated its adoption. At the beginning of 2025, Gemini had fewer than 100 million weekly active users, as compared to over 200 million for ChatGPT. The gap has now narrowed sharply. Nano Banana has enhanced Gemini’s multi-modal capability. Gemini is preferred by enterprise and research-oriented users who put a premium on accuracy and efficiency, whereas ChatGPT has a creative flair.

There are other challengers in the market — Perplexity and Claude. Both recorded triple-digit growth in 2025. Rather than who was the first, what matters in the market is who is the best now. Besides the newer systems better understand the intent of the user and therefore require fewer prompts.

Perplexity is strengthening its citation-based research. It is a precision search tool, whereas others are ‘needle-in-the-haystack’ versions. Claude has coding depth (especially its 3.5 Sonnet and Opus variants). In enterprise automation and regulated sectors, the factors such as reliability, governance, and compliance matter more than colloquial flair.

Progress is not to be measured in terms of higher reasoning scores only. The industry has to focus on smooth workflows, better planning and natural grasp of intent. This gives rise to on-device AI. Gemini has redefined intelligence. It does not route every action through cloud. Many things are worked locally resulting in lower latency, better privacy and usefulness even when connectivity is so so.

The transitional testbed in India is that we have a variety of devices and a huge mobile-first population. There is zero-swift-cost as GPT Go is free. Gemini is bundled with Jio. Perplexity is boarded with Airtel. If instant intelligence is provided by the model as well as reliability, such a model can work anywhere in the world.

OpenAI is alive to this competition and is trying to improve personalization, reliability, image generation and overall user experience.

AI models must show stronger alignment with user intent. Just better linguistic fluency is not enough. There should be content retention, reasoning framework and task abstraction.

The matter is now for competitive advantage. There should be lock-in with ecosystem through habit formation and workflow intelligence. It is a question that has shifted from the model being the best to a model that knows the user best.

Intelligence should flow effortlessly. Systems should understand intent. They should paraphrase imperfect phrasing. They should be responsive regardless of network conditions. Gemini has entered this battleground. It throws a challenge. Can ChatGPT meet it?

Model depth is okay, but at the same time there should be excellent distribution and usability. OpenAI is working on this by introducing GPT 5.2.

The Department for Promotion of Industry and Internal Trade (DPIIT) released a paper to resolve the issue of the use of copyright data in training the AI models is to be seen as infringement of the copyright laws or just a fair dealing exception. the matter acquires urgency as courts have to decide the matter as there are cases filed about this in India as well as abroad.

It is obvious that governments would like to promote AI and developers cannot be expected to negotiate with a large number of aggrieved since their works are used for commercial gain without consent or compensation. the paper’s aim is to create a framework where developers get lawful access to training data while creators receive a guaranteed payment without litigation.

The jurisdictions in Asia and parts of Europe have chosen a route of broad text-and-data-mining (TDM) allowing AI training without permission. The right holders can have opt-out rights. The DPIIT discards this arrangement on two grounds. The creators would not have a structured way to claim compensation. Besides, considering India’s size and diversity, administration of opt-outs would become complex. The fairness doctrine underlying around unpaid use will have to be sorted out first.

The paper proposes a hybrid model summarised as One Nation, One License, and One Payment. AI developers should get a blanket license (mandatory) to use all lawfully accessed copyrighted work for training models. Creators, on their part, get a statutory right of remuneration, with royalties payable only when AI systems are commercialized and start generating revenue.

These royalties will be set aside by a designated autority, subject to judicial review. The royalties will be collected and distributed through a central mechanism designed as a single window.

There are objections to this since creators do not have the right to refuse the use of their work for AI training. There are constitutional issues of the copy rights being equivalent to property rights. Nasscom objects to centralised agency being created fo channelizing royalties calling this a tax on innovation.

The Committee has said its reasoning is rooted in AI scale and its opacity. There is wide scraping from Internet and data access from billions of data points. It is impossible to negotiate permissions on case-by-case basis or to grant each creator a veto. That will halt innovation or lead to an uneven playing field where Big Tech swallows all the necessary rights.

A single-window statutory license removes the need to sign up hundreds of bilateral contracts. It is conducive for startups.

The paper invites suggestions and feedback over the next month.

AI models affect content creators since their works are absorbed into AI system without compensation. As we know, there has to be some regulation of the use of copyrighted material in AI training. This is an issue both in India and in other jurisdictions. Generative AI models ingest large volumes of copyrighted material. The Department for Promotion of Industry and Internal Trade (DPIIT) has released a working paper on regulating the use of copyrighted works.

The paper rejected the extreme positions. There cannot be free use of text-and-data mining (TDM). There is a suggestion for case-by-case licensing of copyright material. But it is not practical considering the scale and architecture of AI systems.

It is not possible to disaggregate training datasets into neatly identifiable units attributed to specific individual right holders. Granular permissions would be slow, fragmented and unworkable. It works in favour of the elite with resources having the capacity to negotiate at the cost of startups and small developers. Such costs could be absorbed by large firms, but not the startups and small developers.

The DPIIT would like to adopt a hybrid approach. There is statutory blanket license for all lawfully accessed works. Besides there should be remuneration right which should be activated on commercialization. It leads to legal certaintly.AI developers will not have to deal with a large number of individual permissions and yet the right holders would have a provision to claim predictable compensation. What is rejected is upfront payment and what is allowed is linkage to commercial deployment. It prevents prolonged litigation.

Another issue raised in the paper is retrospective effect. In principle, past training activities are to be brought into the remit of the new framework. However, this has several challenges. It is difficult to assess the extent of past training. Such models cannot be unwanted. There has to be calibration of retrospective royalty obligations, which is difficult. It could be arbitrary or prohibitive.

The nature of the license is mandatory. Here there are legitimate concerns. Individual consent is affected when the right holders refuse inclusion of their works. Individual consent is on par with property right. It converts private creative output into public use resource. Even if it attracts compensation, it goes against constitution. There is uncertainty about a central royalty — collection and distribution authority.

Nasscom suggests a broad TDM exception with opt-outs. Compulsory royalties in fact amount to innovation tax.

The paper is in the public domain. The majority view of the committee cannot be the final answer. Suggesation have to be invited from the public. There is a need for coherent legal framework. Blanket license is not an ideal solution. There should be pragmatism. AI training is structural. The legislation should use logic to make it implementable.

The European Union boasted of the sobriquet of lifestyle superpower — it has the highest living and social standards in the world. It is a matter of joke for others who are aware people leaving Europe and moving to Dubai or San Francisco for less tax or better pay. As compared to the US, there is better income inequality, social mobility and life expectancy in the European Union (EU). European living standards are almost on par with those of the USA. Finland is the world’s happiest country ( population 5.6 million).

The issue is despite this, Europe has not become a superpower. Though quality of life is a plus point, but it has not made the geography immune to power politics. Europe has weaknesses as defence-free region, energy importer and technology laggard. Finland is witnessing recession and unemployment is rising. It is not comforting for Europe to see Americans going to Switzerland, with reduced academic freedoms and showing crypto cronyism. The European lifestyle is harder to maintain on weak growth.

Russia has become more aggressive. China wants to export its products aggressively. Europe will have to show dynamism to face this environment. It has to control its deficit and debt. It should focus on strategic investments in defence and human capital. (since the US is no longer a magnet to attract talent). Europe should be resilient and should attract both brains and talent. The gap between productivity of the US and Europe must be reduced. Demark has adopted a policy of hire and fire easily ensuring that there is safety net between jobs. It could lead to an innovating culture in Europe.

The advantage of lifestyle superpower could be lost by Europe unless there is productivity boost. Europe should embrace resilience as its DNA.