For the past month, my inbox has been dominated by a single category of news: another frontier model, another benchmark, another valuation that defies arithmetic. Anthropic shipped Claude Opus 4.8 just 41 days after Opus 4.7, then closed a round that valued it at a record 965 billion dollars, vaulting it past OpenAI. Google launched the Gemini 3.5 family. Alibaba's Qwen3.7-Max broke the million-token barrier and cracked the global top five on the Artificial Analysis index. The temptation, if you lead AI inside a large enterprise, is to treat all of this as the main event.

It is not. After two decades operating at the intersection of technology and large-scale business, I have learned to watch the distribution layer, not the supply layer. And the most consequential thing that happened in May 2026 was not a model release at all. It was the speed and scale at which the world's largest professional-services firms wired these models directly into their delivery engines. That is the development that will actually redraw the enterprise AI map, and it carries a particularly sharp edge for those of us building from India.

The signal is the services layer, not the leaderboard

Consider what landed in a single month. KPMG signed a global alliance with Anthropic to deploy Claude across more than 276,000 employees in 138 countries, launching a "Digital Gateway" aimed first at tax clients and private-equity firms. PwC expanded its own Anthropic alliance, committing to certify 30,000 U.S. professionals on the path to its full 364,000-person workforce, and citing insurance underwriting cycles compressed from ten weeks to ten days. EY and Microsoft committed over a billion dollars across five years, with EY scaling Copilot from 150,000 to more than 400,000 of its own people and reporting a 15% productivity uplift. And ServiceNow and Accenture launched a forward-deployed-engineering program built around 300-plus pre-built agent skills governed by an AI Control Tower.

Read those announcements together and a pattern jumps out. The verbs have changed. We have moved from "experiment" and "pilot" to "scale," "certify," "deploy enterprise-wide." The Big Four are not buying AI to sell advice about AI. They are becoming the channel through which agentic capability reaches every Fortune 500 boardroom. When a firm makes itself "Client Zero" and runs the technology across its own workforce before a single client signs, it is doing something procurement decks rarely capture: it is converting a model into a repeatable, governed, billable workflow. That conversion, not the underlying model, is where enterprise value now accrues.

Why agents crossed the chasm in 2026 specifically

People assume agents finally work because the models got smart enough. That is only half the story. What actually changed is that the missing pieces around the model arrived at the same time, and the services firms had the incentive to assemble them.

Opus 4.8's headline feature is not a benchmark; it is "Dynamic Workflows," which plans work and coordinates hundreds of parallel subagents in a single session, with a claimed roughly fourfold reduction in letting code flaws pass unremarked. Google, at the same moment, shipped a Managed Agents API that spins up an agent inside an isolated Linux container with a single call, on a model with a million-token context window. These are not capability features in the old sense. They are operational features: parallelism, isolation, reliability, auditability. They answer the questions a Chief Risk Officer asks, not the questions a researcher asks.

That is precisely why a services firm can now sell agents with a straight face. The reliability and containment improvements are what make an agent insurable. When Opus 4.8 is claimed to be markedly less likely to let flaws in its own code pass, that is a sentence a partner can put in front of a regulated client. The economics follow: a model that needs less human babysitting is a model whose labour can be marked up. The services layer monetises reliability, and 2026 is the first year the models shipped enough of it.

The structural threat to the offshore model is now explicit

Here is where I have to be candid about the view from India. The same wave that empowers the global SIs threatens the business model that built much of Indian IT. Reporting this month put it bluntly: the new private-equity-backed enterprise ventures from OpenAI and Anthropic are being viewed as the most serious structural threat to India's offshore services model since 1990s outsourcing. When the underwriting cycle collapses from ten weeks to ten days, a meaningful share of the billable hours that flowed to offshore delivery centres simply evaporates.

The reflex response is defensive, and it is wrong. The constructive response is already visible. Wipro's expanded ServiceNow tie-up to scale agentic workflows sent its shares up nearly 5% and lifted the Nifty IT index almost 3% in a single session, because the market understood the move: Indian firms pivoting from selling labour-hours to selling governed agent deployment. The question every Indian services leader must answer is whether their firm becomes the integrator of agents or the casualty of them. There is no neutral middle.

  • The work that compresses is repeatable, rules-based, human-in-the-loop processing: the exact category offshore centres scaled on.
  • The work that expands is agent orchestration, control-plane design, domain fine-tuning, and the governance scaffolding that lets a regulated enterprise trust an autonomous workflow.
  • The firms that win will move people up that value chain faster than the work disappears beneath them.

Coding agents are the leading indicator for everything else

If you want to know what is coming to finance, legal, and operations, watch software engineering, because it is the function furthest down this road. Gartner now sizes the enterprise AI coding-agent market at roughly 9.8 to 11 billion dollars annualised and predicts that by 2027 more than 65% of engineering teams using agentic coding will treat the IDE as optional. Read that last clause slowly. The integrated development environment, the cockpit of the profession for a generation, becomes optional, because the human is supervising fleets of agents rather than typing in an editor.

The capital markets agree. Cognition, maker of the autonomous engineer Devin, raised over a billion dollars at a 26 billion valuation, more than doubling in eight months. Meanwhile the tooling is converging on safety-by-default: Claude integrated into Snyk's security platform, and Opsera embedding DevSecOps guardrails directly inside Cursor's IDE. The lesson for enterprise leaders is not "buy a coding agent." It is that every knowledge function will be reorganised around supervising agents rather than performing tasks, and the organisations that redesign their roles now will not have to do it under duress in 2028.

The governance bill is coming due, and it is not optional

None of this scales without governance, and the regulatory ground is shifting faster than most roadmaps assume. EU negotiators reached a provisional agreement on the Digital Omnibus on AI, the first amendments to the AI Act, postponing some high-risk obligations while adding hard new prohibitions. Colorado's governor signed SB 189, repealing and reenacting its landmark AI Act with effect from January 2027. And in a genuine reversal, the Trump administration is reportedly weighing federal pre-release evaluation of advanced models, driven by national-security alarm over cyber-capable systems.

That alarm is not abstract. Anthropic said it expects to bring its previously restricted, cyber-capable "Mythos"-class models to all customers in the coming weeks. Models that can find and chain software vulnerabilities are about to become a procurement option. The liability surface is widening in parallel: CNN sued Perplexity for copyright infringement, reportedly the first such suit by a television network; five major publishers and author Scott Turow sued Meta over training on pirated books; and Pennsylvania sued Character.AI for chatbots posing as licensed doctors. The through-line for enterprises is simple and unforgiving: an agent acting in your name carries your liability. The control plane, identity, audit, sandboxing, spend limits, is now the product, not a compliance afterthought.

India's moment is real, but it is a window, not a guarantee

I am, despite the services squeeze, genuinely optimistic about India, because the same forces that threaten the old model create the new one. The physical build-out is underway: after the India AI Impact Summit's 200-billion-dollar-plus in pledges, gigawatt-scale data centres are under construction in Gujarat, with around 120 MW expected online by year-end and plans to add 20,000-plus GPUs toward a 100,000-GPU national target. The startup base is thickening, from AI-chip firms like Tsavorite and BigEndian raising pre-Series-A rounds to the hard, unglamorous problem of voice AI across India's languages and accents. Even the hyperscalers are courting the ecosystem, with Google's AI Immersion Program shortlisting Indian startups for a Bengaluru cohort.

But infrastructure and capital are necessary, not sufficient. The model layer is global and commoditising; the compute is being financed at a scale, KKR's 10-billion-dollar Helix infrastructure venture, US data-centre private-equity hitting a five-year high near 45.7 billion dollars, that India cannot match dollar-for-dollar. India's edge is not capital. It is the dense layer of domain talent that knows how a bank, an insurer, or a manufacturer actually runs. That is exactly the layer agents need to be wrapped in to be trustworthy. The window is the next eighteen months, and it closes the moment the global SIs finish encoding that domain knowledge into their own agent libraries.

My take

I am telling my own organisation to stop reading model releases as scorecards and start reading them as supply-chain updates. The model is a commodity input arriving from a handful of suppliers; the durable advantage sits in the orchestration, governance, and domain wrapping above it. Concretely, that means three commitments this year. First, build the control plane before you scale the agents, because liability now travels with autonomy. Second, retrain aggressively up the value chain, because the work that compresses and the work that expands are different work, and people do not migrate between them by accident. Third, treat India's domain depth as a strategic asset to be encoded into governed agents now, while the window is open, rather than ceded to someone else's library later.

The firms that mistook this moment for a model race will spend 2027 wondering why their cleverest pilot never reached production. The ones that understood it was a distribution and governance race will have turned agents into auditable, dependable colleagues, and turned India's domain talent into the moat that the commoditised model layer can never be. That is the work in front of us. I intend to do it before the next board cycle, not after.