MUMBAI: The past few months have produced an unusual alignment. Anthropic, OpenAI, Meta, Amazon and Microsoft have all moved past the model race and into the messier business of getting AI to work in production.
Anthropic teamed up with Blackstone, Goldman Sachs and Hellman & Friedman to launch a roughly $1.5 billion enterprise services firm, then bought applied AI company Fractional AI. OpenAI followed with the OpenAI Deployment Company and the acquisition of Tomoro, folding forward-deployed engineers into the house. Meta and Amazon have taken a similar path, embedding engineers inside customer organisations so AI becomes part of daily operations rather than another line item. Microsoft’s Frontier Company, backed by thousands of engineers, consultants and industry specialists, points the same way.
Taken one at a time, these look like ordinary expansion plays. Lined up together, they tell a single story. Enterprise AI has left the “can we build a better model” phase. The harder test is whether the model does anything for the bottom line.
The implementation gap
For two years the conversation was about the models: what they could do, how they scored on benchmarks, which lab shipped the latest leap. Inside companies a blunter question has taken over. How do we get this thing to work for us?
That is a different problem.
Deploying AI is rarely a plug-in exercise. Data has to be cleaned and re-engineered. Workflows have to be rethought. Processes have to be redesigned. Teams have to live with the change. The technology has to line up with what the business is actually trying to do. For many organisations, that turns out to be harder than choosing a model.
The research is unkind. MIT’s 2025 report, The GenAI Divide: State of AI in Business, found that nearly 95 per cent of organisations had seen little or no financial return or even negative returns from generative AI. The technology had not failed. Companies had treated AI like traditional software and layered it onto existing ways of working instead of redesigning those ways of working.
Where the wheels come off
The bruises show up in the same places. Data is scattered across systems that were never meant to talk to one another, incomplete, inconsistently labelled, or locked behind departments that treat it as territory. A model trained on that mess does not get cleverer. It gets faster at being wrong.
Then come the pipes. Legacy ERPs, core banking stacks, shop-floor software and customer systems were built for a different era. Bolting a large language model onto them without rewriting the process around it produces a shiny chatbot sitting beside the real work. Integration, latency, access control and audit trails eat the pilot. So does the human bit: middle managers who will not surrender a workflow, frontline staff who do not trust the output, and legal teams staring at hallucinations, bias, data leakage and a regulator who has not finished writing the rules.
Measurement is another trap. Outfits count pilots launched and seats licensed, not cycle time cut, error rates down or rupees saved. Without a baseline and an owner, AI becomes theatre. Security and compliance tighten the knot further especially in banking, healthcare and the public sector, where one sloppy prompt can push sensitive data into a third-party model. Skills are scarce on both sides of the table: people who know the vertical, and people who can make the model behave inside it. That is why specialised implementation is becoming the scarce skill.
Wiring a model to an API is increasingly commodity work. Understanding how a manufacturing plant, a retail supply chain, a bank or a healthcare network actually runs is not. Verticals do not resemble one another.
Hence the new titles: forward-deployed engineer, AI deployment specialist, AI solutions architect, AI success engineer. Different badges. Same job. Close the gap between what AI can do and what the business needs it to deliver.
What the shop floor teaches
The shift is familiar. Enterprises do not just need access to AI. They need someone in the room who can apply it to problems that matter sitting with customers, learning how the organisation really works, and dragging a proof of concept into something that survives production.
The problems differ by outfit. The pattern does not. The sophistication of the model is rarely the difference. What matters is whether the people implementing it understand the business, have the right people aligned, and weave AI into how work gets done rather than parking it beside the work.
That is why implementation is becoming the real differentiator in enterprise AI.
As models become more accessible and more alike, advantage will sit elsewhere: in an organisation’s ability to deploy AI, fold it into processes and produce outcomes that can be measured.
The next phase
The Anthropic, OpenAI, Meta, Amazon and Microsoft announcements matter less for the technology than for the priority they signal.
Value is not created the moment a model ships. It is created when that model changes how an organisation works. For enterprises, the scoreboard is shifting. Success will no longer be counted in pilots launched or foundation models adopted. It will be counted in how deeply AI is embedded and what it returns.
The models will keep improving. The bigger challenge, and the bigger opportunity, is helping outfits turn those capabilities into business transformation that shows up on the P&L.