Walk into almost any business in India in 2026 and you will find AI somewhere. A chatbot on the website, a copilot in the finance team, an agent drafting first replies. What you will not always find is a result anyone can point to. That gap, between using AI and getting value from it, is the most expensive thing in the market right now. This is a map of where the money goes, why so much of it misses, and what we would build first.
The ground moved
For years, AI in India was a side project you ran on the edge of the real work. That era is over. The domestic AI market was worth around 9.5 billion dollars in 2024 and is forecast to pass 130 billion by 2032. Roughly 88 percent of organisations now use AI in at least one part of the business, and India ranks third in the world for AI activity on the Stanford index, behind only the United States and China. AI stopped being an experiment and became part of the plumbing.
That is the good news, and it is also the trap. Once something becomes infrastructure, spending on it stops being optional, and the cost of doing it badly stops being small.
Adoption is not the same as impact
Here is the number that should stop you. Adoption is nearly universal, yet by one widely cited 2026 index only about 27 percent of companies report a measurable business impact from AI. Separately, industry estimates put the share of AI projects that fail before they ever reach production at around 71 percent. The problem is not a shortage of AI. It is a surplus of AI that does not move a number anyone is accountable for.
The cause is rarely the model. It is the process around it. Most deployments bolt an agent onto a workflow that was already broken, and a faster broken process is still broken. The companies that actually see returns are, by a McKinsey estimate, about 2.8 times more likely to redesign the underlying process than to decorate it. The lesson for an owner is blunt. If you are not willing to change how the work is done, do not pay for AI to do the same work faster.
The cost nobody priced in
There is a second reason the money misses, and it hides in the invoice. The economics of AI flipped when the field moved from simple generation to agents.
A single generative customer reply in 2023 cost about four cents. By 2026, one agentic run, where the system plans, calls tools, fetches data, and spawns sub-agents, costs around a dollar twenty. That is roughly a thirtyfold jump for what looks, from the outside, like the same chat box. Multiply it across every employee running it all day, and the bill behaves nothing like a tidy software subscription.
The wider picture matches. Industry research has infrastructure and model-consumption costs climbing 20 to 30 percent a year, while the labour they replace saves perhaps 5 to 10 percent, and infrastructure is heading past 45 percent of enterprise technology spend. Agentic systems are non-deterministic, which means their cost is not a fixed line you can look up. It is a design decision you make, or fail to make, every time you pick a model, write a prompt, and decide how many times an agent may loop. Price it at the whiteboard, not at the end of the quarter.
The compliance clock is a build constraint
While all of this unfolds, a hard deadline is moving toward every business that touches the personal data of an Indian resident. The Digital Personal Data Protection Act and its 2025 Rules come into full force on 13 May 2027, with no grace period and penalties that reach 250 crore rupees, around 30 million dollars, for a single failure.
It is tempting to file that under legal and move on. That is the mistake. For anyone who builds software, the Act reads as a list of engineering requirements:
- Consent has to be free, specific, and unbundled. Pre-ticked boxes and buried terms are gone, and the notice has to be plain and available across the scheduled languages of India.
- You cannot protect data you cannot see. The first job is a real map of where personal data lives and how it flows, before anything else gets built on top of it.
- Retention has to be automated. Rule 8 even requires telling a person at least 48 hours before their data is auto-deleted, so deletion becomes a small workflow rather than a quiet cron job.
- A breach starts two clocks at once. Six hours to report to CERT-In, and seventy-two hours to inform the Data Protection Board and every affected person. You want that runbook written before you need it.
- And the hard one for AI: erasure. When a person asks to be forgotten, you can delete their record, but the model has already learned from it. Honouring erasure inside a trained model is an architecture problem, not a delete button.
Build these in from the start and they cost almost nothing. Retrofit them under deadline and you face the kind of overhaul that runs a mid-sized company several crore. The deadline is not the enemy. Leaving it to the last minute is.
You are not stuck with the obvious models
One more habit the market gets wrong is defaulting every project to a large Western model. India now has a serious home-grown stack. Models such as Sarvam, Krutrim, and BharatGPT, along with the government Bhashini language rails, handle 22 official languages, keep data resident in India, and often beat the global default on culturally specific tasks at a lower cost. For a business serving customers in Marathi, Bengali, or Tamil, or one that must keep data onshore for the Act, that is not a patriotic choice. It is the better engineering one. Match the model to the job, not to the brand on the box.
What we would build first
Put it together and the playbook is deliberately unglamorous. It is the same order every time:
- Start with a named outcome. If you cannot say which number should move, you are not ready to spend a rupee on the model.
- Map the data. Both the value and the compliance live in knowing exactly where it sits and how it moves.
- Pick the simplest system that solves it. Build is the rarest answer, not the default one.
- Design for the deadline. Privacy belongs in the foundation, not bolted on in a 2027 panic.
- Instrument the cost and the result. If you cannot see both the cost per run and the outcome, you cannot tell whether the AI is earning its keep.
That is the whole game. Not the biggest model or the most pilots. The clearest problem, the simplest build, priced and measured, and compliant before it has to be. It is the reason we diagnose before we build, and why the question we ask first is never which model, but which problem.
This piece is a working read of public industry research and the text of the DPDP Act, not legal advice. For your own obligations under the Act, talk to qualified counsel.