AI Makes Software Cheaper to Build, Not to Own

For decades, the build-or-buy decision in enterprise software followed a familiar logic. Companies built software when their needs were strategic or unusual enough to justify the cost. They bought standard products when a vendor could supply them more cheaply and reliably.

That calculation is starting to change. Nearly a third of respondents to McKinsey’s latest AI report said their organizations had decided against buying at least one software product or feature because their own teams could build it instead.

That matters particularly in India. IT-services firms have long earned their living writing and running software for clients abroad. Multinationals also operate large technology hubs in the country, known as global capability centers, or GCCs. These handle engineering, analytics and back-office work for the parent company. 

A May 2026 report by Nasscom and Zinnov counted more than 2,100 of them, with many moving beyond execution toward owning products and platforms. Coding agents touch both groups. They change what clients buy from the services firms, and they give the captive centers more power to build in-house.

As more software can be built internally, both models come under pressure. Services firms may find clients buying less outside work, while GCCs gain more scope to take on projects themselves.

But the cost of writing code is only part of the equation. Software still has to be secured, connected to other systems, maintained, patched and eventually retired. Someone also has to be accountable when it fails. When a company buys a product, much of that responsibility sits with the vendor. When it builds one itself, more of the burden stays inside.

That is increasingly where the build-or-buy decision will be made.

Proprietary Software Is Moving In-House First

Nakul Jain sees the shift first in engineering-led companies whose teams are ready to look after software long after it launches. Jain is co-founder and CEO of APLYD, which helps governments use AI in public services, and a partner at Athena Infonomics. “The shift is not universal,” he says. “Where that readiness does not exist, buying continues to win, because a vendor also brings accountability that a hastily built internal tool does not.”

Shub Bhowmick co-founded and runs Tredence, a data science and AI services firm. He sees the strongest case for building in applications that depend on a company’s own data or reflect processes unique to the business. He cites AI agents built for a particular domain, personalization systems, and predictive analytics. A retailer might forecast demand from its own supply-chain data. A bank might spot fraud in its own transactions.

Bhowmick sees a weaker case for standard systems such as payroll, human resources, or core customer-management platforms. Established vendors already provide the security, scale, and compliance those systems need, and they carry the upkeep. “Build what makes you unique, buy what keeps the lights on,” he says.

Anupam Kumar Jha, who leads enterprise solutions engineering for Datadog in India, expects many companies to settle in between. Their own engineers will work on applications tied to proprietary workflows, he says. Specialists will supply the infrastructure, security, and operations tools underneath. Datadog, which sells software for monitoring cloud systems, is one of those specialists.

Microsoft found in September that Indian AI users were twice as likely as the global average to have redesigned their work around AI agents. Years of investment in cloud and digital engineering have left many Indian companies with teams that can build more themselves, Jha says. “Software is becoming the infrastructure for AI,” he adds.

Many GCCs began as offshore units carrying out work defined elsewhere. For a multinational whose center now owns products, the team that could replace a vendor’s product is already on the payroll. Coding agents let that team take on projects that once looked too small or too costly to build. Some of that work used to go to Indian IT firms.

Jain’s readiness test and Bhowmick’s data test add up to a simple screen. Build where the software runs on your own data or way of working, and only if a team will look after it for years. Buy the rest. Either way, someone has to understand each application, connect it to other systems, keep it secure, and answer for it when something breaks.

Faster Development Doesn’t Mean Lower Ownership Costs

Bhowmick estimates AI coding tools cut early development time by 40% to 50%, in work such as first drafts of code, testing, and documentation. Architecture, integration, security, monitoring, and maintenance still have to be handled after that. “Speed without discipline is just a liability,” he says.

Datadog found this year that almost 9 in 10 organizations were running software with at least one known, exploitable security flaw. Many of the services it examined relied on code libraries that were no longer being maintained. “Writing code, however, was never the whole cost,” Jha says. Each application a company builds adds to that load, and someone has to patch, monitor, and eventually retire it.

An enterprise that builds with coding agents saves on the first release. It inherits every patch, audit, integration, and retirement that follows.

Jain warns against assuming that faster coding means cheaper software. He argues the bigger change is that AI makes small, specialized applications worth building where custom work used to cost too much. That brings more tools to build and more to maintain. A tool that took days to build can still consume years of engineering time if it is hard to keep up or connect.

Retool, which sells tools for building internal software, surveyed builders late last year. More than half had made something outside IT’s oversight in the previous 12 months. Companies can end up trading a sprawl of software subscriptions for a sprawl of internal applications. Bhowmick warns that unchecked development breeds shadow IT, meaning software that runs without the IT department’s knowledge or shared standards.

Jha points to a slower drift. Separate teams may solve the same problem twice, pile up overlapping dependencies, and ship tools that work at first but grow costly to change. In India, ServiceNow found that only about 1 in 5 enterprises had processes to test, audit, and assess the risks of their AI. That was true even as their spending on AI soared.

Jan Wuppermann, who heads service assurance and data and analytics for NTT DATA in Asia-Pacific, points to a related constraint. Companies, he says, are introducing AI into systems never designed for the security, governance, and data demands now placed on them. “AI is running into architectural limits,” Wuppermann says. NTT DATA’s own research rates only a small minority of companies as AI leaders. Faster coding does not speed up integration, security approval, or production support. Those become the new queues.

From May 2027, India’s data protection rules will require safeguards such as access controls, monitoring, and a year of logs around personal data. The company that decides how that data is used stays responsible even when a vendor does the processing. Buying software never removed that duty. Building in-house adds the work of engineering each safeguard into every tool, and of proving it is there.

Vendors Now Compete With Their Own Customers

Products that mainly save customers from writing simple code are the most exposed. Basic reporting tools and approval workflows sit in that group. An internal team with AI tools can now rebuild much of what they do. Security, infrastructure, monitoring, integration, and compliance are harder to replace, because few companies want more responsibility in those areas. 

Jha expects consolidation rather than a retreat from vendors. As companies build more in-house, he says, they lose interest in buying a separate product for every technology problem. They want fewer platforms that can tie systems together and keep them secure and visible. Broad platforms gain, because more of a company’s software runs on top of them.

India’s IT-services firms have long priced much of their work by the person-month. Clients pay for the time of the engineers assigned to them. Wipro’s chief technology officer, Sandhya Arun, said in September that the company’s AI initiatives had increased productivity by the equivalent of 20,000 employees’ output. Those employees have since been redeployed within Wipro, and Arun said the gain did not necessarily mean a person-for-agent swap. When agents do part of the work, there is less to bill, even if the client needs just as much done.

Clients will still need outside technology expertise, but the work they will pay for is changing. Routine coding is a weaker offer when a client’s own engineers can produce more of it with AI. Jha expects demand to shift from labor-intensive routine coding toward architecture, modernization, integration, security, and governance. Bhowmick says clients want services firms that bring industry expertise and take more responsibility for outcomes.

For Indian firms that grew by adding engineers, the link between headcount and value to the client is weakening. They will have to sell fewer hours and more judgment. That shift favors firms with deep industry knowledge over those built mainly on scale.

A typical software contract spells out uptime targets, deadlines for security fixes, and credits when the vendor falls short. An in-house tool carries none of those promises unless the company writes them for itself. Jain says companies are not necessarily spending less on technology or dropping vendors. They are choosier about what deserves a license, what to build, and when an outside partner is worth the fee. Each vendor now has to show a client why it should not build the thing itself.