Companies are starting to turn down software purchases on the still-unproven bet that AI can help their own teams build what they need at a lower cost. Coding tools have made internal development plausible for businesses that once considered it beyond their reach, giving software vendors a new competitor in their own customers.
McKinsey’s 2026 State of AI study found that 32% of respondents whose organizations regularly use AI say their employers have decided against buying at least one software product or feature because they could build the functionality internally with agentic coding tools. Among technology respondents, the figure reaches 41%. In healthcare, it is 39%. Those decisions arrive alongside a stubborn gap in AI’s financial record. Eighty percent of respondents who use AI at work report better individual productivity, yet only 37% report a positive contribution to organizational earnings before interest and taxes, or EBIT, essentially unchanged from last year.
“A decision to build software is one of the few places where AI’s impact shows up immediately on the P&L, in the form of a purchase you decide not to make,” Tara Balakrishnan, an associate partner at McKinsey and part of QuantumBlack, AI by McKinsey, tells me. “So it is not surprising that it is an early area where companies feel value.”
The survey, which gathered responses from 1,719 participants across 97 countries, does not quantify the spending those decisions displaced.
However, the avoided bill is visible immediately. The cost of maintaining the replacement emerges over time, raising the possibility that AI could disrupt software purchasing well before companies know whether they made the right trade.
Where AI-Built Software Is Starting to Replace Vendors
Balakrishnan, who works with software companies and has contributed to McKinsey’s annual AI research since 2019, says the familiar build-versus-buy debate understates how limited the options once were for many businesses.
“Many companies that previously lacked the capability to build software in-house now have that option,” she says. “Historically, the choice for many companies has been ‘buy’ versus ‘do nothing.’ Agentic coding tools have lowered the threshold for what is worth engineering time, which means a category of work that sat below the funding bar is now getting built.”
Some of that development happens outside engineering. Balakrishnan shared the example of a sales leader whose frontline team used AI to create a tool that automatically updates forecasts for individual accounts in its customer relationship management (CRM) system. Employees previously handled that work manually, when they did it at all. The sales team still uses its CRM platform while creating functionality around it. Balakrishnan sees the shift most clearly among companies that develop software as their core product and those whose proprietary processes fit commercial tools poorly. “It’s a shift but it’s still closer to companies reclaiming spend at the margins of their stack,” she says.
An internal application may eliminate manual work without reducing the resources a company needs to run a department. The financial result, however, depends on what happens to that recovered capacity, and whether the benefit exceeds the cost of operating the new software.
“Individual productivity gains are real, but they’re diffuse. The time saved gets absorbed unless the organization and the work are redesigned around it,” Balakrishnan explained. “An AI transformation requires as much focus on change management as on technology, and rewiring how work gets done is slower than adopting a tool.”
The flat EBIT figure does not establish that internal development loses money, either. Respondents can report savings within individual business functions without attributing positive EBIT impact to AI across the company. The survey does not connect each rejected software purchase to a measured financial outcome. On whether companies are choosing to build too early, Balakrishnan says, “The jury is still out. Very few organizations have run the experiment long enough to know how the total cost of ownership for internally built software compares.”
Cheaper AI Tokens Don’t Guarantee Lower Dev Costs
Balakrishnan argues that the economics of AI coding agents extend well beyond producing code, which she says was never the costliest part of software development. Their broader value comes from changing how teams work throughout the development process, with implications for the cost of bringing software into use and maintaining it over time.
“We see agents close the knowledge gap around legacy codebases, retire technical debt, automate testing pipelines, and shorten the time it takes to prototype an idea and validate it against real market feedback,” Balakrishnan says. “For some organizations, the quality of products has improved with AI, with fewer customer-escaped defects and less technical debt. And what’s fascinating is that much of this work happens around the clock, with agents making progress on the ‘night shift.'”
Yet the organizations reporting the strongest AI results also encounter a distinctive expense problem. McKinsey’s high performers are twice as likely as others to scale software coding agents. Nearly half have passed on software purchases in favor of internal development, compared with 31% of other respondents. Moreover, those same high performers are about three times as likely to report that operating costs constrain their use of coding agents. The report puts the shares at 18% and 6%, respectively. That disparity does not extend to chatbots or other AI agents.
“You do not feel a cost constraint until you are operating at a scale at which it bites,” Balakrishnan says. “High performers report cost pressure because they are running these systems in production, at scale. It’s a symptom of being further along in the AI usage journey rather than a sign that the underlying economics have broken down.”
In the report, McKinsey senior fellow Michael Chui shared that token consumption has grown faster than per-token prices have fallen.
Balakrishnan points out that high performers continue investing. They are more than twice as likely as others to spend over 15% of their IT budgets on AI, and most intend to increase investment. She says some software companies have stabilized AI spending while expanding agent use. The emerging discipline involves tracking agent usage and measuring spending against results. A growing AI bill can accompany better economics if useful output grows faster. Continued investment alone does not establish that this is happening, or that building beats buying.
AI’s Threat to SaaS Could Start With Add-Ons
For software-as-a-service (SaaS) vendors, the immediate challenge may arrive through smaller purchasing decisions rather than wholesale platform replacements. A customer that keeps its core system can still question whether it needs to pay for an additional feature.
“The build-versus-buy framework may need to be applied at a feature-by-feature level going forward,” Balakrishnan says. “That said, there is a meaningful difference between a vibe-coded solution and a production-grade one. It takes real effort to make a tool secure and compliant before it can be used across a broader organization.”
Established vendors retain advantages in data access and network effects, where a product benefits from wider participation. Connections to other systems also matter. A tool that performs one task convincingly still needs to operate within the customer’s wider software environment, with appropriate security and regulatory compliance.
That complicates the “SaaSpocalypse” argument. Vendors may retain the underlying platform while facing tougher questions about paid extensions, as the sales-team example suggests. Their defense increasingly depends on the value of operating and connecting software, beyond whether customers can reproduce a particular feature. However, McKinsey’s findings do not establish which product categories face the greatest revenue exposure.
Only about 6% of respondents qualify as McKinsey’s AI high performers. They attribute at least 5% of organizational EBIT to AI and report significant value from its use. Nearly three-quarters say they have fundamentally redesigned workflows because of AI, compared with about one-quarter of others. “The fact that the share of high performers has stayed essentially flat tells us that the constraint is not access to technology,” Balakrishnan says. “What separates the 6% is that they have changed how work is organized around AI.”
The Control Layer Verdict
A software vendor can lose the sale before its customer discovers that building was the expensive option. Rapid AI innovations are now giving more companies the confidence to reject purchases before they understand the obligations they are taking on. Once colleagues depend on an internal tool, someone must own its failures and defend the cost of keeping it running. That is where greater freedom to build becomes a test of an organization’s ability to govern its own software. Every bug fix draws on the company’s own engineering budget, while continued use of coding agents adds to the bill. Unless the business counts those costs against the vendor’s price, it risks claiming subscription savings while spending more to maintain its replacement.
