A product manager opens a Jira ticket on a Tuesday morning: "Evaluate CRM add-on, $18k/year." By Thursday the engineering lead has closed it with a comment: "Claude Code knocked out a working integration in six hours. Skipping the renewal." That scene played out - in some form - at roughly one in three organizations over the past year.
McKinsey's State of AI: Global Survey 2026, released August 25 after surveying 1,719 participants across 97 nations, found that 32% of organizations have decided against buying one or more software products or features.
They built the functionality themselves with agentic coding tools. In the technology sector, the share hit 41%. The number traveled fast. What did not travel with it is the second number buried deeper in the same report.
The number McKinsey buried next to the headline
In the same survey, the share of organizations reporting that AI contributed to their EBIT sat at 37%, unchanged from a year earlier. Companies are building more and buying less, and the financial results have not moved.
That is not a contradiction - it is a lag. But it points to something real: cancelled purchase orders are not the same as saved money. Coding agents compress the initial development phase, which is roughly 30% of a system's lifetime cost, and do almost nothing to the remaining 70% spent on maintenance, security, integration drift, and eventual replacement.
There is a second cost most build analyses skip entirely. There is a new cost that most build models omit: AI operating and inference expense. McKinsey's 2026 survey found one in five organizations already limiting AI use because of operating costs, with the heaviest builders hitting cost constraints on coding agents about three times as often as everyone else.
McKinsey's agent economics guide estimates $20,000 to $30,000 annual run cost for a single-agent workflow at some banks, and $100,000 to $200,000 for a multi-agent team, driven by fixed infrastructure and orchestration - and that is a per-workflow figure, not per-company.
What the tooling actually looks like right now
Two releases this week made the build option more credible than it was six months ago - and also sharper in its tradeoffs.
Open-source OpenHands, an autonomous coding agent, reached its 1.0 release with production-ready Docker sandboxing, built-in security policies, resource limits, a plugin system, and benchmarks showing it can autonomously complete about 68% of SWE-bench Verified tasks. That is a meaningful number: SWE-bench Verified uses real GitHub issues, not toy problems. A self-hosted OpenHands stack costs roughly $0.20 to $1.05 per resolved task at H100 GPU rates, depending on which model you put behind it. That is the agent harness cost. Your model API bill sits on top if you use hosted models, or disappears entirely if you run open-weight models on your own GPUs.
GitHub Copilot Workspace now supports multiple specialized AI agents working simultaneously on different parts of a codebase, with separate agents for implementation, testing, and documentation that coordinate via a shared context window.
Each agent session runs in its own isolated git worktree, so several agents can work the same repository in parallel without overwriting one another.
And from Apple: Xcode 26.3 introduced support for agentic coding, a new way for developers to build apps using coding agents such as Anthropic's Claude Agent and OpenAI's Codex. With agentic coding, Xcode can work with greater autonomy toward a developer's goals - from breaking down tasks to making decisions based on the project architecture and using built-in tools.
Agents can search documentation, explore file structures, update project settings, and verify their work visually by capturing Xcode Previews and iterating through builds and fixes.
The tooling is no longer a research preview problem. The question is whether your team has thought through what happens after the agent ships the code.
Where the build decision actually pays off - and where it doesn't
Not all software is equally rebuildable. The McKinsey data points to a pattern even if the report does not call it out plainly: the technology sector leads at 41%, followed by healthcare payers and providers at 39%, and professional services and energy at 38%.
Insurance and the public sector, at 19% and 17%, are buying audit evidence and liability - not code.
That gap is the real signal. Internal builds beat a SaaS purchase when:
- The workflow is narrow and stable - a bespoke data pipeline, a Slack integration, a one-way sync between two internal tools
- Your team can own the security posture and has a static analysis gate before anything ships
- The replacement vendor charges for features you do not use, making the surface area of what you actually need small
- You can model three years of maintenance at 15-25% of the build cost annually and still come out ahead
Buying still wins when:
- The vendor absorbs compliance (SOC 2, HIPAA, GDPR) you would otherwise certify yourself
- The product category evolves quickly - you would be chasing a moving target with every sprint
Veracode's 2026 GenAI Code Security Report tested more than 100 AI models across standardized code-generation tasks; the average security pass rate was 56%, meaning 44% of AI code-generation tasks introduced a security vulnerability in testing
- and your team does not have a review layer to catch that
The non-obvious consequence of the McKinsey finding is what it does to pricing power. McKinsey's own 2024 analysis anticipated something close to this shape, projecting generative AI would add 1 to 3 percentage points to total SaaS churn and shift 2 to 4 percentage points of spend from buy to build over three to four years, roughly $35 to $40 billion reallocated. A few points of reallocation is close to invisible in top-line software spend numbers - but it concentrates on the vendors selling features that are easiest to replicate with a well-scoped prompt and an afternoon.
Beagle in action#engineering, Thursday 11:02amAI coding agents and build vs. buy: common questions
What does the McKinsey 32% figure actually mean?
Nearly a third of respondents (32%) report that their organizations have decided against buying one or more software products or features because they could be built internally with agentic coding tools. The survey drew 1,719 responses across 97 countries, weighted by each country's share of global GDP, so it is not a niche US tech-sector read. It measures intent and past decisions, not measured ROI.
Does building with coding agents actually save money?
Not automatically. Coding agents cut the initial development phase, but the cost of writing software fell while the cost of owning software did not. Security reviews, maintenance, and integration drift are not automated away. Model before you build: estimate three years of upkeep at 15-25% of initial build cost annually before concluding a build beats a subscription.
How good are agentic coding tools right now?
Agentic systems operate in persistent loops, breaking down high-level objectives into executable steps, and use a suite of tools - such as the file system, terminal, and version control - to explore codebases, recover from errors, and manage complex, multi-step tasks. OpenHands 1.0 hits 68% on SWE-bench Verified autonomously; GitHub Copilot Workspace now runs separate specialized agents for implementation, testing, and documentation simultaneously. For well-scoped internal tooling, they are production-viable. For novel architecture or security-sensitive systems, human review is still the gate.
What kinds of software are hardest for coding agents to replace?
Anything where the vendor's value is compliance certification, liability, or proprietary data - not the code itself. Software vendors face increasing pressure to justify licenses with capabilities that are hard to replicate as agent scripts, such as proprietary data, specialized workflows, or guaranteed compliance and support. A SOC 2-certified SaaS product is not just an API with a nice UI; it is an audited set of controls your internal build will have to match independently.
Should high-performing AI teams build more than they buy?
The trend is most pronounced among what McKinsey identifies as "high performers" - the 6% of respondents who attribute at least 5% of their EBIT to AI. Nearly half of these high performers are skipping software purchases, compared to 31% of their peers. That correlation is real, but causation runs both ways: teams with the AI maturity to build reliably are also the teams with the maturity to evaluate when not to.