The Prototype Shipped Overnight. Year Two Is the Real Invoice.

The Question the Headlines Didn't Answer

Picture a familiar scene: a program manager walks into a budget meeting with a demo. An agentic coding tool produced a working case management prototype in a matter of hours, perhaps overnight. The team wants to build the whole system. Why pay for a platform license when you can own the thing outright and have something to show in 24 hours? It is a reasonable question, and it is being asked in technology offices across the country right now.

In March, we published "Are AI Tools Really a Threat to SaaS?" and argued that agentic coding tools were more likely to strengthen established platforms than replace them. Production readiness is a qualitatively different property, one that platforms like Salesforce, Dynamics 365, Power Platform, and ServiceNow have accumulated over years of investment in security, compliance, and upgrade management. Our closing line: the end of SaaS is not coming; a more iterative way of building on top of it is.

Now McKinsey has published its State of AI 2026 survey, and the headlines have been loud. Nearly a third of respondents (32 percent) report that their organizations decided against purchasing at least one software product or feature because they could build the functionality in-house using agentic coding tools. Among AI high performers, the figure approaches half. Forty percent of respondents at organizations over one billion dollars in revenue are scaling agents in at least one function, up from 27 percent a year earlier. McKinsey also reports that a meaningful share of respondents say AI operating costs, including token costs, are already constraining use of the technology. Financial press coverage of the findings has framed the data as a significant shift in enterprise software strategy.

These numbers deserve careful reading. The survey drew 1,719 responses across 97 countries, weighted by each country's share of global GDP. The weighting methodology and the survey's focus on organizations large enough to have formal AI programs means the results skew toward larger commercial enterprises rather than the state and local agencies or mid-sized organizations that make up much of our readership. The unit of measurement matters: 32 percent decided against buying at least one product or feature, at least once. This is a survey of decisions, not an analysis of outcomes, and that distinction matters more than the headline number. It does not record what those internally built replacements cost to run, secure, and maintain, because in most cases nobody has owned one long enough to know. The TCO analysis for "we built it instead" will show up in the 2027 and 2028 surveys. The organizations that skipped purchases in 2026 are the sample. Gartner has separately estimated that $234 billion in enterprise application software spend is now exposed to what it calls agentic arbitrage. The question is not whether the shift is real. The question is whether the organizations driving it have priced the full cost of the decision.

The Invoice That Hasn't Arrived Yet

The McKinsey survey is a snapshot of purchasing behavior, not a ledger of total cost. What it cannot see is the ownership tail that follows every custom build.

Start with token and compute run cost. Analysis of 2.4 billion enterprise API calls found that organizations routing every workload to frontier models paid $18.40 per million tokens while those using a tiered architecture paid a median of $2.31 per million tokens, an 87 percent cost gap driven by a single architectural decision. To put that in budget terms: a team processing the equivalent of 100 million tokens per month would pay roughly $1,840 per month under the frontier-only approach versus $231 under a tiered model, a difference of over $19,000 per year before any other infrastructure costs. Add ongoing security review and patching, which a platform vendor handles as part of its engineering operation and a custom builder must staff independently (see our September post on ownership costs for the full breakdown). For public sector organizations, add accessibility remediation: bringing a custom application into Section 508 and WCAG compliance can run $100,000 to $200,000 in year one alone. Then add documentation, dependency and upgrade management, and the cost of the next team that inherits the code. Academic research on technical debt demonstrates that dependency cycles cause changes to become costlier over time, compounding operational costs in ways invisible at the point of build. Deloitte's 2026 Global Technology Leadership Study, as cited by the Software Improvement Group, estimates that technical debt absorbs 21 to 40 percent of total IT spending.

We have laid out the full ownership argument in our September post, "Code Delivery Is No Longer the Bottleneck. Ownership Is." The structural point is that what a platform license fee has always paid for is the vendor carrying most of these costs on behalf of the buyer. The license is not just access to features. It is a transfer of maintenance liability.

The 32 percent figure measures what organizations decided to stop buying. The 2027 and 2028 surveys will begin to measure what that decision cost.

A Faster Build Is Not a Reason to Jump to the Custom End

To understand where agentic coding tools actually deliver value, place them on the solution continuum from "Making Sense of the AI Solution Landscape." That continuum runs from general purpose AI tools through embedded AI in existing applications, through low-code builders like Power Platform, through enterprise AI platforms, to purpose-built custom systems. Agentic coding tools change cost and speed at every point on this spectrum, and the nature of that change varies depending on where you are.

At the embedded AI and low-code end, agents accelerate configuration, generate platform-native components, and compress the feedback loop without adding ownership burden. At the custom end, they reduce initial build time but do not reduce the surface area of what must be owned afterward. The organizations most at risk are those treating a faster build as a reason to jump straight to the custom end, as if the speed of construction changes the cost of occupancy.

The data on what happens next is sobering. Gartner predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, and inadequate risk controls. Independent analysis finds that 88 percent of enterprise AI agent pilots never reach production, with failure driven not by model quality but by the absence of governance controls that production systems require. This is the gap we examined in "The Last Twenty Percent Problem in GenAI Solutions": AI has made the prototype cheaper and faster, but it has not shortened the last twenty percent. KPMG's advisory framework for the build, buy, or borrow decision captures a meaningful signal from its own proprietary AI Pulse Survey: 57 percent of organizations surveyed by KPMG now favor a blended approach to building and buying AI agents, up from 51 percent the prior quarter. If KPMG's survey is directionally representative, the market may already be moving toward the middle of the spectrum.

The Platform You Already Own Is Already Using AI

One thing is easy to overlook when a build-from-scratch proposal is on the table: the organization may already own much of what it is proposing to build. Many enterprises hold Microsoft 365 and Power Platform entitlements that cover a large share of the functionality their teams are now scoping as custom work. In state and local government, that overlap is unusually well documented. StateScoop reporting shows that 3,866 state and local government agencies have access to Microsoft Office 365, a figure that reflects the scale of existing Microsoft entitlements in the sector even if the precise count has shifted since that data was published. Gartner research, as reported by Insentra Group, finds that organizations without centralized SaaS management overspend by at least 25 percent due to unused entitlements, with Microsoft 365 representing both the biggest consolidation opportunity and the most common source of underutilization. The NASCIO 2026 State CIO Top 10 Priorities report makes the context explicit: artificial intelligence ranks first among state CIO priorities, cloud solutions rank third, and SaaS ranks fifth. This audience is actively navigating exactly this tension.

The comparison that matters is not between an agentic coding tool and a static platform. As we argued in March, the platforms themselves have been absorbing AI into their architecture. Microsoft's Copilot in Power Platform is generally available, with capabilities that include natural language search, instant visualization, and AI-generated summaries across model-driven apps, with additional AI-assisted build paths continuing to expand. According to Microsoft's March 2026 Power Platform feature update, Copilot capabilities are enabled at the environment level, making natural language search, instant visualization, and AI-generated summaries available across every model-driven app. Salesforce Agentforce, ServiceNow Now Assist, and Dynamics 365 Copilot follow the same pattern, each embedding AI capabilities into the platform environments where governance and compliance controls already operate. The point applies equally to whatever governed platform the organization has standardized on. The platforms most likely to be displaced are the ones that chose not to integrate AI. Those that did are now offering an AI-assisted build path that does not carry a custom ownership tail.

What the Convergence Model Looks Like in Practice

The optimal future state is AI-assisted prototyping to collapse the feedback loop with working software in front of real users within days, AI-accelerated specification that turns stakeholder sessions into testable acceptance criteria and data models, then implementation on the platform the organization already operates and can sustain.

The before-and-after view of a typical workflow modernization makes the compression concrete, based on the delivery patterns we observe across our engagements. Before: requirements gathering takes four to six weeks, wireframes and approval another three to four, development eight to twelve, UAT and deployment four to six, total roughly six to eight months. After: a working prototype in front of users in week one, a vetted specification from stakeholder sessions in week two, platform-native build in four to eight weeks, UAT and deployment in two to three weeks, total roughly two to three months. Platform-native components, including Power Apps canvas and model-driven apps, Power Automate flows, Dataverse schemas, and connectors, are generated and refined rather than built from scratch. The speed advantage is real, but it shows up most cleanly at the platform end, where AI compresses delivery without adding to what must be maintained afterward.

A second convergence model is worth noting briefly: an AI-built experience and workflow layer on top of platform-provided backend components that handle data integrity, security, and compliance. The platform provides the governed foundation; the custom layer handles what the platform cannot express. Fully custom build does win on a full TCO view in specific cases, including differentiated capability no platform can express, integration requirements outside any platform's connector surface, and situations where per-user platform licensing is itself the cost problem at scale. These cases are real. They are not the majority of what we see proposed.

Rethinking the Ownership Math When Build Costs Change

AI-assisted development meaningfully lowers the initial build cost. A system that previously required $200,000 in development labor might now be assembled for considerably less. That is a genuine shift, but it introduces a question worth examining: if AI lowers the cost to build, what does that do to the ongoing cost of ownership?

The long-standing industry benchmark places annual maintenance at roughly 15 to 25 percent of the original development budget. When AI lowers the initial build cost, the absolute dollar figure of that percentage drops in raw terms as well. The complicating factor is that the percentage itself may not hold. AI-generated codebases can carry patterns optimized for speed of generation rather than long-term readability, along with dependency choices made without full context of the organization's upgrade cadence. Whether those characteristics raise the effective maintenance rate above historical benchmarks is an open empirical question that the systems built in 2025 and 2026 will begin to answer.

For AI-developed code built on an established platform, the platform absorbs security patching, compliance posture, infrastructure, and upgrade cycles, narrowing the ownership obligation to the application layer. For code built from the ground up, the full ownership obligation applies to the entire codebase, and token and compute run cost adds a variable that traditional maintenance models did not include. The McKinsey survey's finding that a significant share of respondents are already reporting AI operating cost constraints confirms this variable is not theoretical. The value of any ownership model is in forcing the question before the decision is made: what is the projected annual cost, who owns it, and does the organization have the capacity to sustain it over the system's expected life?

A Decision Method That Works Before the Evidence Is In

Score each candidate initiative across seven dimensions: (1) whether the capability creates competitive or mission advantage no available platform can express; (2) what compliance and accessibility obligations apply and who carries them if built custom; (3) whether the system will need to be maintained for five or more years and by whom; (4) how many systems it must connect to and whether the platform's connector coverage can reach them; (5) what the organization already owns and what share of the stated requirement it satisfies today; (6) whether the team has the skills and bandwidth to own the system through its full lifecycle; and (7) what the estimated annual ownership cost is, including inference costs, maintenance labor, and sustainment overhead, across both platform-native and fully custom scenarios.

Beyond the scoring, one gate is non-negotiable: a named sustainment owner must be identified before any build proceeds. If no one can be named to own the system after delivery, it should not be built. That single requirement is the most reliable leading indicator of whether a project will deliver lasting value or become the next system that costs more to maintain than it cost to build. KPMG's advisory framework reaches a similar conclusion: the governance and ownership question must precede the architectural one.

We will revisit this question when TCO data from early builders becomes available. The McKinsey survey has confirmed that the decision is being made at scale. The evidence on whether it was the right decision is still forming, and the organizations making it now are the ones who will generate that evidence. Until it arrives, the best available use of these tools is to deliver faster on platforms the organization already operates, trusts, and can sustain. That is where the speed shows up without the ownership bill attached.

Sources

1. The state of AI in 2026: On the road to ROI — McKinsey (2026)
2. The Build-vs-Buy Shift: 32% of Enterprises Bet on Agentic Coding Tools — Forkast News via Yahoo Finance (2026)
3. A Third of Companies Skipped Buying Software and Built It — Digital Applied (2026)
4. Gartner Says $234 Billion in Enterprise Application Software Spend Is at Risk from Agentic AI — Gartner (2026)
5. AI Token Costs: Why Enterprise AI Bills Keep Rising in 2026 — Optimum Partners (2026)
6. Government Website Accessibility: Complete Section 508 & ADA Title II Compliance Guide — AllAccessible
7. Technical Debt Management: The Road Ahead for Successful Software Delivery — arXiv (2024)
8. Cost of Technical Debt - SIG — Software Improvement Group
9. Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 — Gartner (2025)
10. Enterprise AI coding agent deployment in 2026 — Northflank (2026)
11. Agentic AI untangled: Navigating the build, buy, or borrow decision — KPMG (2026)
12. Cloud adoption gains traction with state, local agencies — StateScoop
13. The Great SaaS Consolidation 2026 - United States — Insentra Group (2026)
14. 2026 STATE CIO TOP 10 PRIORITIES - NASCIO — NASCIO (2025)
15. What's new in Power Platform: March 2026 feature update — Microsoft (2026)
16. Software Maintenance Costs 2026: Complete Pricing Guide — Adevs (2026)