Build Instead of Buy Under Agentic Coding

Sources#

• 2026 State of Scaling: The Great Sorting
• The state of AI in 2026: On the road to ROI

Summary#

The state of AI in 2026: On the road to ROI (McKinsey / QuantumBlack, 2026-08-25, online survey of 1,719 participants in 97 nations fielded May 4 – June 8 2026, GDP-weighted, empirical but wholly self-reported) reports a procurement effect that no other source in this corpus measures at population scale:

"Nearly one-third of respondents (32 percent) report that their organizations have decided against purchasing at least one software product or feature because they were able to build the functionality in-house using agentic coding tools."

This is the buy-side counterpart of the supply-side thesis Implementation Abundance Inverts Product Work argues from inside a frontier lab: if standing up a feature is cheap, the make-or-buy line moves. Here it has moved for about a third of a global respondent panel, in a single year, without anyone announcing it.

The distribution, and what it is doing (Exhibit 4)#

Industry	% deciding against a purchase	n
Technology	41	199
Healthcare payers and providers	39	47
Professional services	38	205
Energy and materials	38	153
Financial institutions	36	148
Media and telecom	34	71
Pharma and medical products	33	55
Travel and logistics	31	56
Consumer goods and retail	29	116
Engineering and construction	28	72
Advanced manufacturing	27	95
Insurance	19	40
Public and social sector	17	71
Total	32	1,719

The spread is 17 to 41 points — narrower than the adoption spreads elsewhere in the same survey, which is the first thing worth noticing. This is not a technology-sector story with a long tail; energy and materials (38) and professional services (38) sit within three points of technology, and eleven of thirteen industries clear 27%. The two laggards are the two most procurement-constrained sectors in the panel (insurance, public and social sector), which is consistent with the decision being gated by permission to build rather than by ability.

The high-performer split is the largest in the exhibit set: nearly half of AI high performers (the 6% of respondents attributing ≥5% of EBIT to AI and calling its value "significant") report forgoing a purchase, against 31% of everyone else. That gap is small relative to the same cohort's gaps on workflow redesign (~72% vs ~25%, Exhibit 11, approximate) — build-instead-of-buy is one of the least concentrated high-performer behaviors in the survey, i.e. it has already diffused past the leading edge.

The precondition: who can actually do this (Exhibit 3)#

The decision requires a coding-agent capability, and that capability is size-gated:

Software coding agents	≥$1B revenue	<$1B revenue
At least scaling	31	17
Piloting	16	15
Experimenting	21	21
Not at all	21	40
Don't know	11	7

About two in ten organizations overall report scaling software coding agents. So roughly a third of organizations report forgoing a purchase while roughly a fifth report scaling the tool that makes it possible — the decision is running slightly ahead of the scaled capability, which fits a world where one team's coding agents are enough to kill one renewal. Among AI high performers the scaling rate doubles (40% vs 20%, Exhibit 13).

Note the 11% "don't know" rate on coding agents at large enterprises — the highest of any tool in Exhibit 3 except other AI technologies. The panel is C-level-heavy (654 of 1,490 role-identified respondents are C-level; 195 are individual contributors), and executives are measurably less certain about coding agents than about chatbots. Treat the coding-agent rows as the survey's softest, not its hardest.

The bound: the build option is not free (Exhibits 8 and 14)#

About 20% of respondents report that AI-related operating costs, including token costs, have constrained their organization's AI use — broadly flat across company sizes and spread 12% (consumer goods and retail, public sector) to 25% (technology) across industries. Tool-by-tool, roughly one in ten report constraint on each of chatbots, agents and coding agents.

The cut that matters here is Exhibit 14: AI high performers report cost-constrained use of software coding agents at 18%, against 6% of all others — a 3× gap, and the only tool where high performers report more constraint than everyone else (chatbots 10 vs 9, other agents 8 vs 9, specialized tools 6 vs 6). The orgs that build instead of buying are the orgs hitting the token bill. Michael Chui states the mechanism in the article's own commentary:

"They've discovered that AI isn't 'too cheap to meter' when it comes to agentic software development and complex reasoning tasks that benefit most from frontier models: Even as per-token costs have declined, the number of tokens consumed and generated has increased even faster."

That is the same shape Cost-per-Task Over Cost-per-Token argues about at the task level, arriving at the budget level: a build decision made because implementation got cheap is settled in a currency whose total is going up.

What is actually being reallocated#

Chui names the budget line: "AI influences other parts of the IT budget, too: Agentic coding tools are creating a more capable option for moving some software development in-house, rather than buying packaged software." Two surrounding figures set the scale — 28% of respondents say their organizations spend more than 10% of enterprise-wide ICT budget on AI, and 60% expect AI investment to rise in the coming year (Exhibit 9).

Lieven Van der Veken frames the same finding as a posture change rather than a spend change:

"Until recently, many leaders assumed that AI was beyond the capabilities of their own technology teams… Now, the tone is changing. Leaders are asking what their organizations need to build AI tools themselves. The rise of software coding agents and in-house development is one clear sign of this broader shift."

with the explicit caveat that this "does not mean companies should build everything themselves" — the claim is that buyers now negotiate "from a position of agency," which is a consulting frame, not a measurement. Weight accordingly: McKinsey sells the AI transformation work that the posture change implies.

Not the same decision as forking a harness#

Harness Build-vs-Buy prices the other build-versus-buy decision in this corpus — whether to own the coding agent itself, where the measured cost of ownership is ~4,600 merged PRs (≈13/day) of upstream drift per year per forked harness. The two decisions point opposite ways and compose badly:

• Harness layer: buying (or customizing at the highest layer that works) is the argued default, because the artifact churns faster than any fork can track.
• Application layer: building is the newly available option, because the harness you bought makes application code cheap.

An organization can rationally rent its coding agent and build the software that agent writes — and the McKinsey figure says a third of them are doing exactly the second half. Nothing in either source measures the composition: no one has asked whether the orgs forgoing SaaS purchases are the ones running vendor harnesses or the ones maintaining forks.

What the instrument cannot support#

The question asked was whether the organization decided against buying at least one product or feature. Everything downstream of the decision is unmeasured:

• No dollar value. "One or more products or features" includes a killed $2M ERP module and a declined $40/month add-on identically. Nothing separates a budget shift from a procurement anecdote.
• No completion. The survey records a decision at decision time. Whether the in-house build shipped, matched the product's scope, or was still running a year later is outside the instrument entirely — which is the specific way this figure could be an artifact of Pilot-to-Production Gap's mechanism rather than a counterexample to it.
• No maintenance horizon. A purchased product's cost is a renewal; a built product's cost is a permanent team. The survey prices neither.
• No counterfactual. Some of these purchases would not have happened anyway. Nothing controls for budget pressure, which was independently tight in the same fielding window.
• Self-report, one respondent per organization, GDP-weighted, from McKinsey's own panel, in the survey where the same population reported enterprise-level EBIT impact essentially flat at 37%. Respondents who report an unrealized headline elsewhere in the same instrument reported this one too.

The same thesis, priced by the public market (ICONIQ, September 2026)#

Every reading above is a buyer-survey or a practitioner account. ICONIQ's 2026 State of Scaling (ICONIQ Venture & Growth, September 2026, empirical) supplies the vendor-side consequence as a valuation fact, and states the causal claim without hedging: "The rise of frontier AI models is making in-house application development more accessible, putting pressure on NTM revenue multiples across all software sub-sectors" (p.13).

The surrounding numbers: the S&P software index fell 3% LTM, trailing every other sector while semiconductors gained 47%; within software, NTM revenue multiples split to 9.2x for high-growth companies (20%+ growers) against 2.7x for the rest in Q2 2026. Infrastructure and security are named as the resilient sub-sectors, on the reasoning that companies deploying AI must first buy the infrastructure, data systems, governance and security to do it reliably — which is build-instead-of-buy at the application layer financing buy-not-build one layer down.

How much weight this carries. The multiple data is real and public; the attribution is ICONIQ's editorial judgment, offered with no decomposition separating in-housing pressure from rate environment, growth deceleration, or AI-native competitive entry — all of which compress software multiples and none of which is build-instead-of-buy. What it establishes is narrower than it sounds and still worth having: a major growth investor now names in-house development as a first-order explanation for software multiple compression, which is this page's thesis being priced rather than surveyed. Treat it as the market's expectation of the thesis, not as its confirmation.

Connections#

• Implementation Abundance Inverts Product Work — the supply-side thesis this is the demand-side reading of: implementation abundance reaching the purchase order. Its "does the inversion survive outside a frontier lab with unlimited tokens?" question is partly answered here, and partly answered against it
• Harness Build-vs-Buy — the same verb one layer down, pointing the other way; see the composition gap above
• Standardize the Infrastructure, Not the Tools — the cost-visibility layer that makes a token bill a budget line an executive can act on; the 20% cost-constrained figure is what that meter reads at population scale
• Cost-per-Task Over Cost-per-Token — what the build decision is denominated in once it is made, and Chui's tokens-grow-faster-than-price-falls observation
• The AI-Native Safe-Choice Inversion — the vendor-side threat: the buyer's "safe" option inverted once, toward AI-native challengers; this is the third option, where the buyer stops buying
• Forward-Deployed Engineering as a Delivery Layer — the same demand resolved the other way. An enterprise declining a purchase because it can build the thing, and an enterprise paying a vendor to embed engineers who build the thing for it, both want a fitted system rather than a product; which resolution it picks turns on whether it believes it has the capability in-house, and no source measures the split
• Pilot-to-Production Gap — the reason to hold the figure loosely: a decision not to buy is exactly the kind of commitment a pilot's success conditions can produce and production cannot honor
• Seven Powers Applied to AI — the same procurement shift one level of severity apart: 32% declining a purchase outright is the terminal case of the switching-cost erosion that page tracks, whose milder and far more common face is Madrona's 77% of enterprises re-evaluating their AI vendors every six months. Note the two figures are not additive — different instruments, different populations, and one is about software the buyer could build while the other is about AI vendors it keeps rebuying

Open Questions#

• Does the forgone purchase stay forgone? No source records whether an in-house replacement survives a maintenance cycle, or whether the vendor is re-engaged at renewal + 24 months. The decision is captured at decision time only, and the re-buy is the falsifying event. A follow-up wave asking the same respondents about the specific product they declined would settle it.
• Is the 32% moving real money? The instrument has no dollar value and no size floor, so the figure is compatible with a large budget reallocation and with thirteen declined add-ons. A payment-rail instrument applied to software-vendor spend — the Ramp aperture pointed at SaaS line items rather than AI vendors — would separate the two, since a genuine displacement shows up as decelerating net-new software vendor spend in the industries reporting the highest rates.
• Does the build option get rationed by tokens? High performers report build-instead-of-buy at ~48% and cost-constrained coding-agent use at 18% (3× everyone else), but the survey publishes no cross-tab, so it cannot say whether the same organizations hold both. If they do, the procurement effect has a price ceiling and will plateau rather than compound.

Sources#

• The state of AI in 2026: On the road to ROI — Dan Tinkoff, Lieven Van der Veken & Michael Chui with Tara Balakrishnan, The state of AI in 2026: On the road to ROI (McKinsey / QuantumBlack, 2026-08-25, empirical, self-reported): the key-takeaways list and §"Use of AI is deepening" for the 32% headline; Exhibit 4 (industry breakdown with per-industry n); Exhibit 3 (coding-agent phase by revenue band); Exhibits 13 and 14 (high-performer scaling and cost constraint by tool); Exhibit 8 (cost-constrained use by industry); Exhibit 9 (investment expectation); Van der Veken's and Chui's commentary blocks; §"About the research" for the panel. All 18 exhibits are rasterized SVG charts transcribed at ingest — Exhibit 11 carries no printed labels and its values are read off gridlines (approximate, flagged wherever cited). COI: McKinsey sells the AI transformation consulting this finding implies demand for, and the "position of agency" framing is the consulting product. Full evidence handling at Source Notes
• 2026 State of Scaling: The Great Sorting — ICONIQ Venture & Growth, 2026 State of Scaling: The Great Sorting (September 2026, empirical; 52-page PDF, docling-parsed). Cited here only for the public-market section (p.11-14): the -3% LTM software index against semiconductors' +47%, the 9.2x-vs-2.7x NTM revenue-multiple split, and ICONIQ's attribution of multiple compression to frontier models making in-house application development more accessible. The market data is public; the causal attribution is the publisher's editorial judgment with no decomposition. Publisher COI at ICONIQ