As AI collapses the cost of writing code in 2026, does the build-vs-outsource calculus actually shift — and what does the verifiable data say a PM should do?

The one-paragraph answer

What got cheap in 2026 is writing code, not shipping software — so the build-vs-outsource decision does not collapse into "AI made building cheap, bring it all in-house." AI adoption among software professionals is now near-universal (90% use it at work, up ~14 points year-on-year, a median of two hours a day; VERIFIED, DORA 2025 via cloud.google.com), and more than 80% believe it lifts their productivity (VERIFIED, DORA 2025). But the DORA research — Google Cloud's own — is blunt that individual speed does not automatically become team delivery: in 2024 a 25% rise in AI adoption was associated with an estimated 1.5% drop in delivery throughput and a 7.2% drop in delivery stability (VERIFIED, DORA 2024, Oct 22 2024), and even in 2025, when throughput finally turned positive, "AI adoption does continue to have a negative relationship with software delivery stability" (VERIFIED, DORA 2025). DORA's headline framing is the load-bearing idea for a build-vs-outsource call: "AI doesn't fix a team; it amplifies what's already there. Strong teams use AI to become even better and more efficient. Struggling teams will find that AI only highlights and intensifies their existing problems" (VERIFIED, DORA 2025). So the marginal-cost story is real but partial — Uber's report that >70% of its pull requests now come from AI agents (SECONDARY; uber.com blocked this session) is what "cheap code" looks like inside a team that already had elite delivery capability, not proof that the capability itself is now free. The honest read for a PM: AI lowers the price of the code, and thereby raises the premium on the delivery capability around it — the version control, small-batch, platform and data practices DORA names — which is exactly the capability a build-vs-outsource decision has always turned on. Ask not "is building cheap now?" but "which side of this contract actually has the seven capabilities that make cheap code turn into shipped software?" — and note that the outsourcing-market numbers that would size this shift (Deloitte, Gartner, Statista) could not be read from a primary source this session and are reported as unverified.

---

Method — what could and could not be read

Primary spine read directly; the live receipt could not be. The DORA findings below were read this session on cloud.google.com — the blog of Google Cloud, which owns and publishes DORA — so they are first-party and graded VERIFIED. The Uber "software factory" numbers are graded SECONDARY: www.uber.com is EGRESS_BLOCKED at this session's proxy, so its engineering blog post could not be opened; the figures come only from search-result extracts and third-party write-ups. The outsourcing-market dimension has no reachable primary source this session — www.deloitte.com, www2.deloitte.com, gartner.com, and survey.stackoverflow.co are all blocked, and Statista-style market-size numbers surface only through vendor/aggregator pages, which this repo does not publish. Blocked hosts are named again under Sources.

Asymmetric confidence, stated up front: the developer-productivity half of this question is strongly sourced (DORA, first-party). The outsourcing-market half is not — every market-structure claim here is SECONDARY at best, and that is the finding's ceiling, not a rounding error.

---

§1 — What actually got cheap: individual code generation

VERIFIED, DORA 2025 (State of AI-Assisted Software Development), read on cloud.google.com — the report's own publisher. The 2025 report is built on "survey responses from nearly 5,000 technology professionals from around the world" plus "over 100 hours of qualitative data," and its adoption numbers are unambiguous:

• 90% of respondents report using AI at work — described as a ~14-point rise year-on-year, with professionals "now dedicating a median of two hours daily to working with AI."
• More than 80% believe it has increased their productivity.
• 30% report little or no trust in the code generated by AI — "a slightly lower percentage than last year" (the 2024 figure was 39%, VERIFIED DORA 2024).

This is the collapse the question is about, and it is genuine: the act of producing a unit of code is faster and more widely automated than at any prior measurement. If the build-vs-outsource decision were only about the cost of typing code, it would now tilt hard toward "just build it."

§2 — What did NOT get cheap: turning code into shipped software

VERIFIED, DORA 2024 and 2025. DORA's whole contribution is that it measures delivery, not output — and delivery did not move the way raw code speed did.

• DORA 2024 (published Oct 22, 2024, cloud.google.com): "As AI adoption increased, it was accompanied by an estimated decrease in delivery throughput by 1.5%, and an estimated reduction in delivery stability by 7.2%." The same report's verdict: "AI has positive impacts on many important individual and organizational factors which foster the conditions for high software delivery performance. But, AI does not appear to be a panacea."
• DORA 2025 (published Sep 23, 2025): the throughput picture improved — "Unlike last year, we observe a positive relationship between AI adoption on both software delivery throughput and product performance." But stability did not: "However, AI adoption does continue to have a negative relationship with software delivery stability."

Read together, the two reports say the same structural thing across a year: AI reliably speeds the writing of code and unreliably — at best — improves the delivery of software, and it still makes the running system less stable. That gap is precisely where a build-vs-outsource decision lives. The cost that fell is the cheap part; the expensive part (review, testing, integration, keeping production stable) did not fall, and by the stability metric got harder.

This matches the studio's existing "AI coding productivity paradox" concept: individual throughput up, system-level delivery flat-to-worse, because the bottleneck moves downstream to review and integration.

§3 — The DORA amplifier thesis is the actual decision rule

VERIFIED verbatim, DORA 2025. The single most useful sentence for a PM making this call:

"AI doesn't fix a team; it amplifies what's already there. Strong teams use AI to become even better and more efficient. Struggling teams will find that AI only highlights and intensifies their existing problems."

The report also notes 90% of organizations have adopted at least one platform (VERIFIED, DORA 2025), reinforcing that the differentiator is no longer access to AI or tooling — nearly everyone has it — but the capability that surrounds it.

For build-vs-outsource this converts a cost question into a capability question. "Should we build in-house now that AI is cheap?" only has a good answer if you know whether your team is the strong team AI amplifies or the struggling team it exposes — and the same test applies to a vendor you would outsource to. Cheap code in the hands of a low-capability team (yours or a vendor's) produces more of the thing DORA measured: unstable delivery, faster.

§4 — The Uber receipt: what "cheap code" looks like inside an elite-capability team

SECONDARY throughout — www.uber.com is EGRESS_BLOCKED this session, so the engineering blog post "Running a Software Factory Efficiently at Uber Scale" could not be opened. The figures below come from search-result extracts and third-party write-ups only, and several aggregators report them consistently; none is a primary read.

Reported figures (SECONDARY), attributed to a post dated August 27, 2026:

• More than 70% of pull requests are attributed to local or cloud agents.
• Engineers have built over 3,600 agent skills, with more than 30,000 agent skill executions per day.
• Weekly active users of the agentic tools grew ~7x between February and August 2026; weekly agent requests grew ~9.4x over the same window.
• Cost per session down ~52% from a June peak; cost per 1,000 model requests down ~34% from a February–July peak.

The temptation is to read "70% of PRs from agents" as proof that building software is now nearly free and the outsource option is dead. The DORA data says read it the opposite way: Uber is the DORA "strong team" — a company with elite delivery infrastructure — and the amplifier thesis predicts exactly this outcome for that kind of team. The 3,600 hand-built skills, the platform, and the cost governance are the capability; the 70% is what the capability produces when you point cheap code at it. It is not transferable evidence that a team without that infrastructure — or a low-cost vendor bought purely on rate — would get the same result. (If uber.com becomes reachable, this section should be re-graded and the exact numbers confirmed against the primary post.)

§5 — The framework a PM can actually apply: the seven DORA AI Capabilities

VERIFIED, cloud.google.com, "From adoption to impact: Putting the DORA AI Capabilities Model to work" (Dec 10, 2025). DORA distilled, from the same 2025 respondents, "seven foundational capabilities that are proven to amplify the positive impact of AI on organizational performance." This is the checklist that answers "which side has the capability":

#	Capability
1	Clear and communicated AI stance
2	Healthy data ecosystems
3	AI-accessible internal data
4	Strong version control practices
5	Working in small batches
6	User-centric focus
7	Quality internal platforms

Notice what is not on the list: "hire cheaper developers," "write more code," "adopt more AI tools." Every item is a delivery-and-context capability — the expensive half from §2. This is the concrete form of "which side has the capability" in a build-vs-outsource decision.

§6 — The outsourcing-market half I could NOT verify (the ceiling)

The question named the "custom software vs outsourcing market," and honesty requires stating plainly: no primary source for the market side was reachable this session. What surfaced, and its grade:

• The direction most cited — that cost is no longer the top stated reason organizations outsource, with talent access and speed-to-capability overtaking it (attributed to Deloitte's Global Outsourcing Survey) — is SECONDARY only; www.deloitte.com and www2.deloitte.com are EGRESS_BLOCKED, so the exact figure and year could not be read from Deloitte directly. Directionally it aligns with the DORA thesis (you outsource for capability, not to save on now-cheap code), but it is not verified here.
• The Gartner projection that by 2030 all IT work will involve AI (some split of augmentation vs autonomous) is SECONDARY; gartner.com was not reachable.
• Global IT-outsourcing and custom-software market-size dollar figures (the various Statista-style totals) appear only through vendor and aggregator estimate pages. Per this repo's rule they are not publishable and are excluded rather than repeated with a hedge.

So the market-structure claim in the one-paragraph answer is deliberately soft. What is solid is the mechanism (AI cheapens code, raises the premium on delivery capability); what is not solid this session is the market accounting of how buyers are responding.

---

How a PM applies this — what to try this week

1. Change the question you ask in a build-vs-outsource meeting. Not "can AI let us build this cheaply in-house?" but "which side of this — our team or the vendor — actually has the seven DORA capabilities?" Run the §5 list against both. Cheap code amplifies whichever side is stronger; if neither is strong, outsourcing the code just relocates the instability.
2. Stop pricing the decision on code volume or developer rate. DORA's data says the code is the part that got cheap; the review/test/integration/stability capability is the part that did not. Price the decision on the expensive half.
3. Use the amplifier thesis as a vendor-selection filter. When evaluating an outsourcing partner, ask for evidence of capabilities 4–7 (version control discipline, small batches, user-centricity, internal platforms) — not headcount or AI-tool logos. A vendor selling "we use AI agents" is selling the cheap half.
4. If you are the in-house team, do the honest self-assessment before you insource on the strength of "AI makes it cheap now." Uber's 70%-of-PRs number is a strong-team result; if your delivery stability is already shaky, adding agent-generated code is the DORA failure mode, faster.
5. Track delivery stability, not PR count, as the AI success metric. It is the one metric DORA says AI still hurts across both years — so it is the early-warning signal that "cheap building" is quietly costing you downstream.

Course relevance

Feeds the DevOps-literacy leg directly (the DORA delivery metrics and the seven-capability model are exactly what a PM needs to read an engineering conversation) and the Offshore/Remote thread (build-vs-outsource is the client/vendor boundary; AI moves the scarce skill from writing code to owning delivery capability). It also gives the cohort a clean, checkable demonstration of the studio's own method: a hot claim ("AI made building cheap, 70% of Uber's PRs are agents") audited against the primary research that reframes it.

Open questions for a future session

• What does the Uber post actually say, read directly, once uber.com is reachable — and does it describe the delivery-stability cost DORA predicts, or only the volume/cost wins?
• What is the primary Deloitte figure and year for "cost is no longer the top reason to outsource," read from Deloitte rather than an aggregator?
• Does DORA 2025's full report quantify the build-vs-buy or in-house-vs-vendor split anywhere, beyond the platform-adoption number?
• Do the seven capabilities correlate differently for in-house vs outsourced teams in DORA's data?

Parked candidates (not researched today — no reachable primary source)

• IT-outsourcing and custom-software market sizing. Every dollar figure found routed through vendor/aggregator estimate pages; excluded per repo rule rather than published with a hedge.
• Stack Overflow Developer Survey 2025 AI section. survey.stackoverflow.co is EGRESS_BLOCKED; would have added a second independent adoption dataset alongside DORA.
• GitHub Octoverse 2024/2025. github.blog is EGRESS_BLOCKED; would have added first-party data on AI-repository growth and developer counts.

---

Cross-References

Related: frameworks/dora-metrics, frameworks/dora-ai-capabilities-model, concepts/ai-amplifier-effect, concepts/ai-coding-productivity-paradox, concepts/build-vs-outsource-under-ai, concepts/verification-debt, concepts/review-as-constraint, concepts/bridge-se-role, people/nicole-forsgren, Research Digests/2026Sep/2026Sep02_BrSE_vs_FDE_Origins_And_Convergence, Research Digests/2026Mar/2026Mar20_DORA_AI_Productivity_Paradox

Sources

Primary (VERIFIED — read directly this session, cloud.google.com, the DORA publisher)

• DORA 2024, Accelerate State of DevOps Report, announced on the Google Cloud blog, dated October 22, 2024 — https://cloud.google.com/blog/products/devops-sre/announcing-the-2024-dora-report. Source of: 25% AI-adoption increase → +7.5% documentation quality / +3.4% code quality / +3.1% code-review speed; −1.5% delivery throughput and −7.2% delivery stability; "AI does not appear to be a panacea"; 39% little-to-no trust in AI-generated code.
• DORA 2025, State of AI-Assisted Software Development, announced on the Google Cloud blog, dated September 23, 2025 — https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report. Source of: ~5,000 respondents + 100+ hours qualitative; 90% use AI at work; >80% report productivity gains; 30% little-to-no trust; 90% of orgs adopted ≥1 platform; the amplifier quote; "positive relationship … on both software delivery throughput and product performance" but "continue to have a negative relationship with software delivery stability."
• DORA AI Capabilities Model — Google Cloud blog, "From adoption to impact: Putting the DORA AI Capabilities Model to work," dated December 10, 2025 — https://cloud.google.com/blog/products/ai-machine-learning/from-adoption-to-impact-putting-the-dora-ai-capabilities-model-to-work. Source of the seven capabilities and the "seven foundational capabilities … proven to amplify the positive impact of AI" framing.

Search extracts (SECONDARY — page not opened this session; primary host blocked; URL recorded as it appeared)

• Uber Engineering, "Running a Software Factory Efficiently at Uber Scale," reported dated August 27, 2026 — primary URL https://www.uber.com/us/en/blog/efficient-software-factory/ (also /gb/en/…), EGRESS_BLOCKED. The >70%-of-PRs-from-agents, 3,600 agent skills, 30,000 executions/day, 7x WAU, 9.4x requests, and cost-reduction figures are from search-result extracts and third-party write-ups only.
• Deloitte Global Outsourcing Survey — "cost is no longer the top stated reason to outsource" direction; Deloitte hosts (www.deloitte.com, www2.deloitte.com) blocked; figure and year unverified.
• Gartner — "by 2030 all IT work will involve AI" projection; gartner.com not reachable.

Not reachable this session (recorded, not guessed): www.uber.com, dora.dev, blog.google, github.blog, survey.stackoverflow.co, services.google.com, www.deloitte.com, www2.deloitte.com, gartner.com, web.archive.org — all EGRESS_BLOCKED / proxy CONNECT tunnel failed, response 403 on direct test. Global IT-outsourcing / custom-software market-size dollar figures were found only on vendor and aggregator estimate pages and are excluded as non-publishable, not repeated with a hedge.