Where AI Cuts Software Delivery Costs by 50% — and Where It Doesn't

AI reliably cuts software delivery costs by 40-60% in three phases of the life cycle — writing specifications, generating new code, and producing tests — but it delivers far less in code review, and it can add cost in operations and maintenance if quality is left unmanaged. Across a full project, a blended saving of roughly 35-50% is realistic in 2026, and it depends less on the tools than on senior engineers directing them. That is the honest picture that emerges from Google's 2025 DORA report, Stanford's 2026 AI Index, and METR's randomised controlled trial of AI-assisted developers.

The market is full of two bad numbers. Vendors quote "10x productivity" as if it applied uniformly to everything a delivery team does; sceptics quote a single contrarian study as if AI made everyone slower. Both are wrong in the same way: they treat software delivery as one activity. It is at least five — specification, build, test, review, and operations — and AI's economics are dramatically different in each. We build production software with AI systems every day, so we have a commercial interest in the 50% number being credible rather than inflated. Here is where it actually comes from, phase by phase.

Where does the "50% cheaper" number come from?

No single study says AI halves delivery cost across the board, and anyone claiming otherwise is selling something. What the data shows is compounding. Stanford's 2026 AI Index, released this month, documents measured productivity gains of roughly 26% in software development tasks — larger than customer support, smaller than marketing — with significant variance by task type and experience level. Google's 2025 DORA report, published in September 2025 from a survey of nearly 5,000 technology professionals, found 90% now use AI at work, more than 80% report a productivity increase, and — reversing 2024's finding — AI adoption now correlates with higher software delivery throughput. The 50% figure emerges when phase-level savings compound and when team shape changes: smaller teams of senior engineers directing AI systems, rather than the same headcount typing faster.

Specifications: a quiet 30-40% saving nobody markets

The writing gets cheap; the deciding does not: AI compresses the mechanical half of the specification phase — drafting PRDs, expanding user stories, enumerating edge cases, generating API contracts and acceptance criteria — from weeks of documentation effort into days. McKinsey's analysis of the AI-enabled product development life cycle argues this is precisely where organisations leave money on the table by confining AI to coding: gains compound when it spans discovery through delivery, freeing product managers and engineers for higher-value work. But the expensive part of specification was never typing. Stakeholder alignment, domain judgment, and deciding what not to build are untouched, which is why we score this phase at 30-40% rather than higher. A beautifully drafted specification for the wrong product is now cheaper to produce than ever.

Build: the largest saving on paper — and the most conditional

Code generation is where AI capability has moved fastest. Stanford's 2026 AI Index records dramatic year-on-year jumps on the SWE-bench Verified benchmark, and agentic tools now complete multi-file changes that would have been science fiction in 2024. On greenfield work — new services, well-scoped features, standard architectures — a 40-60% reduction in build cost is achievable today, and DORA's 2025 finding that AI adoption is now linked to higher delivery throughput supports treating that as mainstream rather than exceptional.

The METR caveat every buyer should know: in July 2025, METR published a randomised controlled trial in which 16 experienced open-source maintainers completed 246 real tasks on large, mature repositories they knew intimately. With AI tools, they were 19% slower — while believing they had been 20% faster. Two lessons follow. First, on dense legacy codebases where the engineer's tacit context exceeds the model's, savings shrink and can invert; brownfield build savings are realistically 0-20%, not 50%. Second, self-reported productivity is untrustworthy in both directions — if your delivery partner cannot show measured cycle-time data, they are quoting perception, not economics.

Testing: the most under-priced win in the SDLC

Test authoring is repetitive, pattern-heavy, and verifiable — exactly the profile AI handles best. Generating unit tests, expanding regression suites, and scaffolding integration harnesses routinely cuts 40-50% from QA effort, and teams that historically skipped tests under deadline pressure now have no excuse. The caveat is the oracle problem: an AI writing tests against existing code tends to assert what the code does, not what it should do. A generated suite can score high coverage while enshrining bugs as expected behaviour. The discipline that protects the saving is human-defined acceptance criteria first, generated tests second — which is also why testing savings flow to teams with strong engineering practice and evaporate for teams without it.

Why doesn't code review get cheaper?

Here the honest answer is uncomfortable: review costs barely move, and on some teams they rise. DORA's 2025 report observes that time saved in code creation is frequently re-allocated to auditing and verification — AI accelerates the writing, and the acceleration exposes weaknesses downstream. More code arriving faster makes review the new bottleneck. AI-assisted pre-review genuinely helps — catching obvious defects, summarising diffs, flagging risky changes — worth perhaps 10-20%. But accountable human judgment does not scale down with tooling, and in regulated sectors it legally cannot: defence accreditation regimes and healthcare clinical-safety processes require named humans reviewing and signing off changes, with audit trails to match. Any cost model that halves review effort in those sectors is fiction.

Operations and maintenance: where the bill arrives

Unmanaged AI code is a loan, not a discount: GitClear's 2025 research, analysing 211 million changed lines of code, found duplicated code blocks growing several-fold during 2024, copy-pasted lines exceeding refactored ("moved") lines for the first time in its dataset, and 7.9% of newly added code revised within two weeks — up from 5.5% in 2020. DORA's 2025 report reinforces the pattern: AI adoption continues to show a negative relationship with delivery stability. AI does help operations — incident summarisation, runbook drafting, log triage and observability queries are real, modest savings. But maintenance economics are set upstream, by the quality of what was merged. Teams that ship unreviewed generated code at volume are not cutting cost; they are deferring it at interest.

Which delivery costs does AI not touch at all?

A fully burdened delivery budget contains large line items that AI barely dents in 2026: compliance and accreditation cycles, procurement and contracting, environment access in air-gapped defence settings, data migration off legacy estates, and the organisational latency of getting decisions made. Researchers call the capability pattern a "jagged frontier" — excellent at structured tasks, weak where judgment dominates — and these are the judgment-dominated residue of every project. There is also a workforce cost accruing quietly: the same Index reports employment for software developers aged 22-25 down nearly 20% since 2024. Organisations that stop hiring juniors are saving payroll today and liquidating their pipeline of future senior reviewers — the one role this entire analysis shows becoming more critical, not less.

What does a realistic 2026 cost model look like?

Put the phases together and the honest arithmetic reads: specifications 30-40% cheaper; build 40-60% on greenfield but 0-20% on complex brownfield; testing 40-50%; review 10-20%; operations roughly flat, and negative where quality is unmanaged. Weight those by a typical project's effort distribution and a well-run team lands between 35% and 50% blended savings — nearer 50% for greenfield delivery under senior direction, nearer 20-30% for legacy-heavy regulated work. DORA's core conclusion explains the spread: the value of AI is unlocked not by the tools but by surrounding technical practices and culture — small batches, strong version control, fast quality feedback, and a clear organisational stance on AI use. The teams capturing 50% treat AI output as an untrusted contribution from a very fast junior: reviewed, tested, and directed by engineers who could have written it themselves. That is why "up to 50% lower delivery cost" is a defensible claim in April 2026 — and why any partner quoting a flat 50% discount on everything, including your compliance burden and your legacy integration, has not read the data. Ask them which phase the saving comes from. The good ones can tell you, line by line.