Can AI Reduce Engineering Errors?

AI can catch the coordination and specification errors that quietly consume a tenth of every project, and just as easily manufacture new ones that look perfectly plausible. Whether errors fall or merely change shape depends on what happens after the model answers.

Updated July 20263 min readEvidence-graded6 sources

On this page

The BFCS verdictEvery article scored on the same four dimensions

Better

PartlyModerate confidence

Catches old error classes, creates new onesAI flags coordination and specification errors well, but injects plausible-looking mistakes of its own. Net error reduction is not automatic; it is a property of the review process.

Note and confidence

Moderate confidenceCatches old error classes, creates new onesAI flags coordination and specification errors well, but injects plausible-looking mistakes of its own. Net error reduction is not automatic; it is a property of the review process.

Faster

YesHigh confidence

First-pass drafts, checks and iteration compressGenerating candidate designs, code and checks is dramatically faster; the time simply moves downstream to verification.

Note and confidence

High confidenceFirst-pass drafts, checks and iteration compressGenerating candidate designs, code and checks is dramatically faster; the time simply moves downstream to verification.

Cheaper

PartlyModerate confidence

Rework is the prize; review is the new costErrors caught in design are far cheaper than errors caught on site or in production, but only if someone competent still checks.

Note and confidence

Moderate confidenceRework is the prize; review is the new costErrors caught in design are far cheaper than errors caught on site or in production, but only if someone competent still checks.

Safer

ConditionalGovernance-gated

Unverified output is dangerous in safety-critical workA confident, wrong answer inserted into structural or safety-critical work is worse than no answer. Licensed accountability does not transfer to a model.

Note and confidence

Governance-gatedUnverified output is dangerous in safety-critical workA confident, wrong answer inserted into structural or safety-critical work is worse than no answer. Licensed accountability does not transfer to a model.

Two error economies

Engineering error shows up in two very different places, and AI touches both. In the physical world, design changes, errors and omissions drive the majority of rework cost; one detailed study of industrial projects put deviations at 12.4% of installed cost, with design accounting for roughly 78% of it.^{Source 1} In software, defects escape into production every day. The common thread: an error caught at the design stage is cheap; the same error caught on site or in a live system is not.

There is genuine precedent for tools reducing error. Since the 1990s, digitisation and BIM have cut the cost impact of design errors from around 9% of project value to 1 to 2%.^{Source 3} AI is the next increment: automated clash detection, specification and consistency checking, and first-pass drafting. Where it reviews, it genuinely helps.

The evidence, in four numbers

~78%

of rework cost from design errors

design changes, errors & omissions across nine industrial projects (CII)

4 to 12%

of project cost lost to rework

CII field-rework average ~5%; UK GIRI puts avoidable error at 10 to 25%

91.5% / 28.7%

AI code valid vs fully correct

GitHub Copilot on the HumanEval task set (Yetistiren et al.)

40%

of AI programs with security flaws

1,689 Copilot-generated programs (Pearce et al.)

A fluent, confident, wrong answer is the most expensive kind of error, and it is exactly the kind AI is best at producing.

Where it falls short

Plausible but wrong. AI output is fluent and confident even when incorrect. On the HumanEval benchmark, Copilot produced valid-looking code 91.5% of the time but only ~28.7% was fully correct.^{Source 4}
New error classes. In one study, 40% of Copilot-generated programs contained security vulnerabilities, a failure mode that barely existed before.^{Source 5}
Verification debt. Time saved drafting is owed back in review. Where that review is skipped, defect and churn rates rise rather than fall.^{Source 6}

Where governance decides the outcome

Error reduction is a verification outcome, not a generation outcome. The discipline that decides it is old and unglamorous: independent review, checking against requirements, and a named human who is accountable for the result. In safety-critical engineering, an AI proposal that enters a structural calculation or a control system unchecked is a liability, not a productivity gain. Treat AI as the proposer and a qualified engineer as the disposer; the licence, and the responsibility, do not transfer to the model.

How to think about adoption

Point AI at the design phase, where errors are cheap to catch and ruinous to miss, and treat it as a tireless first reviewer rather than a final author. Mandate human verification proportional to consequence, instrument your defect-escape rate so you can see whether errors are actually falling, and never let unchecked AI output flow into a safety-critical deliverable. The prize (a tenth of project cost) is real, but only for teams that keep the review discipline they already know they need.

How we read this evidence

BFCS.ai does not fabricate figures. Software and construction evidence are drawn from separate literatures and should not be pooled; each is cited to its source.

Sources

1
Construction Industry Institute, “Costs of Quality Deviations in Design and Construction”:
What we took from it
across nine industrial projects, rework, redesign and repair averaged 12.4% of installed project cost; design changes, errors and omissions were ~78 to 79% of deviation costs.
Back to text
2
CII IR-153 and the UK Get It Right Initiative (GIRI, 2015/16):
What we took from it
direct field rework averages ~5% of project value (90th percentile 12.4%); GIRI estimates UK avoidable error at 10 to 25% (~21% average), roughly £5bn per year.
3
PlanRadar review of rework studies (1991 to 2024):
What we took from it
the cost impact of design-related error fell from ~9% to ~1 to 2% of project cost since the 1990s, attributed to digitisation and BIM, direct evidence that better tools can reduce error.
Back to text
4
Yetistiren et al. (HumanEval, 164 tasks):
What we took from it
GitHub Copilot produced valid code in 91.5% of cases but only ~28.7% was fully correct; Nguyen & Nadi found correctness varied by language (Java 57%, JavaScript 27%).
Back to text
5
Pearce et al.:
What we took from it
in a study of 1,689 GitHub Copilot-generated programs, 40% contained security vulnerabilities.
Back to text
6
GitClear, “AI Copilot Code Quality 2025” (211 million lines analysed):
What we took from it
rising code churn and duplication, and a higher defect rate, correlated with copilot adoption.
Back to text