AI in legal and software development: productivity and risk benchmarks (with sources)

In knowledge work, AI is both the most adopted and the most ambiguous. The same Copilot that speeds you up can slow you down; the same legal AI that summarizes one case can invent another. Here, the context detail flips the sign of the result.

Among developers, AI use rose from 76% to 84% in 2025, but trust in accuracy fell from 40% to 29%: they use it more and trust it less (Stack Overflow). In legal, active generative-AI use jumped from 14% to 26% in a year, and more than 95% of professionals expect it to be central to their work within five years (Thomson Reuters).

Where AI already delivers

In BigLaw, Allen & Overy was the first to deploy generative AI firmwide: about 3,500 lawyers across 43 offices ran roughly 40,000 queries to Harvey during the beta, across 250 practice areas. Those are the primary-sourced figures; the specific time savings that circulate come only from marketing material and are left out.

Where AI broke (and the contrast that matters)

Copilot's +55% came from a simple task done by beginners. In a controlled METR study (2025), the result flipped: 16 experienced developers, in codebases they knew well, were 19% slower with AI, despite believing they had sped up by about 20%. Same technology, opposite sign. Context decides.

• Mata v. Avianca (2023): lawyers sanctioned $5,000 for filing fake cases invented by ChatGPT, the reference case for every hallucination ruling since.
• Legal hallucination: even purpose-built tools still err, about 17% for Lexis+, 33% for Westlaw and 43% for GPT-4, per Stanford HAI.
• Scale of the problem: an independent tracker already counted more than a thousand US decisions with AI-fabricated citations in 2026.
• Deskilling: higher confidence in AI correlates with less critical thinking (Microsoft/CMU, CHI 2025), and AI-generated code tends to be less secure while the developer feels more confident it is safe (Stanford).

In Brazil: the bar and the courts already reacted

The Brazilian bar (OAB) approved recommendations for AI use in legal practice in 2024: the lawyer must disclose AI use to the client and partners must supervise associates and interns. And courts already fine fake citations, with cases at the TJSC (10% of the case value), the TST (1%), a federal court in Londrina (20 minimum wages) and the TJPR with a seven-figure fine. The principle is always the same: responsibility for verifying information rests with the attorney.

The lesson for anyone implementing

The failure pattern repeats: more confidence precisely when there should be more scrutiny. The rule that works is "assistant, not authority": every AI output, whether a citation, a number or code, is a draft until verified. Gartner projects legal-tech budgets will double by 2028, but also that more than 40% of agentic-AI projects will be canceled by 2027. That is why Reche ships development with quality gates that treat AI-generated code as a draft to validate, never as ready-made truth.

1. 1.Stack Overflow — 2025 Developer Survey (AI)
2. 2.Thomson Reuters — Future of Professionals / legal AI adoption
3. 3.Artificial Lawyer — LawGeex 94% vs 85% on NDA review
4. 4.Peng et al. — Impact of GitHub Copilot on productivity (arXiv)
5. 5.Thomson Reuters — AI to save professionals 12 hours per week by 2029
6. 6.Chien & Kim — Generative AI and Legal Aid (LMU field study)
7. 7.Google/DORA — Accelerate State of DevOps 2024
8. 8.A&O Shearman — firmwide Harvey deployment
9. 9.METR — Early-2025 AI and experienced developer productivity
10. 10.Mata v. Avianca — Seyfarth analysis
11. 11.Stanford HAI — legal models hallucinate in 1 of 6+ queries
12. 12.Damien Charlotin — AI hallucination cases database
13. 13.Lee et al. — Impact of generative AI on critical thinking (CHI 2025)
14. 14.OAB — recommendations for AI use in legal practice
15. 15.Conjur — Brazilian courts fine AI-fabricated case law