Somebody told your board that AI makes engineers 40% faster, and now it wants a release date. Which 40%, though? In AI fintech software development, backlog work and the coding itself do shrink, and so does the reading that comes before every integration. Other stages won't budge, because a PCI assessor or a payment provider sets the pace.
TL;DR
AI shortens backlog work, context gathering, and coding,
PCI DSS assessments and PSP certification don't get any shorter,
Lead time shows gains that commit counts hide,
Payments logic still needs a human reviewer before it merges.
Why do "X% faster" claims fall apart between the IDE and the release?
Because a number from a single task gets quoted as if it described a release. The classic case is GitHub's 2023 Copilot experiment (Peng et al.). Developers asked to build an HTTP server in JavaScript finished 55.8% faster with AI. Nobody had to get that server through review.
METR's 2025 randomized trial makes the gap uncomfortable. Its 16 experienced open-source developers worked through 246 real issues and took 19% longer with AI. Asked afterward, they estimated AI had made them 20% faster. A 2026 follow-up leans the other way, on weak evidence by METR's own account.
DORA sees it at team level. Its 2024 report linked more AI adoption to 1.5% lower delivery throughput and 7.2% lower stability. Your lead time and change failure rate sit in those two numbers.
A year later, 90% of respondents were using AI and more than 80% felt more productive, yet only throughput had recovered. Spendesk's CTO, Alan Wright, sees the same split. AI-assisted PRs went from 16% to 70% of all PRs, but "what reaches the customer hasn't sped up to match yet."
Trust lags behind use. In Stack Overflow's 2025 survey, 46% of developers said they distrust the accuracy of AI output, and only 3% highly trust it. Their top frustration, cited by 66%, was answers that are "almost right, but not quite."
Finance is catching up fast. Cyberhaven's telemetry from financial services firms shows developer use of AI coding assistants rising from 16.7% to 42.1% over 2025. In Wolters Kluwer's Q1 2026 survey of 148 institutions, though, just 26.4% felt confident their AI initiatives would meet regulatory requirements. Explainability was the concern they cited most.
Which stages of the AI software development lifecycle actually get faster?
The ones your engineers control, code review included. The gain depends on how many manual checks you keep. Agentic Collections, an open-source AI delivery framework used by 300+ engineers on 50+ commercial projects, shows how big the gains get. It runs four phases (Research, Plan, Implement, Review), each needing a human sign-off.
Discovery and backlog
Hand the agents your workshop notes and the PSP's documentation, and tickets and specs come back. In Agentic Collections, that cut backlog work from 1-2 days to 15 minutes. Each manual pass over generated tickets claws some of that back, so spend it on sprint selection.
Context gathering
Much of a PSP integration is reading the legacy module and the provider's API reference, then checking which compliance rules apply. Agentic Collections gives that reading to Business Analyst and Context Engineer agents. The framework's fintech delivery results put context gathering at 3 minutes, down from 60.
Coding
Coding is where task-level numbers like GitHub's 55.8% come from. Assistants feel fastest on integration glue, tests, and boilerplate. The catch shows up in security review. When Veracode gave 80 coding tasks to more than 100 models in 2025, 45% of test cases came back with OWASP Top 10 vulnerabilities.
Code review and QA
AI review takes the first pass. Agentic Collections runs its own checks on every PR (acceptance criteria, security, performance), but nothing merges until an engineer clicks approve. Each manual check buys safety and costs time, and payments logic is worth the cost.
Black Duck's survey of 831 engineers (March 2026) found bottlenecks forming around AI-generated code, led by manual review (52%) and security testing (51%). At Spendesk, Wright says code is "being generated at a volume that's outpacing the team's ability to review it accurately." His next step is AI review of AI-written code.
Where does AI stop speeding up fintech delivery?
At compliance review, a separate gate from code review. PCI assessors, auditors, and payment providers set its pace.
Compliance and release
Assessors check, among other things, that custom code was reviewed before release (PCI DSS 6.2.3). Where that review was manual, 6.2.3.1 adds a second reviewer and management sign-off. AI makes a decent clerk, drafting change logs and pulling test reports into the evidence package. The reviewers take as long as ever, and assuming otherwise is one of the risks of AI coding tools in regulated fintech.
How do you measure software developer performance when AI writes part of the code?
If you're working out how to measure software developer performance with AI in the loop, start before rollout with a baseline on your own team. Then track three sets of numbers:
Delivery (DORA). Lead time for changes, deployment frequency, change failure rate, and failed deployment recovery time.
Productivity (SPACE). The SPACE framework for developer productivity (Forsgren et al., 2021) looks at satisfaction, performance, activity, communication, and efficiency.
Experience (DevEx surveys). Ask how your developers experience feedback loops, cognitive load, and focus time (Noda et al., 2023).
Stage by stage, those numbers become these signals:
Stage | Green flag | Red flag |
|---|---|---|
Discovery and backlog | Lead time from ticket to release drops | Tickets closed as the success metric |
Context gathering | Less time lost to context switching | Unchecked AI summaries of compliance rules |
Coding | Cycle time drops and fewer defects reach production | Faster coding, same release date |
Code review and QA | Review queues stay flat as PR volume grows | More PRs, longer reviews, unreviewed payments logic |
Compliance and release | Evidence ready before the assessor asks | Board date assumes faster approvals |
Leave out volume metrics such as lines of code, PR count, tickets closed, and accepted AI suggestions. They all rise after a rollout whether or not anything ships sooner. METR's developers fell for the same illusion. Agentic Collections measures against a baseline too: average lead time is down by up to 40% compared with the same teams before adoption. When someone promises your board 40%, ask for that comparison.
What delivery date can you promise the board?
Build it from two estimates. AI can bring engineering time down. Approval time gets no AI discount, so plan certification and audit at today's length. In custom fintech software development, speed over a multi-year roadmap also depends on engineers who keep growing with the product.
For xpate, that has meant working with the same engineering partner since 2022. Its CTO, Ruben Vosmeer, says that partner has "invested in developing their engineers and bringing them to a new level, adjusting to the market's needs."
Key insights
A faster task and a faster release are different measurements,
Reviewer capacity, not coding speed, sets the delivery ceiling,
Finance adopts AI faster than it gains regulatory confidence.
Fintech teams that hit their dates told the board two things: how much faster engineering got, and how long approvals still take.
Authors
Jakub Matuszak
Marketing Specialist at The Software House, focused on B2B tech insights and turning complex topics into actionable guidance for engineering leaders.
Andrzej Wysoczański
Frontend developer with 10 years of experience. With The Software House for almost 7 years, going from a regular dev to the Head of Frontend. He loves keeping tabs on the latest frontend technologies, especially React-related. Regular of the Taby & Spacje podcast (tsh.io/taby-vs-spacje) for Polish speaking programmers.