BLOG
Almost every developer who uses AI says it saves time. In Atlassian’s State of Developer Experience 2025, 99% did. The DORA report 2025 tells a different story: AI acts as an amplifier. Without a strong internal platform and a user-centric focus, the individual gains rarely reach team or organisational performance.
That gap was the starting point for one of the afternoon talks at Agile Tour Vienna 2026. Gojko Adzic and Ana Nad, AI Adoption Lead at Raiffeisen Bank International (RBI), had spent the months before working on exactly that question at RBI. Their answer: AI adoption is process improvement. And we have done this before.
Workshop with Gojko Adzic,
30 November 2026 in Vienna
one day on moving agentic work from individual speed to team performance.
Gojko opened with a parallel many in the room recognised. When iterative processes were introduced, the pitch was speed: six-month cycles instead of three years, then one month, then two weeks. But research from that time showed that raw speed only mattered in one situation – high-growth markets with very short product lifetimes. Everywhere else, building a better product or keeping development cost under control mattered more.
„We make it very, very easy for ourselves to move faster, but moving faster is not the same as moving further.„
His own example: two AI workflows of 20 minutes each, run in parallel – followed by a week of merge conflicts. The same mistake as two developers disappearing for six months and then trying to merge.
The metrics repeat as well. Token leaderboards and bragging about spend are the story points of this decade: „Every 20 years we relearn the lesson that measuring effort is idiotic.“ What counts is cycle time and the rework it causes. Twenty minutes that cause a week of rework are not progress.
Individual benefits are cheap: a licence and a credit card. Organisational benefits need decisions – making teams user-centric, building a platform, setting budgets. That is why the two worlds drift apart.
Ana described how RBI started on the individual side: GitHub Copilot, an LLM gateway, Claude for everyone. Teams reported being faster at coding, faster at proofs of concept – and very good at spending tokens.
In May, Ana called Gojko with a plan: build a dark factory. His first question: how dark is dark? In a bank, the answer is clear. There are compliance requirements and checks, and some decisions need a human who is accountable. The project was renamed „software factory„, and the real question came up: what are we actually trying to achieve?
The answers from IT and business were concrete: shorter cycle time, better knowledge retention, more standardisation, faster onboarding of new team members – „both humans and agents“ – and more consistent delivery quality. The productivity gained should free capacity for new business ideas.
When the pilot teams were asked for their cycle time and their bottlenecks, the first reaction was irritation: weren’t you supposed to teach us agentic development? But the bottlenecks were there: code reviews, missing automation, incomplete regression tests. One team wanted agent-driven operations and had no logging or monitoring.
Gojko referred to John Seddon’s work on process improvement: first understand the process, then improve it, and only then bring in automation. Otherwise a bottleneck that was harmless in manual work simply explodes. Seddon distinguishes value demand, the work customers actually asked for, from failure demand, the work caused by what went wrong earlier. „One of the best ways to improve throughput is to eliminate stupid work.„
An insight Gojko called almost philosophical: writing skills, rules and prompts for agents means documenting a team’s real process in more detail than ever before. That makes it visible – and improvable. His recommendation for the last step: once a process is stable, move it out of the LLM into deterministic tools – a lint rule, a faster test run, a workflow orchestrator. They help humans too, and they keep working even if you switch the AI off tomorrow.
This is the core of Gojko’s one-day workshop on 30 November: how to move control and orchestration out of the LLM, keep humans in control and measure whether you deliver better, not just more.
Join the workshop
RBI defined an AI-native product development life cycle for its eleven core markets – deliberately including product discovery, not just delivery. Three principles are non-negotiable: complete upfront specification, dual verification and a human in the loop. Each phase is documented in Markdown with inputs, outputs, evidence and controls, readable by humans and agents. Two gates stay with humans: approving the specification before an agent builds, and validating the evidence before anything goes to production – for example a recorded video of all automated tests.
For the teams, RBI packaged the standard as a „getting started kit„: skills, hooks, tools and agents. Off-the-shelf frameworks had frustrated teams before. As Ana put it, they didn’t work „because it’s someone else’s process„. One practical bottleneck turned out to be context fragmentation – a team working across 20 or 50 repositories. „This is not an agent problem. This is a process problem.“ It hurts new developers just as much.
To show teams where they stand, RBI uses a maturity model. Many teams see themselves at level 4 or 5 – until it turns out pull requests go to production without checks. Gojko compared the path to agile adoption: first unit tests, then test-driven development, then continuous integration, then shorter iterations. Step by step, with a clear next level.
Gojko’s closing advice: there is budget and excitement for AI right now. Use it.
Use AI as an excuse to improve the process and you will get in a much better place.
If you want to work this out for your own teams: on 30 November 2026 Gojko runs a one-day workshop at TechTalk in Vienna – on moving agentic work from individual productivity to team and organisational performance, spec-driven approaches against rework, humans in control, and lower token cost through deterministic tools. For tech leads, architects, platform and engineering leads and senior developers who already experiment with coding agents.
The slides from the talk are available on the Agile Tour Vienna website.