Two formats for two different jobs: a single-process pilot in 6–8 weeks, or a full rebuild of one or two processes in 12–16 weeks. Our team works alongside yours — ownership stays with you.
Led by the authors of the method — Sergei Lipchanskii and Askhat Urazbaev.
Who this is for
Implementation closes one specific gap: the process is chosen, but you have no team able to build the flywheel by hand. If you are somewhere else on the path, we’ll say so and point you at the right thing.
Your format ✓
We’ll point you elsewhere
The main filter
Implementation works when a process has three things: an owner, accountable for the outcome; accessible data (or a way to assemble it in a week or two); a willing process owner to give half their time for twelve weeks on a rebuild, or five to eight hours a week on a pilot. Without all three, implementation drags, the metrics sag, and the flywheel stops turning three months after go-live.
How it runs
Before we start, we check that the process is genuinely a fit. Every phase ends at a checkpoint with one of three answers: go, conditional go, or stop. If the potential doesn’t hold up at any checkpoint, we say so and close the project — saving you months and the cost of a full rebuild.
Included in the project
Five conditions:
If a process fails the checkpoint, we help you pick another one or run a full discovery. That saves 8–16 weeks and a substantial budget on a pilot that would otherwise have gone into the wrong process.
Weeks 1–4
In-depth interviews with the people who run the process (5–10 of them), observation and data collection (20–30 hours), a baseline measurement, and log analysis. A value-stream map with every step tagged by autonomy level (A1–A5) and by reshape-versus-deploy. Then the economics, the architecture, and a prototype of the new AI process.
Weeks 5–9
The heart of the project: the AI works alongside the expert in shadow mode — it recommends, nothing it decides is applied. We log every divergence, track the disagreement rate, and grow the agent’s Knowledge Base. The threshold to move on is a divergence rate below 30%. Then comes canary: the AI handles 5–10% of real traffic. To go to production: a Human Intervention Rate below 15% and zero critical errors.
Weeks 10–12 (on a pilot, weeks 5–8, compressed)
Daily monitoring, and a first wave of improvements driven by production data. Then handover: the process owner takes direct responsibility for the outcome, and our team moves to on-request advisory.
What does not fit into 6–16 weeks
Those are separate engagements — worth discussing if it comes to that.
What you end up with
The AI has handled at least 500 cases, the Human Intervention Rate (HIR) is recorded, and the kill switch has been tested in the field.
Efficiency (cycle time, Cost-per-Outcome, annual saving), quality (error rate, correct-escalation rate, incidents), and speed (throughput, SLA, response time). Measured before and after — not forecast.
Red lines (what the AI never does), guardrails (where it escalates automatically), and guidance across autonomy levels A1–A5. The runbook covers how to monitor, how to roll back, how to add a rule, how to escalate. Both carry over to the next process.
Five to seven pages for the board, built on actual data. Used both to make the case internally for the next process and to brief the board.
Formats and timelines
The pilot is a board-ready case built on a single process. The rebuild is a production result across one or two. The difference is depth and scale, not the quality of the work.
One process, a full Dual Run, before-and-after metrics, and a case you can take to the board. You come out with a working solution running as a canary on 5–10% of the flow. A rebuild is the next step — or we hand the tuned solution to your team.
The full cycle, from choosing the process to go-live and handover. A process in production, a Rulebook, a runbook, and recommendations for scaling. Two processes can run in parallel across different teams.
Authors
Founder of the AI consulting and platform practice, since 2023.
Co-founder of ScrumTrek, since 2008 — product and Agile transformations.
Where this format came from
Across three years of AI projects we kept hitting the same failure: a pilot started without a proper choice of process, and four weeks in it turned out the process was wrong — no data, no feedback loop, nobody who cared. On the first two projects we lost six to eight weeks each before closing them down. After the third, we made discovery a mandatory stage before anyone signs a contract.
The programme is three years of practice repackaged into a 6–16 week format, alongside the book, the workshop and the diagnostics.
Get in touch
Within one working day: a 45-minute call with one of the people who runs this.
The open method: the Data Flywheel, the five elements of the flywheel, the autonomy spectrum.
Read the methodChoosing the process for a first pilot: self-assessment, or a one-to-two-week audit.
Request diagnosticsTwo days, a team of 8–15 — flywheel architecture and an eight-week roadmap.
See the workshopWe’ll look at your problem, pick the right format, and show where AI actually strengthens your operations.
50+
companies have gone through AI transformation with us — from diagnostics to working solutions.
© 2026 Kiberry — the European practice of ScrumTrek
