- 01Trigger
- What starts the work?
- 02Steps
- What are the main steps? A rough sequence is enough.
- 03Inputs & output
- What goes in, and what should come out?
- 04Frequency
- How often does the work happen?
- 05Rough baseline
- How long does it usually take? Note any obvious quality, uncertainty, or rework issue already known. An estimate is enough.
- 06Friction
- Where is it slow, repetitive, uncertain, frustrating, or error-prone?
- 07Examples
- Bring 1–2 real cases that can safely be used in the lab.
Learn to redesign improvement work with AI — one real workflow at a time
A learning-by-doing Sprint for improvement and development professionals. Each participant or pair brings one recurring workflow from real work. Step by step, the group decides what should stay human, what is better handled with deterministic automation, and where AI should contribute — within the tools, data, permissions, and constraints the organization actually has. Then each person builds a bounded first experiment and learns from real use.
Prepare a workshop · Analyze customer feedback · Turn a transcript into actions · Prepare a project update
1 organization · 1 month · 1 bounded workflow per person or pair · ½-day design lab · 2 learning clinics · 2-hour review
Improvement workflow allocation map
For every concrete step, decide whether the work stays human, follows deterministic rules, or uses AI in a designed role.
How the one-month program unfolds
One workflow from real work moves from preparation to a first experiment, real use, and the next version.
-
Before the lab
Prepare one workflow
Bring one recurring workflow, a rough baseline, and 1–2 usable examples.
-
Week 1 · design lab
Design the first experiment
Clarify the workflow → decide the role of human judgment, deterministic automation, and AI step by step → bound the experiment.
-
Weeks 2–3 · two clinics
Use it and learn in real work
Use the design, bring observations back to the cohort, and improve it from experience.
-
Week 4 · review
Choose the next version
Decide what to keep, change, or test next and what should be considered from the organization’s perspective.
Throughout · available tools · data · permissions · organizational constraints
Bring one recurring workflow from real work
Once the Sprint is agreed and the cohort is set, each person or pair prepares one recurring workflow. The goal is a lightweight current-state picture — not a polished process map and not a pre-selected AI idea.
You do not need to decide where AI belongs. Bring the work — we make the human, rules, automation, and AI choices together in the design lab.
The design lab works through three design steps
Each person or pair arrives with one recurring workflow, a rough baseline, and usable examples. The cohort uses the same three design steps: clarify the workflow end to end, decide the right role for each concrete step, and define one experiment worth taking into real use.
Map the selected workflow end to end
Blank workflow step 1
Blank workflow step 2
Blank workflow step 3
Blank workflow step 4
Use the prepared workflow and real examples to check the trigger, sequence, inputs, outputs, rough timing, handoffs, and edge cases before redesigning anything.
Make one decision per step
- Stays human
Judgment, unclear cases, and decisions with consequences.
- Deterministic automation
Explicit, repeatable logic best handled with deterministic automation.
- AI — with a designed role
AI can support one bounded step or work across a connected sequence. Make its scope, method, context, tools, limits, stop conditions, and human action gates explicit.
Bound the first experiment
- Inside one step ↔ across a sequence
- Ad hoc ↔ defined ↔ controlled & repeatable
- Manual ↔ connected ↔ integrated
- Good output · prevent known errors · stop abnormal autonomy · human judgment / approval
- Time · quality · uncertainty · intervention · failures / friction
Poka-yoke: prevent or detect predictable errors at the source. Jidoka: autonomation with a designed stop or hold, visible abnormal condition, and handback when autonomy should not continue.
Each working surface captures one named decision participants make in the room.
One bounded workflow can carry a different answer at every step
In this example, an improvement professional turns customer comments into themes and reporting. Unclear cases stay human, selected steps use controlled AI, data can move manually now, and richer connections or integration can remain later.
Organizational operating envelope
The organization is part of the workflow design
The lab asks what context the workflow may use, what capability it needs, and what action may follow before the first target is selected.
-
01 / Context
What may the workflow use?
Approved ERP export · permitted data files · Microsoft 365 / SharePoint / Teams · Google Workspace · Jira / Confluence · Miro
Organizational conditionPermissions · data access · security & privacy
-
02 / Capability
What should AI contribute?
Classify · extract · compare · summarize · draft · challenge · use approved context and tools
Available and approved capabilities: Copilot · Claude · ChatGPT / Codex · Gemini
Organizational conditionApproved tools · governance · suitable access
-
03 / Action & handoff
What may happen next?
Human-reviewed output · manual transfer in the first test · later Power Automate / API / MCP / structured storage
Organizational conditionAllowed actions · integration · support capacity
Autonomation & human control across the pathImprovement professional starts the run · autonomation / Jidoka: the workflow continues while agreed conditions hold; if a required source is missing or categorization criteria are not met, it stops or holds and hands the case back to a person · output is checked before use or reporting
Together these conditions defineWorkable now
The horizontal scale runs from Manual through Connected to Integrated. The shaded Workable now band covers the Manual column.
- Get ERP comments First target Later
- Categorize comments First target: controlled AI step, manual enablement Later: richer context
- Handle unclear cases Stays human
- Summarize themes First target: controlled AI step, manual enablement Later: structured storage
- Update reporting Not yet
These working conditions define the shaded Workable now band. A controlled manual workflow can be the strongest first experiment, while later connections remain mapped for deliberate development.
Lean Six Sigma Master Black Belt · sprint facilitator
Several real workflows — one shared way to design and learn
Participants redesign the recurring work they know. Facilitation makes each workflow's choices visible, while the cohort compares evidence, recognizes shared constraints, and identifies patterns worth adapting and testing elsewhere in the organization.
- Participant or pair
- Owns one recurring workflow, names its steps and edge cases, and uses the first design in real work during the month
- The cohort
- Compares evidence across workflows and distinguishes candidate patterns worth testing elsewhere from choices that only make sense in one case
- Internal IT or data specialist — when useful
- Clarifies approved tools, permitted data, allowed AI actions, human-control requirements, and available connections
- Facilitator
- Structures the decisions, bounds the experiment, and keeps the learning tied to evidence from real use
One shared method; different workflows, experiments, and constraints.
Each prototype returns to real work — the evidence returns to the cohort
After the design lab, each participant or pair uses the prototype they designed. The clinics bring the cohort back together to compare evidence across different workflows, refine controls, and recognize organizational enablers that may help more than one case.
Before the labParticipant pre-work
Prepare one recurring workflow
Bring one recurring workflow, a rough baseline, and 1–2 usable examples. The goal is a lightweight current-state picture; decisions about where AI belongs are made in the design lab.
Week 1½-day design lab
Map, allocate, design and prototype
Using the prepared workflow and rough baseline, each person or pair clarifies one selected workflow, makes the human, rules, and AI roles visible, and builds one controlled first experiment.
Week 260-minute clinic
Share what happened
Participants bring real cases from their prototypes: what AI handled, where a stop or human handback occurred, and what surprised them.
Week 360-minute clinic · Yokoten
Compare evidence across the cohort
Compare repeated cases from the different workflows: what became reliable, where controls or organizational constraints shaped the result, and which tested learning another workflow could adapt. Yokoten shares learning across the organization; it does not copy a solution unchanged.
Week 42-hour review
Reflect, standardize and choose the next version
Compare each workflow with its baseline and decide what stays human, what is worth standardizing, what needs another iteration, and which shared learning should be tested in another context next.
Customize the Sprint to the organization
Who should take part, and what kinds of improvement or development work do they own?
Which terminology, examples, or internal methods should the Sprint use?
Which AI tools and workspaces are approved?
Which systems and data may participants use?
Which permissions, privacy, security, or governance constraints matter?
What should the cohort be better able to do after the month?
This setup customizes the shared method. Participants then prepare their own recurring workflow using the pre-work shown earlier on the page.
Let’s talk
If an AI Workflow Sprint could be relevant for your organization, let’s talk about your situation and what a useful implementation could look like.
Let’s talk about the Sprint